src/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_ElecGB_VdwLJ_GeomP1P1_avx_256_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012, by the GROMACS development team, led by
   5  * David van der Spoel, Berk Hess, Erik Lindahl, and including many
   6  * others, as listed in the AUTHORS file in the top-level source
   7  * 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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gmx_math_x86_avx_256_single.h"
  50 #include "kernelutil_x86_avx_256_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_256_single
  54  * Electrostatics interaction: GeneralizedBorn
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_256_single
  61                     (t_nblist * gmx_restrict                nlist,
  62                      rvec * gmx_restrict                    xx,
  63                      rvec * gmx_restrict                    ff,
  64                      t_forcerec * gmx_restrict              fr,
  65                      t_mdatoms * gmx_restrict               mdatoms,
  66                      nb_kernel_data_t * gmx_restrict        kernel_data,
  67                      t_nrnb * gmx_restrict                  nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrE,jnrF,jnrG,jnrH;
  78     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  79     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  80     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  81     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  82     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  83     real             rcutoff_scalar;
  84     real             *shiftvec,*fshift,*x,*f;
  85     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  86     real             scratch[4*DIM];
  87     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  88     real *           vdwioffsetptr0;
  89     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  90     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  91     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  92     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  93     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
  94     real             *charge;
  95     __m256i          gbitab;
  96     __m128i          gbitab_lo,gbitab_hi;
  97     __m256           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
  98     __m256           minushalf = _mm256_set1_ps(-0.5);
  99     real             *invsqrta,*dvda,*gbtab;
 100     int              nvdwtype;
 101     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 102     int              *vdwtype;
 103     real             *vdwparam;
 104     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 105     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 106     __m256i          vfitab;
 107     __m128i          vfitab_lo,vfitab_hi;
 108     __m128i          ifour       = _mm_set1_epi32(4);
 109     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 110     real             *vftab;
 111     __m256           dummy_mask,cutoff_mask;
 112     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 113     __m256           one     = _mm256_set1_ps(1.0);
 114     __m256           two     = _mm256_set1_ps(2.0);
 115     x                = xx[0];
 116     f                = ff[0];
 117
 118     nri              = nlist->nri;
 119     iinr             = nlist->iinr;
 120     jindex           = nlist->jindex;
 121     jjnr             = nlist->jjnr;
 122     shiftidx         = nlist->shift;
 123     gid              = nlist->gid;
 124     shiftvec         = fr->shift_vec[0];
 125     fshift           = fr->fshift[0];
 126     facel            = _mm256_set1_ps(fr->epsfac);
 127     charge           = mdatoms->chargeA;
 128     nvdwtype         = fr->ntype;
 129     vdwparam         = fr->nbfp;
 130     vdwtype          = mdatoms->typeA;
 131
 132     invsqrta         = fr->invsqrta;
 133     dvda             = fr->dvda;
 134     gbtabscale       = _mm256_set1_ps(fr->gbtab.scale);
 135     gbtab            = fr->gbtab.data;
 136     gbinvepsdiff     = _mm256_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 137
 138     /* Avoid stupid compiler warnings */
 139     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 140     j_coord_offsetA = 0;
 141     j_coord_offsetB = 0;
 142     j_coord_offsetC = 0;
 143     j_coord_offsetD = 0;
 144     j_coord_offsetE = 0;
 145     j_coord_offsetF = 0;
 146     j_coord_offsetG = 0;
 147     j_coord_offsetH = 0;
 148
 149     outeriter        = 0;
 150     inneriter        = 0;
 151
 152     for(iidx=0;iidx<4*DIM;iidx++)
 153     {
 154         scratch[iidx] = 0.0;
 155     }
 156
 157     /* Start outer loop over neighborlists */
 158     for(iidx=0; iidx<nri; iidx++)
 159     {
 160         /* Load shift vector for this list */
 161         i_shift_offset   = DIM*shiftidx[iidx];
 162
 163         /* Load limits for loop over neighbors */
 164         j_index_start    = jindex[iidx];
 165         j_index_end      = jindex[iidx+1];
 166
 167         /* Get outer coordinate index */
 168         inr              = iinr[iidx];
 169         i_coord_offset   = DIM*inr;
 170
 171         /* Load i particle coords and add shift vector */
 172         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 173
