src/gromacs/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) 2012,2013,2014,2015,2017, 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 "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_256_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_256_single
  51  * Electrostatics interaction: GeneralizedBorn
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_256_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrE,jnrF,jnrG,jnrH;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  77     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  78     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  83     real             scratch[4*DIM];
  84     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  85     real *           vdwioffsetptr0;
  86     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  88     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     __m256i          gbitab;
  93     __m128i          gbitab_lo,gbitab_hi;
  94     __m256           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
  95     __m256           minushalf = _mm256_set1_ps(-0.5);
  96     real             *invsqrta,*dvda,*gbtab;
  97     int              nvdwtype;
  98     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  99     int              *vdwtype;
 100     real             *vdwparam;
 101     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 102     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 103     __m256i          vfitab;
 104     __m128i          vfitab_lo,vfitab_hi;
 105     __m128i          ifour       = _mm_set1_epi32(4);
 106     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 107     real             *vftab;
 108     __m256           dummy_mask,cutoff_mask;
 109     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 110     __m256           one     = _mm256_set1_ps(1.0);
 111     __m256           two     = _mm256_set1_ps(2.0);
 112     x                = xx[0];
 113     f                = ff[0];
 114
 115     nri              = nlist->nri;
 116     iinr             = nlist->iinr;
 117     jindex           = nlist->jindex;
 118     jjnr             = nlist->jjnr;
 119     shiftidx         = nlist->shift;
 120     gid              = nlist->gid;
 121     shiftvec         = fr->shift_vec[0];
 122     fshift           = fr->fshift[0];
 123     facel            = _mm256_set1_ps(fr->ic->epsfac);
 124     charge           = mdatoms->chargeA;
 125     nvdwtype         = fr->ntype;
 126     vdwparam         = fr->nbfp;
 127     vdwtype          = mdatoms->typeA;
 128
 129     invsqrta         = fr->invsqrta;
 130     dvda             = fr->dvda;
 131     gbtabscale       = _mm256_set1_ps(fr->gbtab->scale);
 132     gbtab            = fr->gbtab->data;
 133     gbinvepsdiff     = _mm256_set1_ps((1.0/fr->ic->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 134
 135     /* Avoid stupid compiler warnings */
 136     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 137     j_coord_offsetA = 0;
 138     j_coord_offsetB = 0;
 139     j_coord_offsetC = 0;
 140     j_coord_offsetD = 0;
 141     j_coord_offsetE = 0;
 142     j_coord_offsetF = 0;
 143     j_coord_offsetG = 0;
 144     j_coord_offsetH = 0;
 145
 146     outeriter        = 0;
 147     inneriter        = 0;
 148
 149     for(iidx=0;iidx<4*DIM;iidx++)
 150     {
 151         scratch[iidx] = 0.0;
 152     }
 153
 154     /* Start outer loop over neighborlists */
 155     for(iidx=0; iidx<nri; iidx++)
 156     {
 157         /* Load shift vector for this list */
 158         i_shift_offset   = DIM*shiftidx[iidx];
 159
 160         /* Load limits for loop over neighbors */
 161         j_index_start    = jindex[iidx];
 162         j_index_end      = jindex[iidx+1];
 163
 164         /* Get outer coordinate index */
 165         inr              = iinr[iidx];
 166         i_coord_offset   = DIM*inr;
 167
 168         /* Load i particle coords and add shift vector */
 169         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 170
 171         fix0             = _mm256_setzero_ps();
 172         fiy0             = _mm256_setzero_ps();
 173         fiz0             = _mm256_setzero_ps();
 174
 175         /* Load parameters for i particles */
 176         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 177         isai0            = _mm256_set1_ps(invsqrta[inr+0]);
 178         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 179
 180         /* Reset potential sums */
 181         velecsum         = _mm256_setzero_ps();
 182         vgbsum           = _mm256_setzero_ps();
 183         vvdwsum          = _mm256_setzero_ps();