 174         fix0             = _mm256_setzero_ps();
 175         fiy0             = _mm256_setzero_ps();
 176         fiz0             = _mm256_setzero_ps();
 177
 178         /* Load parameters for i particles */
 179         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 180         isai0            = _mm256_set1_ps(invsqrta[inr+0]);
 181         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 182
 183         /* Reset potential sums */
 184         velecsum         = _mm256_setzero_ps();
 185         vgbsum           = _mm256_setzero_ps();
 186         vvdwsum          = _mm256_setzero_ps();
 187         dvdasum          = _mm256_setzero_ps();
 188
 189         /* Start inner kernel loop */
 190         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 191         {
 192
 193             /* Get j neighbor index, and coordinate index */
 194             jnrA             = jjnr[jidx];
 195             jnrB             = jjnr[jidx+1];
 196             jnrC             = jjnr[jidx+2];
 197             jnrD             = jjnr[jidx+3];
 198             jnrE             = jjnr[jidx+4];
 199             jnrF             = jjnr[jidx+5];
 200             jnrG             = jjnr[jidx+6];
 201             jnrH             = jjnr[jidx+7];
 202             j_coord_offsetA  = DIM*jnrA;
 203             j_coord_offsetB  = DIM*jnrB;
 204             j_coord_offsetC  = DIM*jnrC;
 205             j_coord_offsetD  = DIM*jnrD;
 206             j_coord_offsetE  = DIM*jnrE;
 207             j_coord_offsetF  = DIM*jnrF;
 208             j_coord_offsetG  = DIM*jnrG;
 209             j_coord_offsetH  = DIM*jnrH;
 210
 211             /* load j atom coordinates */
 212             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 213                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 214                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 215                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 216                                                  &jx0,&jy0,&jz0);
 217
 218             /* Calculate displacement vector */
 219             dx00             = _mm256_sub_ps(ix0,jx0);
 220             dy00             = _mm256_sub_ps(iy0,jy0);
 221             dz00             = _mm256_sub_ps(iz0,jz0);
 222
 223             /* Calculate squared distance and things based on it */
 224             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 225
 226             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 227
 228             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 229
 230             /* Load parameters for j particles */
 231             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 232                                                                  charge+jnrC+0,charge+jnrD+0,
 233                                                                  charge+jnrE+0,charge+jnrF+0,
 234                                                                  charge+jnrG+0,charge+jnrH+0);
 235             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 236                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 237                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 238                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 239             vdwjidx0A        = 2*vdwtype[jnrA+0];
 240             vdwjidx0B        = 2*vdwtype[jnrB+0];
 241             vdwjidx0C        = 2*vdwtype[jnrC+0];
 242             vdwjidx0D        = 2*vdwtype[jnrD+0];
 243             vdwjidx0E        = 2*vdwtype[jnrE+0];
 244             vdwjidx0F        = 2*vdwtype[jnrF+0];
 245             vdwjidx0G        = 2*vdwtype[jnrG+0];
 246             vdwjidx0H        = 2*vdwtype[jnrH+0];
 247
 248             /**************************
 249              * CALCULATE INTERACTIONS *
 250              **************************/
 251
 252             r00              = _mm256_mul_ps(rsq00,rinv00);
 253
 254             /* Compute parameters for interactions between i and j atoms */
 255             qq00             = _mm256_mul_ps(iq0,jq0);
 256             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 257                                             vdwioffsetptr0+vdwjidx0B,
 258                                             vdwioffsetptr0+vdwjidx0C,
 259                                             vdwioffsetptr0+vdwjidx0D,
 260                                             vdwioffsetptr0+vdwjidx0E,
 261                                             vdwioffsetptr0+vdwjidx0F,
 262                                             vdwioffsetptr0+vdwjidx0G,
 263                                             vdwioffsetptr0+vdwjidx0H,
 264                                             &c6_00,&c12_00);
 265
 266             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 267             isaprod          = _mm256_mul_ps(isai0,isaj0);
 268             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 269             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 270
 271             /* Calculate generalized born table index - this is a separate table from the normal one,
 272              * but we use the same procedure by multiplying r with scale and truncating to integer.