 184         dvdasum          = _mm256_setzero_ps();
 185
 186         /* Start inner kernel loop */
 187         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 188         {
 189
 190             /* Get j neighbor index, and coordinate index */
 191             jnrA             = jjnr[jidx];
 192             jnrB             = jjnr[jidx+1];
 193             jnrC             = jjnr[jidx+2];
 194             jnrD             = jjnr[jidx+3];
 195             jnrE             = jjnr[jidx+4];
 196             jnrF             = jjnr[jidx+5];
 197             jnrG             = jjnr[jidx+6];
 198             jnrH             = jjnr[jidx+7];
 199             j_coord_offsetA  = DIM*jnrA;
 200             j_coord_offsetB  = DIM*jnrB;
 201             j_coord_offsetC  = DIM*jnrC;
 202             j_coord_offsetD  = DIM*jnrD;
 203             j_coord_offsetE  = DIM*jnrE;
 204             j_coord_offsetF  = DIM*jnrF;
 205             j_coord_offsetG  = DIM*jnrG;
 206             j_coord_offsetH  = DIM*jnrH;
 207
 208             /* load j atom coordinates */
 209             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 210                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 211                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 212                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 213                                                  &jx0,&jy0,&jz0);
 214
 215             /* Calculate displacement vector */
 216             dx00             = _mm256_sub_ps(ix0,jx0);
 217             dy00             = _mm256_sub_ps(iy0,jy0);
 218             dz00             = _mm256_sub_ps(iz0,jz0);
 219
 220             /* Calculate squared distance and things based on it */
 221             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 222
 223             rinv00           = avx256_invsqrt_f(rsq00);
 224
 225             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 226
 227             /* Load parameters for j particles */
 228             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 229                                                                  charge+jnrC+0,charge+jnrD+0,
 230                                                                  charge+jnrE+0,charge+jnrF+0,
 231                                                                  charge+jnrG+0,charge+jnrH+0);
 232             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 233                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 234                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 235                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 236             vdwjidx0A        = 2*vdwtype[jnrA+0];
 237             vdwjidx0B        = 2*vdwtype[jnrB+0];
 238             vdwjidx0C        = 2*vdwtype[jnrC+0];
 239             vdwjidx0D        = 2*vdwtype[jnrD+0];
 240             vdwjidx0E        = 2*vdwtype[jnrE+0];
 241             vdwjidx0F        = 2*vdwtype[jnrF+0];
 242             vdwjidx0G        = 2*vdwtype[jnrG+0];
 243             vdwjidx0H        = 2*vdwtype[jnrH+0];
 244
 245             /**************************
 246              * CALCULATE INTERACTIONS *
 247              **************************/
 248
 249             r00              = _mm256_mul_ps(rsq00,rinv00);
 250
 251             /* Compute parameters for interactions between i and j atoms */
 252             qq00             = _mm256_mul_ps(iq0,jq0);
 253             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 254                                             vdwioffsetptr0+vdwjidx0B,
 255                                             vdwioffsetptr0+vdwjidx0C,
 256                                             vdwioffsetptr0+vdwjidx0D,
 257                                             vdwioffsetptr0+vdwjidx0E,
 258                                             vdwioffsetptr0+vdwjidx0F,
 259                                             vdwioffsetptr0+vdwjidx0G,
 260                                             vdwioffsetptr0+vdwjidx0H,
 261                                             &c6_00,&c12_00);
 262
 263             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 264             isaprod          = _mm256_mul_ps(isai0,isaj0);
 265             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 266             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 267
 268             /* Calculate generalized born table index - this is a separate table from the normal one,
 269              * but we use the same procedure by multiplying r with scale and truncating to integer.
 270              */
 271             rt               = _mm256_mul_ps(r00,gbscale);
 272             gbitab           = _mm256_cvttps_epi32(rt);
 273             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 274             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 275             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 276             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 277             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 278             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 279             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 280                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 281             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 282                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 283             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 284                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 285             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 286                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 287             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 288             Heps             = _mm256_mul_ps(gbeps,H);
 289             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 290             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 291             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 292
 293             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 294             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 295             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 296             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 297             fjptrA           = dvda+jnrA;
 298             fjptrB           = dvda+jnrB;
 299             fjptrC           = dvda+jnrC;
 300             fjptrD           = dvda+jnrD;
 301             fjptrE           = dvda+jnrE;
 302             fjptrF           = dvda+jnrF;
 303             fjptrG           = dvda+jnrG;
 304             fjptrH           = dvda+jnrH;
 305             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 306                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 307             velec            = _mm256_mul_ps(qq00,rinv00);
 308             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 309
 310             /* LENNARD-JONES DISPERSION/REPULSION */
 311
 312             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 313             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 314             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 315             vvdw             = _mm256_sub_ps( _mm256_mul_ps(vvdw12,one_twelfth) , _mm256_mul_ps(vvdw6,one_sixth) );
 316             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 317
 318             /* Update potential sum for this i atom from the interaction with this j atom. */
 319             velecsum         = _mm256_add_ps(velecsum,velec);
 320             vgbsum           = _mm256_add_ps(vgbsum,vgb);
 321             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 322
 323             fscal            = _mm256_add_ps(felec,fvdw);
 324
 325             /* Calculate temporary vectorial force */
 326             tx               = _mm256_mul_ps(fscal,dx00);
 327             ty               = _mm256_mul_ps(fscal,dy00);
 328             tz               = _mm256_mul_ps(fscal,dz00);
 329
 330             /* Update vectorial force */
 331             fix0             = _mm256_add_ps(fix0,tx);
 332             fiy0             = _mm256_add_ps(fiy0,ty);
 333             fiz0             = _mm256_add_ps(fiz0,tz);
 334
 335             fjptrA             = f+j_coord_offsetA;
 336             fjptrB             = f+j_coord_offsetB;
 337             fjptrC             = f+j_coord_offsetC;
 338             fjptrD             = f+j_coord_offsetD;
 339             fjptrE             = f+j_coord_offsetE;
 340             fjptrF             = f+j_coord_offsetF;
 341             fjptrG             = f+j_coord_offsetG;
 342             fjptrH             = f+j_coord_offsetH;
 343             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 344
 345             /* Inner loop uses 70 flops */
 346         }
 347
 348         if(jidx<j_index_end)
 349         {
 350
 351             /* Get j neighbor index, and coordinate index */
 352             jnrlistA         = jjnr[jidx];
 353             jnrlistB         = jjnr[jidx+1];
 354             jnrlistC         = jjnr[jidx+2];
 355             jnrlistD         = jjnr[jidx+3];
 356             jnrlistE         = jjnr[jidx+4];
 357             jnrlistF         = jjnr[jidx+5];
 358             jnrlistG         = jjnr[jidx+6];
 359             jnrlistH         = jjnr[jidx+7];
 360             /* Sign of each element will be negative for non-real atoms.
 361              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 362              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 363              */
 364             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 365                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 366
 367             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 368             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 369             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 370             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 371             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 372             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 373             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 374             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 375             j_coord_offsetA  = DIM*jnrA;
 376             j_coord_offsetB  = DIM*jnrB;
 377             j_coord_offsetC  = DIM*jnrC;
 378             j_coord_offsetD  = DIM*jnrD;
 379             j_coord_offsetE  = DIM*jnrE;
 380             j_coord_offsetF  = DIM*jnrF;
 381             j_coord_offsetG  = DIM*jnrG;
 382             j_coord_offsetH  = DIM*jnrH;
 383
 384             /* load j atom coordinates */
 385             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 386                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 387                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 388                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 389                                                  &jx0,&jy0,&jz0);
 390
 391             /* Calculate displacement vector */
 392             dx00             = _mm256_sub_ps(ix0,jx0);
 393             dy00             = _mm256_sub_ps(iy0,jy0);
 394             dz00             = _mm256_sub_ps(iz0,jz0);
 395
 396             /* Calculate squared distance and things based on it */
 397             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 398
 399             rinv00           = avx256_invsqrt_f(rsq00);
 400
 401             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 402
 403             /* Load parameters for j particles */
 404             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 405                                                                  charge+jnrC+0,charge+jnrD+0,
 406                                                                  charge+jnrE+0,charge+jnrF+0,
 407                                                                  charge+jnrG+0,charge+jnrH+0);
 408             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 409                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 410                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 411                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 412             vdwjidx0A        = 2*vdwtype[jnrA+0];
 413             vdwjidx0B        = 2*vdwtype[jnrB+0];
 414             vdwjidx0C        = 2*vdwtype[jnrC+0];
 415             vdwjidx0D        = 2*vdwtype[jnrD+0];
 416             vdwjidx0E        = 2*vdwtype[jnrE+0];
 417             vdwjidx0F        = 2*vdwtype[jnrF+0];
 418             vdwjidx0G        = 2*vdwtype[jnrG+0];
 419             vdwjidx0H        = 2*vdwtype[jnrH+0];
 420
 421             /**************************
 422              * CALCULATE INTERACTIONS *
 423              **************************/
 424
 425             r00              = _mm256_mul_ps(rsq00,rinv00);
 426             r00              = _mm256_andnot_ps(dummy_mask,r00);
 427
 428             /* Compute parameters for interactions between i and j atoms */
 429             qq00             = _mm256_mul_ps(iq0,jq0);
 430             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 431                                             vdwioffsetptr0+vdwjidx0B,
 432                                             vdwioffsetptr0+vdwjidx0C,
 433                                             vdwioffsetptr0+vdwjidx0D,
 434                                             vdwioffsetptr0+vdwjidx0E,
 435                                             vdwioffsetptr0+vdwjidx0F,
 436                                             vdwioffsetptr0+vdwjidx0G,
 437                                             vdwioffsetptr0+vdwjidx0H,
 438                                             &c6_00,&c12_00);
 439
 440             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 441             isaprod          = _mm256_mul_ps(isai0,isaj0);
 442             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 443             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 444
 445             /* Calculate generalized born table index - this is a separate table from the normal one,
 446              * but we use the same procedure by multiplying r with scale and truncating to integer.