 273              */
 274             rt               = _mm256_mul_ps(r00,gbscale);
 275             gbitab           = _mm256_cvttps_epi32(rt);
 276             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 277             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 278             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 279             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 280             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 281             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 282             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 283                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 284             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 285                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 286             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 287                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 288             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 289                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 290             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 291             Heps             = _mm256_mul_ps(gbeps,H);
 292             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 293             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 294             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 295
 296             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 297             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 298             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 299             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 300             fjptrA           = dvda+jnrA;
 301             fjptrB           = dvda+jnrB;
 302             fjptrC           = dvda+jnrC;
 303             fjptrD           = dvda+jnrD;
 304             fjptrE           = dvda+jnrE;
 305             fjptrF           = dvda+jnrF;
 306             fjptrG           = dvda+jnrG;
 307             fjptrH           = dvda+jnrH;
 308             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 309                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 310             velec            = _mm256_mul_ps(qq00,rinv00);
 311             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 312
 313             /* LENNARD-JONES DISPERSION/REPULSION */
 314
 315             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 316             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 317             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 318             vvdw             = _mm256_sub_ps( _mm256_mul_ps(vvdw12,one_twelfth) , _mm256_mul_ps(vvdw6,one_sixth) );
 319             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 320
 321             /* Update potential sum for this i atom from the interaction with this j atom. */
 322             velecsum         = _mm256_add_ps(velecsum,velec);
 323             vgbsum           = _mm256_add_ps(vgbsum,vgb);
 324             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 325
 326             fscal            = _mm256_add_ps(felec,fvdw);
 327
 328             /* Calculate temporary vectorial force */
 329             tx               = _mm256_mul_ps(fscal,dx00);
 330             ty               = _mm256_mul_ps(fscal,dy00);
 331             tz               = _mm256_mul_ps(fscal,dz00);
 332
 333             /* Update vectorial force */
 334             fix0             = _mm256_add_ps(fix0,tx);
 335             fiy0             = _mm256_add_ps(fiy0,ty);
 336             fiz0             = _mm256_add_ps(fiz0,tz);
 337
 338             fjptrA             = f+j_coord_offsetA;
 339             fjptrB             = f+j_coord_offsetB;
 340             fjptrC             = f+j_coord_offsetC;
 341             fjptrD             = f+j_coord_offsetD;
 342             fjptrE             = f+j_coord_offsetE;
 343             fjptrF             = f+j_coord_offsetF;
 344             fjptrG             = f+j_coord_offsetG;
 345             fjptrH             = f+j_coord_offsetH;
 346             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 347
 348             /* Inner loop uses 70 flops */
 349         }
 350
 351         if(jidx<j_index_end)
 352         {
 353
 354             /* Get j neighbor index, and coordinate index */
 355             jnrlistA         = jjnr[jidx];
 356             jnrlistB         = jjnr[jidx+1];
 357             jnrlistC         = jjnr[jidx+2];
 358             jnrlistD         = jjnr[jidx+3];
 359             jnrlistE         = jjnr[jidx+4];
 360             jnrlistF         = jjnr[jidx+5];
 361             jnrlistG         = jjnr[jidx+6];
 362             jnrlistH         = jjnr[jidx+7];
 363             /* Sign of each element will be negative for non-real atoms.
 364              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 365              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 366              */
 367             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 368                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 369
 370             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 371             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 372             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 373             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 374             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 375             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 376             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 377             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 378             j_coord_offsetA  = DIM*jnrA;
 379             j_coord_offsetB  = DIM*jnrB;
 380             j_coord_offsetC  = DIM*jnrC;
 381             j_coord_offsetD  = DIM*jnrD;
 382             j_coord_offsetE  = DIM*jnrE;
 383             j_coord_offsetF  = DIM*jnrF;
 384             j_coord_offsetG  = DIM*jnrG;
 385             j_coord_offsetH  = DIM*jnrH;
 386
 387             /* load j atom coordinates */
 388             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 389                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 390                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 391                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 392                                                  &jx0,&jy0,&jz0);
 393
 394             /* Calculate displacement vector */
 395             dx00             = _mm256_sub_ps(ix0,jx0);
 396             dy00             = _mm256_sub_ps(iy0,jy0);
 397             dz00             = _mm256_sub_ps(iz0,jz0);
 398
 399             /* Calculate squared distance and things based on it */
 400             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 401
 402             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 403
 404             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 405
 406             /* Load parameters for j particles */
 407             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 408                                                                  charge+jnrC+0,charge+jnrD+0,
 409                                                                  charge+jnrE+0,charge+jnrF+0,
 410                                                                  charge+jnrG+0,charge+jnrH+0);
 411             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 412                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 413                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 414                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 415             vdwjidx0A        = 2*vdwtype[jnrA+0];
 416             vdwjidx0B        = 2*vdwtype[jnrB+0];
 417             vdwjidx0C        = 2*vdwtype[jnrC+0];
 418             vdwjidx0D        = 2*vdwtype[jnrD+0];
 419             vdwjidx0E        = 2*vdwtype[jnrE+0];
 420             vdwjidx0F        = 2*vdwtype[jnrF+0];
 421             vdwjidx0G        = 2*vdwtype[jnrG+0];
 422             vdwjidx0H        = 2*vdwtype[jnrH+0];
 423
 424             /**************************
 425              * CALCULATE INTERACTIONS *
 426              **************************/
 427
 428             r00              = _mm256_mul_ps(rsq00,rinv00);
 429             r00              = _mm256_andnot_ps(dummy_mask,r00);
 430
 431             /* Compute parameters for interactions between i and j atoms */
 432             qq00             = _mm256_mul_ps(iq0,jq0);
 433             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 434                                             vdwioffsetptr0+vdwjidx0B,
 435                                             vdwioffsetptr0+vdwjidx0C,
 436                                             vdwioffsetptr0+vdwjidx0D,
 437                                             vdwioffsetptr0+vdwjidx0E,
 438                                             vdwioffsetptr0+vdwjidx0F,
 439                                             vdwioffsetptr0+vdwjidx0G,
 440                                             vdwioffsetptr0+vdwjidx0H,
 441                                             &c6_00,&c12_00);
 442
 443             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 444             isaprod          = _mm256_mul_ps(isai0,isaj0);
 445             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 446             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 447
 448             /* Calculate generalized born table index - this is a separate table from the normal one,
 449              * but we use the same procedure by multiplying r with scale and truncating to integer.
 450              */
 451             rt               = _mm256_mul_ps(r00,gbscale);
 452             gbitab           = _mm256_cvttps_epi32(rt);
 453             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 454             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 455             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 456             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 457             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 458             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 459             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 460                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 461             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 462                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 463             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 464                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 465             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 466                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 467             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 468             Heps             = _mm256_mul_ps(gbeps,H);
 469             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 470             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 471             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 472
 473             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 474             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 475             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 476             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 477             /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
 478             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 479             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 480             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 481             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 482             fjptrE             = (jnrlistE>=0) ? dvda+jnrE : scratch;
 483             fjptrF             = (jnrlistF>=0) ? dvda+jnrF : scratch;
 484             fjptrG             = (jnrlistG>=0) ? dvda+jnrG : scratch;
 485             fjptrH             = (jnrlistH>=0) ? dvda+jnrH : scratch;
 486             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 487                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 488             velec            = _mm256_mul_ps(qq00,rinv00);
 489             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 490
 491             /* LENNARD-JONES DISPERSION/REPULSION */
 492
 493             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 494             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 495             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 496             vvdw             = _mm256_sub_ps( _mm256_mul_ps(vvdw12,one_twelfth) , _mm256_mul_ps(vvdw6,one_sixth) );
 497             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 498
 499             /* Update potential sum for this i atom from the interaction with this j atom. */
 500             velec            = _mm256_andnot_ps(dummy_mask,velec);