 447              */
 448             rt               = _mm256_mul_ps(r00,gbscale);
 449             gbitab           = _mm256_cvttps_epi32(rt);
 450             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 451             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 452             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 453             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 454             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 455             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 456             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 457                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 458             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 459                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 460             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 461                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 462             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 463                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 464             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 465             Heps             = _mm256_mul_ps(gbeps,H);
 466             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 467             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 468             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 469
 470             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 471             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 472             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 473             dvdatmp          = _mm256_andnot_ps(dummy_mask,dvdatmp);
 474             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 475             /* 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. */
 476             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 477             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 478             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 479             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 480             fjptrE             = (jnrlistE>=0) ? dvda+jnrE : scratch;
 481             fjptrF             = (jnrlistF>=0) ? dvda+jnrF : scratch;
 482             fjptrG             = (jnrlistG>=0) ? dvda+jnrG : scratch;
 483             fjptrH             = (jnrlistH>=0) ? dvda+jnrH : scratch;
 484             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 485                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 486             velec            = _mm256_mul_ps(qq00,rinv00);
 487             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 488
 489             /* LENNARD-JONES DISPERSION/REPULSION */
 490
 491             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 492             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 493             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 494             vvdw             = _mm256_sub_ps( _mm256_mul_ps(vvdw12,one_twelfth) , _mm256_mul_ps(vvdw6,one_sixth) );
 495             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 496
 497             /* Update potential sum for this i atom from the interaction with this j atom. */
 498             velec            = _mm256_andnot_ps(dummy_mask,velec);
 499             velecsum         = _mm256_add_ps(velecsum,velec);
 500             vgb              = _mm256_andnot_ps(dummy_mask,vgb);
 501             vgbsum           = _mm256_add_ps(vgbsum,vgb);
 502             vvdw             = _mm256_andnot_ps(dummy_mask,vvdw);
 503             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 504
 505             fscal            = _mm256_add_ps(felec,fvdw);
 506
 507             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 508
 509             /* Calculate temporary vectorial force */
 510             tx               = _mm256_mul_ps(fscal,dx00);
 511             ty               = _mm256_mul_ps(fscal,dy00);
 512             tz               = _mm256_mul_ps(fscal,dz00);
 513
 514             /* Update vectorial force */
 515             fix0             = _mm256_add_ps(fix0,tx);
 516             fiy0             = _mm256_add_ps(fiy0,ty);
 517             fiz0             = _mm256_add_ps(fiz0,tz);
 518
 519             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 520             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 521             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 522             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 523             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 524             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 525             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 526             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 527             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 528
 529             /* Inner loop uses 71 flops */
 530         }
 531
 532         /* End of innermost loop */
 533
 534         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 535                                                  f+i_coord_offset,fshift+i_shift_offset);
 536
 537         ggid                        = gid[iidx];
 538         /* Update potential energies */
 539         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 540         gmx_mm256_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 541         gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 542         dvdasum = _mm256_mul_ps(dvdasum, _mm256_mul_ps(isai0,isai0));
 543         gmx_mm256_update_1pot_ps(dvdasum,dvda+inr);
 544
 545         /* Increment number of inner iterations */
 546         inneriter                  += j_index_end - j_index_start;
 547
 548         /* Outer loop uses 10 flops */
 549     }
 550
 551     /* Increment number of outer iterations */
 552     outeriter        += nri;
 553
 554     /* Update outer/inner flops */
 555
 556     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*71);
 557 }
 558 /*
 559  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_single
 560  * Electrostatics interaction: GeneralizedBorn
 561  * VdW interaction:            LennardJones
 562  * Geometry:                   Particle-Particle
 563  * Calculate force/pot:        Force
 564  */
 565 void
 566 nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_single
 567                     (t_nblist                    * gmx_restrict       nlist,
 568                      rvec                        * gmx_restrict          xx,
 569                      rvec                        * gmx_restrict          ff,
 570                      struct t_forcerec           * gmx_restrict          fr,
 571                      t_mdatoms                   * gmx_restrict     mdatoms,
 572                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 573                      t_nrnb                      * gmx_restrict        nrnb)
 574 {
 575     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 576      * just 0 for non-waters.