 501             velecsum         = _mm256_add_ps(velecsum,velec);
 502             vgb              = _mm256_andnot_ps(dummy_mask,vgb);
 503             vgbsum           = _mm256_add_ps(vgbsum,vgb);
 504             vvdw             = _mm256_andnot_ps(dummy_mask,vvdw);
 505             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 506
 507             fscal            = _mm256_add_ps(felec,fvdw);
 508
 509             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 510
 511             /* Calculate temporary vectorial force */
 512             tx               = _mm256_mul_ps(fscal,dx00);
 513             ty               = _mm256_mul_ps(fscal,dy00);
 514             tz               = _mm256_mul_ps(fscal,dz00);
 515
 516             /* Update vectorial force */
 517             fix0             = _mm256_add_ps(fix0,tx);
 518             fiy0             = _mm256_add_ps(fiy0,ty);
 519             fiz0             = _mm256_add_ps(fiz0,tz);
 520
 521             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 522             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 523             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 524             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 525             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 526             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 527             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 528             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 529             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 530
 531             /* Inner loop uses 71 flops */
 532         }
 533
 534         /* End of innermost loop */
 535
 536         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 537                                                  f+i_coord_offset,fshift+i_shift_offset);
 538
 539         ggid                        = gid[iidx];
 540         /* Update potential energies */
 541         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 542         gmx_mm256_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 543         gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 544         dvdasum = _mm256_mul_ps(dvdasum, _mm256_mul_ps(isai0,isai0));
 545         gmx_mm256_update_1pot_ps(dvdasum,dvda+inr);
 546
 547         /* Increment number of inner iterations */
 548         inneriter                  += j_index_end - j_index_start;
 549
 550         /* Outer loop uses 10 flops */
 551     }
 552
 553     /* Increment number of outer iterations */
 554     outeriter        += nri;
 555
 556     /* Update outer/inner flops */
 557
 558     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*71);
 559 }
 560 /*
 561  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_single
 562  * Electrostatics interaction: GeneralizedBorn
 563  * VdW interaction:            LennardJones
 564  * Geometry:                   Particle-Particle
 565  * Calculate force/pot:        Force
 566  */
 567 void
 568 nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_single
 569                     (t_nblist * gmx_restrict                nlist,
 570                      rvec * gmx_restrict                    xx,
 571                      rvec * gmx_restrict                    ff,
 572                      t_forcerec * gmx_restrict              fr,
 573                      t_mdatoms * gmx_restrict               mdatoms,
 574                      nb_kernel_data_t * gmx_restrict        kernel_data,
 575                      t_nrnb * gmx_restrict                  nrnb)
 576 {
 577     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 578      * just 0 for non-waters.
 579      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 580      * jnr indices corresponding to data put in the four positions in the SIMD register.
 581      */
 582     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 583     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 584     int              jnrA,jnrB,jnrC,jnrD;
 585     int              jnrE,jnrF,jnrG,jnrH;
 586     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 587     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 588     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 589     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 590     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 591     real             rcutoff_scalar;
 592     real             *shiftvec,*fshift,*x,*f;
 593     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 594     real             scratch[4*DIM];
 595     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 596     real *           vdwioffsetptr0;
 597     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 598     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 599     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 600     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 601     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 602     real             *charge;
 603     __m256i          gbitab;
 604     __m128i          gbitab_lo,gbitab_hi;
 605     __m256           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 606     __m256           minushalf = _mm256_set1_ps(-0.5);
 607     real             *invsqrta,*dvda,*gbtab;
 608     int              nvdwtype;
 609     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 610     int              *vdwtype;
 611     real             *vdwparam;
 612     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 613     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 614     __m256i          vfitab;
 615     __m128i          vfitab_lo,vfitab_hi;
 616     __m128i          ifour       = _mm_set1_epi32(4);
 617     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 618     real             *vftab;
 619     __m256           dummy_mask,cutoff_mask;
 620     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 621     __m256           one     = _mm256_set1_ps(1.0);
 622     __m256           two     = _mm256_set1_ps(2.0);
 623     x                = xx[0];
 624     f                = ff[0];
 625
 626     nri              = nlist->nri;
 627     iinr             = nlist->iinr;
 628     jindex           = nlist->jindex;
 629     jjnr             = nlist->jjnr;
 630     shiftidx         = nlist->shift;
 631     gid              = nlist->gid;
 632     shiftvec         = fr->shift_vec[0];
 633     fshift           = fr->fshift[0];
 634     facel            = _mm256_set1_ps(fr->epsfac);
 635     charge           = mdatoms->chargeA;
 636     nvdwtype         = fr->ntype;
 637     vdwparam         = fr->nbfp;
 638     vdwtype          = mdatoms->typeA;
 639
 640     invsqrta         = fr->invsqrta;
 641     dvda             = fr->dvda;
 642     gbtabscale       = _mm256_set1_ps(fr->gbtab.scale);
 643     gbtab            = fr->gbtab.data;
 644     gbinvepsdiff     = _mm256_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 645
 646     /* Avoid stupid compiler warnings */
 647     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 648     j_coord_offsetA = 0;
 649     j_coord_offsetB = 0;
 650     j_coord_offsetC = 0;
 651     j_coord_offsetD = 0;
 652     j_coord_offsetE = 0;
 653     j_coord_offsetF = 0;
 654     j_coord_offsetG = 0;
 655     j_coord_offsetH = 0;
 656
 657     outeriter        = 0;
 658     inneriter        = 0;
 659
 660     for(iidx=0;iidx<4*DIM;iidx++)
 661     {
 662         scratch[iidx] = 0.0;
 663     }
 664
 665     /* Start outer loop over neighborlists */
 666     for(iidx=0; iidx<nri; iidx++)
 667     {
 668         /* Load shift vector for this list */
 669         i_shift_offset   = DIM*shiftidx[iidx];
 670
 671         /* Load limits for loop over neighbors */
 672         j_index_start    = jindex[iidx];
 673         j_index_end      = jindex[iidx+1];
 674
 675         /* Get outer coordinate index */
 676         inr              = iinr[iidx];
 677         i_coord_offset   = DIM*inr;
 678
 679         /* Load i particle coords and add shift vector */
 680         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 681
 682         fix0             = _mm256_setzero_ps();
 683         fiy0             = _mm256_setzero_ps();
 684         fiz0             = _mm256_setzero_ps();
 685
 686         /* Load parameters for i particles */
 687         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 688         isai0            = _mm256_set1_ps(invsqrta[inr+0]);
 689         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 690
 691         dvdasum          = _mm256_setzero_ps();
 692
 693         /* Start inner kernel loop */
 694         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 695         {
 696
 697             /* Get j neighbor index, and coordinate index */
 698             jnrA             = jjnr[jidx];
 699             jnrB             = jjnr[jidx+1];
 700             jnrC             = jjnr[jidx+2];
 701             jnrD             = jjnr[jidx+3];
 702             jnrE             = jjnr[jidx+4];
 703             jnrF             = jjnr[jidx+5];
 704             jnrG             = jjnr[jidx+6];
 705             jnrH             = jjnr[jidx+7];
 706             j_coord_offsetA  = DIM*jnrA;
 707             j_coord_offsetB  = DIM*jnrB;
 708             j_coord_offsetC  = DIM*jnrC;
 709             j_coord_offsetD  = DIM*jnrD;
 710             j_coord_offsetE  = DIM*jnrE;
 711             j_coord_offsetF  = DIM*jnrF;
 712             j_coord_offsetG  = DIM*jnrG;
 713             j_coord_offsetH  = DIM*jnrH;
 714
 715             /* load j atom coordinates */
 716             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 717                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 718                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 719                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 720                                                  &jx0,&jy0,&jz0);
 721
 722             /* Calculate displacement vector */
 723             dx00             = _mm256_sub_ps(ix0,jx0);
 724             dy00             = _mm256_sub_ps(iy0,jy0);
 725             dz00             = _mm256_sub_ps(iz0,jz0);
 726
 727             /* Calculate squared distance and things based on it */
 728             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 729
 730             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 731
 732             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 733
 734             /* Load parameters for j particles */
 735             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 736                                                                  charge+jnrC+0,charge+jnrD+0,
 737                                                                  charge+jnrE+0,charge+jnrF+0,
 738                                                                  charge+jnrG+0,charge+jnrH+0);
 739             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 740                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 741                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 742                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 743             vdwjidx0A        = 2*vdwtype[jnrA+0];
 744             vdwjidx0B        = 2*vdwtype[jnrB+0];
 745             vdwjidx0C        = 2*vdwtype[jnrC+0];
 746             vdwjidx0D        = 2*vdwtype[jnrD+0];
 747             vdwjidx0E        = 2*vdwtype[jnrE+0];
 748             vdwjidx0F        = 2*vdwtype[jnrF+0];
 749             vdwjidx0G        = 2*vdwtype[jnrG+0];
 750             vdwjidx0H        = 2*vdwtype[jnrH+0];
 751
 752             /**************************
 753              * CALCULATE INTERACTIONS *
 754              **************************/
 755
 756             r00              = _mm256_mul_ps(rsq00,rinv00);
 757
 758             /* Compute parameters for interactions between i and j atoms */
 759             qq00             = _mm256_mul_ps(iq0,jq0);
 760             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 761                                             vdwioffsetptr0+vdwjidx0B,
 762                                             vdwioffsetptr0+vdwjidx0C,
 763                                             vdwioffsetptr0+vdwjidx0D,
 764                                             vdwioffsetptr0+vdwjidx0E,
 765                                             vdwioffsetptr0+vdwjidx0F,
 766                                             vdwioffsetptr0+vdwjidx0G,
 767                                             vdwioffsetptr0+vdwjidx0H,
 768                                             &c6_00,&c12_00);
 769
 770             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 771             isaprod          = _mm256_mul_ps(isai0,isaj0);
 772             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 773             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 774
 775             /* Calculate generalized born table index - this is a separate table from the normal one,
 776              * but we use the same procedure by multiplying r with scale and truncating to integer.