 577      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 578      * jnr indices corresponding to data put in the four positions in the SIMD register.
 579      */
 580     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 581     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 582     int              jnrA,jnrB,jnrC,jnrD;
 583     int              jnrE,jnrF,jnrG,jnrH;
 584     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 585     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 586     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 587     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 588     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 589     real             rcutoff_scalar;
 590     real             *shiftvec,*fshift,*x,*f;
 591     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 592     real             scratch[4*DIM];
 593     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 594     real *           vdwioffsetptr0;
 595     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 596     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 597     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 598     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 599     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 600     real             *charge;
 601     __m256i          gbitab;
 602     __m128i          gbitab_lo,gbitab_hi;
 603     __m256           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 604     __m256           minushalf = _mm256_set1_ps(-0.5);
 605     real             *invsqrta,*dvda,*gbtab;
 606     int              nvdwtype;
 607     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 608     int              *vdwtype;
 609     real             *vdwparam;
 610     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 611     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 612     __m256i          vfitab;
 613     __m128i          vfitab_lo,vfitab_hi;
 614     __m128i          ifour       = _mm_set1_epi32(4);
 615     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 616     real             *vftab;
 617     __m256           dummy_mask,cutoff_mask;
 618     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 619     __m256           one     = _mm256_set1_ps(1.0);
 620     __m256           two     = _mm256_set1_ps(2.0);
 621     x                = xx[0];
 622     f                = ff[0];
 623
 624     nri              = nlist->nri;
 625     iinr             = nlist->iinr;
 626     jindex           = nlist->jindex;
 627     jjnr             = nlist->jjnr;
 628     shiftidx         = nlist->shift;
 629     gid              = nlist->gid;
 630     shiftvec         = fr->shift_vec[0];
 631     fshift           = fr->fshift[0];
 632     facel            = _mm256_set1_ps(fr->ic->epsfac);
 633     charge           = mdatoms->chargeA;
 634     nvdwtype         = fr->ntype;
 635     vdwparam         = fr->nbfp;
 636     vdwtype          = mdatoms->typeA;
 637
 638     invsqrta         = fr->invsqrta;
 639     dvda             = fr->dvda;
 640     gbtabscale       = _mm256_set1_ps(fr->gbtab->scale);
 641     gbtab            = fr->gbtab->data;
 642     gbinvepsdiff     = _mm256_set1_ps((1.0/fr->ic->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 643
 644     /* Avoid stupid compiler warnings */
 645     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 646     j_coord_offsetA = 0;
 647     j_coord_offsetB = 0;
 648     j_coord_offsetC = 0;
 649     j_coord_offsetD = 0;
 650     j_coord_offsetE = 0;
 651     j_coord_offsetF = 0;
 652     j_coord_offsetG = 0;
 653     j_coord_offsetH = 0;
 654
 655     outeriter        = 0;
 656     inneriter        = 0;
 657
 658     for(iidx=0;iidx<4*DIM;iidx++)
 659     {
 660         scratch[iidx] = 0.0;
 661     }
 662
 663     /* Start outer loop over neighborlists */
 664     for(iidx=0; iidx<nri; iidx++)
 665     {
 666         /* Load shift vector for this list */
 667         i_shift_offset   = DIM*shiftidx[iidx];
 668
 669         /* Load limits for loop over neighbors */
 670         j_index_start    = jindex[iidx];
 671         j_index_end      = jindex[iidx+1];
 672
 673         /* Get outer coordinate index */
 674         inr              = iinr[iidx];
 675         i_coord_offset   = DIM*inr;
 676
 677         /* Load i particle coords and add shift vector */
 678         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 679
 680         fix0             = _mm256_setzero_ps();
 681         fiy0             = _mm256_setzero_ps();
 682         fiz0             = _mm256_setzero_ps();
 683
 684         /* Load parameters for i particles */
 685         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 686         isai0            = _mm256_set1_ps(invsqrta[inr+0]);