 777              */
 778             rt               = _mm256_mul_ps(r00,gbscale);
 779             gbitab           = _mm256_cvttps_epi32(rt);
 780             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 781             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 782             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 783             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 784             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 785             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 786             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 787                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 788             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 789                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 790             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 791                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 792             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 793                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 794             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 795             Heps             = _mm256_mul_ps(gbeps,H);
 796             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 797             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 798             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 799
 800             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 801             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 802             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 803             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 804             fjptrA           = dvda+jnrA;
 805             fjptrB           = dvda+jnrB;
 806             fjptrC           = dvda+jnrC;
 807             fjptrD           = dvda+jnrD;
 808             fjptrE           = dvda+jnrE;
 809             fjptrF           = dvda+jnrF;
 810             fjptrG           = dvda+jnrG;
 811             fjptrH           = dvda+jnrH;
 812             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 813                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 814             velec            = _mm256_mul_ps(qq00,rinv00);
 815             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 816
 817             /* LENNARD-JONES DISPERSION/REPULSION */
 818
 819             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 820             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
 821
 822             fscal            = _mm256_add_ps(felec,fvdw);
 823
 824             /* Calculate temporary vectorial force */
 825             tx               = _mm256_mul_ps(fscal,dx00);
 826             ty               = _mm256_mul_ps(fscal,dy00);
 827             tz               = _mm256_mul_ps(fscal,dz00);
 828
 829             /* Update vectorial force */
 830             fix0             = _mm256_add_ps(fix0,tx);
 831             fiy0             = _mm256_add_ps(fiy0,ty);
 832             fiz0             = _mm256_add_ps(fiz0,tz);
 833
 834             fjptrA             = f+j_coord_offsetA;
 835             fjptrB             = f+j_coord_offsetB;
 836             fjptrC             = f+j_coord_offsetC;
 837             fjptrD             = f+j_coord_offsetD;
 838             fjptrE             = f+j_coord_offsetE;
 839             fjptrF             = f+j_coord_offsetF;
 840             fjptrG             = f+j_coord_offsetG;
 841             fjptrH             = f+j_coord_offsetH;
 842             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 843
 844             /* Inner loop uses 63 flops */
 845         }
 846
 847         if(jidx<j_index_end)
 848         {
 849
 850             /* Get j neighbor index, and coordinate index */
 851             jnrlistA         = jjnr[jidx];
 852             jnrlistB         = jjnr[jidx+1];
 853             jnrlistC         = jjnr[jidx+2];
 854             jnrlistD         = jjnr[jidx+3];
 855             jnrlistE         = jjnr[jidx+4];
 856             jnrlistF         = jjnr[jidx+5];
 857             jnrlistG         = jjnr[jidx+6];
 858             jnrlistH         = jjnr[jidx+7];
 859             /* Sign of each element will be negative for non-real atoms.
 860              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 861              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 862              */
 863             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 864                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 865
 866             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 867             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 868             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 869             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 870             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 871             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 872             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 873             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 874             j_coord_offsetA  = DIM*jnrA;
 875             j_coord_offsetB  = DIM*jnrB;
 876             j_coord_offsetC  = DIM*jnrC;
 877             j_coord_offsetD  = DIM*jnrD;
 878             j_coord_offsetE  = DIM*jnrE;
 879             j_coord_offsetF  = DIM*jnrF;
 880             j_coord_offsetG  = DIM*jnrG;
 881             j_coord_offsetH  = DIM*jnrH;
 882
 883             /* load j atom coordinates */
 884             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 885                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 886                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 887                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 888                                                  &jx0,&jy0,&jz0);
 889
 890             /* Calculate displacement vector */
 891             dx00             = _mm256_sub_ps(ix0,jx0);
 892             dy00             = _mm256_sub_ps(iy0,jy0);
 893             dz00             = _mm256_sub_ps(iz0,jz0);
 894
 895             /* Calculate squared distance and things based on it */
 896             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 897
 898             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 899
 900             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 901
 902             /* Load parameters for j particles */
 903             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 904                                                                  charge+jnrC+0,charge+jnrD+0,
 905                                                                  charge+jnrE+0,charge+jnrF+0,
 906                                                                  charge+jnrG+0,charge+jnrH+0);