 687         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 688
 689         dvdasum          = _mm256_setzero_ps();
 690
 691         /* Start inner kernel loop */
 692         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 693         {
 694
 695             /* Get j neighbor index, and coordinate index */
 696             jnrA             = jjnr[jidx];
 697             jnrB             = jjnr[jidx+1];
 698             jnrC             = jjnr[jidx+2];
 699             jnrD             = jjnr[jidx+3];
 700             jnrE             = jjnr[jidx+4];
 701             jnrF             = jjnr[jidx+5];
 702             jnrG             = jjnr[jidx+6];
 703             jnrH             = jjnr[jidx+7];
 704             j_coord_offsetA  = DIM*jnrA;
 705             j_coord_offsetB  = DIM*jnrB;
 706             j_coord_offsetC  = DIM*jnrC;
 707             j_coord_offsetD  = DIM*jnrD;
 708             j_coord_offsetE  = DIM*jnrE;
 709             j_coord_offsetF  = DIM*jnrF;
 710             j_coord_offsetG  = DIM*jnrG;
 711             j_coord_offsetH  = DIM*jnrH;
 712
 713             /* load j atom coordinates */
 714             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 715                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 716                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 717                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 718                                                  &jx0,&jy0,&jz0);
 719
 720             /* Calculate displacement vector */
 721             dx00             = _mm256_sub_ps(ix0,jx0);
 722             dy00             = _mm256_sub_ps(iy0,jy0);
 723             dz00             = _mm256_sub_ps(iz0,jz0);
 724
 725             /* Calculate squared distance and things based on it */
 726             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 727
 728             rinv00           = avx256_invsqrt_f(rsq00);
 729
 730             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 731
 732             /* Load parameters for j particles */
 733             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 734                                                                  charge+jnrC+0,charge+jnrD+0,
 735                                                                  charge+jnrE+0,charge+jnrF+0,
 736                                                                  charge+jnrG+0,charge+jnrH+0);
 737             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 738                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 739                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 740                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 741             vdwjidx0A        = 2*vdwtype[jnrA+0];
 742             vdwjidx0B        = 2*vdwtype[jnrB+0];
 743             vdwjidx0C        = 2*vdwtype[jnrC+0];
 744             vdwjidx0D        = 2*vdwtype[jnrD+0];
 745             vdwjidx0E        = 2*vdwtype[jnrE+0];
 746             vdwjidx0F        = 2*vdwtype[jnrF+0];
 747             vdwjidx0G        = 2*vdwtype[jnrG+0];
 748             vdwjidx0H        = 2*vdwtype[jnrH+0];
 749
 750             /**************************
 751              * CALCULATE INTERACTIONS *
 752              **************************/
 753
 754             r00              = _mm256_mul_ps(rsq00,rinv00);
 755
 756             /* Compute parameters for interactions between i and j atoms */
 757             qq00             = _mm256_mul_ps(iq0,jq0);
 758             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 759                                             vdwioffsetptr0+vdwjidx0B,
 760                                             vdwioffsetptr0+vdwjidx0C,
 761                                             vdwioffsetptr0+vdwjidx0D,
 762                                             vdwioffsetptr0+vdwjidx0E,
 763                                             vdwioffsetptr0+vdwjidx0F,
 764                                             vdwioffsetptr0+vdwjidx0G,
 765                                             vdwioffsetptr0+vdwjidx0H,
 766                                             &c6_00,&c12_00);
 767
 768             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 769             isaprod          = _mm256_mul_ps(isai0,isaj0);
 770             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 771             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 772
 773             /* Calculate generalized born table index - this is a separate table from the normal one,
 774              * but we use the same procedure by multiplying r with scale and truncating to integer.