 907             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 908                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 909                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 910                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 911             vdwjidx0A        = 2*vdwtype[jnrA+0];
 912             vdwjidx0B        = 2*vdwtype[jnrB+0];
 913             vdwjidx0C        = 2*vdwtype[jnrC+0];
 914             vdwjidx0D        = 2*vdwtype[jnrD+0];
 915             vdwjidx0E        = 2*vdwtype[jnrE+0];
 916             vdwjidx0F        = 2*vdwtype[jnrF+0];
 917             vdwjidx0G        = 2*vdwtype[jnrG+0];
 918             vdwjidx0H        = 2*vdwtype[jnrH+0];
 919
 920             /**************************
 921              * CALCULATE INTERACTIONS *
 922              **************************/
 923
 924             r00              = _mm256_mul_ps(rsq00,rinv00);
 925             r00              = _mm256_andnot_ps(dummy_mask,r00);
 926
 927             /* Compute parameters for interactions between i and j atoms */
 928             qq00             = _mm256_mul_ps(iq0,jq0);
 929             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 930                                             vdwioffsetptr0+vdwjidx0B,
 931                                             vdwioffsetptr0+vdwjidx0C,
 932                                             vdwioffsetptr0+vdwjidx0D,
 933                                             vdwioffsetptr0+vdwjidx0E,
 934                                             vdwioffsetptr0+vdwjidx0F,
 935                                             vdwioffsetptr0+vdwjidx0G,
 936                                             vdwioffsetptr0+vdwjidx0H,
 937                                             &c6_00,&c12_00);
 938
 939             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 940             isaprod          = _mm256_mul_ps(isai0,isaj0);
 941             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 942             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 943
 944             /* Calculate generalized born table index - this is a separate table from the normal one,
 945              * but we use the same procedure by multiplying r with scale and truncating to integer.
 946              */
 947             rt               = _mm256_mul_ps(r00,gbscale);
 948             gbitab           = _mm256_cvttps_epi32(rt);
 949             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 950             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 951             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 952             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 953             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 954             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 955             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 956                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 957             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 958                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 959             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 960                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 961             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 962                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 963             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 964             Heps             = _mm256_mul_ps(gbeps,H);
 965             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 966             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 967             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 968
 969             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 970             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 971             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 972             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 973             /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
 974             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 975             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 976             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 977             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 978             fjptrE             = (jnrlistE>=0) ? dvda+jnrE : scratch;
 979             fjptrF             = (jnrlistF>=0) ? dvda+jnrF : scratch;
 980             fjptrG             = (jnrlistG>=0) ? dvda+jnrG : scratch;
 981             fjptrH             = (jnrlistH>=0) ? dvda+jnrH : scratch;
 982             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 983                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 984             velec            = _mm256_mul_ps(qq00,rinv00);
 985             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 986
 987             /* LENNARD-JONES DISPERSION/REPULSION */
 988
 989             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 990             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
 991
 992             fscal            = _mm256_add_ps(felec,fvdw);
 993
 994             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 995
 996             /* Calculate temporary vectorial force */
 997             tx               = _mm256_mul_ps(fscal,dx00);
 998             ty               = _mm256_mul_ps(fscal,dy00);
 999             tz               = _mm256_mul_ps(fscal,dz00);
1000
1001             /* Update vectorial force */
1002             fix0             = _mm256_add_ps(fix0,tx);
1003             fiy0             = _mm256_add_ps(fiy0,ty);
1004             fiz0             = _mm256_add_ps(fiz0,tz);
1005
1006             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1007             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1008             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1009             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1010             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
1011             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
1012             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
1013             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
1014             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
1015
1016             /* Inner loop uses 64 flops */
1017         }
1018
1019         /* End of innermost loop */
1020
1021         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
1022                                                  f+i_coord_offset,fshift+i_shift_offset);
1023
1024         dvdasum = _mm256_mul_ps(dvdasum, _mm256_mul_ps(isai0,isai0));
1025         gmx_mm256_update_1pot_ps(dvdasum,dvda+inr);
1026
1027         /* Increment number of inner iterations */
1028         inneriter                  += j_index_end - j_index_start;
1029
1030         /* Outer loop uses 7 flops */
1031     }
1032
1033     /* Increment number of outer iterations */
1034     outeriter        += nri;
1035
1036     /* Update outer/inner flops */
1037
1038     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*64);
1039 }