 775              */
 776             rt               = _mm256_mul_ps(r00,gbscale);
 777             gbitab           = _mm256_cvttps_epi32(rt);
 778             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 779             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 780             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 781             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 782             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 783             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 784             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 785                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 786             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 787                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 788             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 789                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 790             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 791                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 792             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 793             Heps             = _mm256_mul_ps(gbeps,H);
 794             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 795             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 796             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 797
 798             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 799             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 800             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 801             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 802             fjptrA           = dvda+jnrA;
 803             fjptrB           = dvda+jnrB;
 804             fjptrC           = dvda+jnrC;
 805             fjptrD           = dvda+jnrD;
 806             fjptrE           = dvda+jnrE;
 807             fjptrF           = dvda+jnrF;
 808             fjptrG           = dvda+jnrG;
 809             fjptrH           = dvda+jnrH;
 810             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 811                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 812             velec            = _mm256_mul_ps(qq00,rinv00);
 813             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 814
 815             /* LENNARD-JONES DISPERSION/REPULSION */
 816
 817             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 818             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
 819
 820             fscal            = _mm256_add_ps(felec,fvdw);
 821
 822             /* Calculate temporary vectorial force */
 823             tx               = _mm256_mul_ps(fscal,dx00);
 824             ty               = _mm256_mul_ps(fscal,dy00);
 825             tz               = _mm256_mul_ps(fscal,dz00);
 826
 827             /* Update vectorial force */
 828             fix0             = _mm256_add_ps(fix0,tx);
 829             fiy0             = _mm256_add_ps(fiy0,ty);
 830             fiz0             = _mm256_add_ps(fiz0,tz);
 831
 832             fjptrA             = f+j_coord_offsetA;
 833             fjptrB             = f+j_coord_offsetB;
 834             fjptrC             = f+j_coord_offsetC;
 835             fjptrD             = f+j_coord_offsetD;
 836             fjptrE             = f+j_coord_offsetE;
 837             fjptrF             = f+j_coord_offsetF;
 838             fjptrG             = f+j_coord_offsetG;
 839             fjptrH             = f+j_coord_offsetH;
 840             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 841
 842             /* Inner loop uses 63 flops */
 843         }
 844
 845         if(jidx<j_index_end)
 846         {
 847
 848             /* Get j neighbor index, and coordinate index */
 849             jnrlistA         = jjnr[jidx];
 850             jnrlistB         = jjnr[jidx+1];
 851             jnrlistC         = jjnr[jidx+2];
 852             jnrlistD         = jjnr[jidx+3];
 853             jnrlistE         = jjnr[jidx+4];
 854             jnrlistF         = jjnr[jidx+5];
 855             jnrlistG         = jjnr[jidx+6];
 856             jnrlistH         = jjnr[jidx+7];
 857             /* Sign of each element will be negative for non-real atoms.
 858              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 859              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 860              */
 861             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 862                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 863
 864             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 865             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 866             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 867             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 868             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 869             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 870             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 871             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 872             j_coord_offsetA  = DIM*jnrA;
 873             j_coord_offsetB  = DIM*jnrB;
 874             j_coord_offsetC  = DIM*jnrC;
 875             j_coord_offsetD  = DIM*jnrD;
 876             j_coord_offsetE  = DIM*jnrE;
 877             j_coord_offsetF  = DIM*jnrF;
 878             j_coord_offsetG  = DIM*jnrG;
 879             j_coord_offsetH  = DIM*jnrH;
 880
 881             /* load j atom coordinates */
 882             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 883                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 884                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 885                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 886                                                  &jx0,&jy0,&jz0);
 887
 888             /* Calculate displacement vector */
 889             dx00             = _mm256_sub_ps(ix0,jx0);
 890             dy00             = _mm256_sub_ps(iy0,jy0);
 891             dz00             = _mm256_sub_ps(iz0,jz0);
 892
 893             /* Calculate squared distance and things based on it */
 894             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 895
 896             rinv00           = avx256_invsqrt_f(rsq00);
 897
 898             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 899
 900             /* Load parameters for j particles */
 901             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 902                                                                  charge+jnrC+0,charge+jnrD+0,
 903                                                                  charge+jnrE+0,charge+jnrF+0,
 904                                                                  charge+jnrG+0,charge+jnrH+0);
 905             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 906                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 907                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 908                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 909             vdwjidx0A        = 2*vdwtype[jnrA+0];
 910             vdwjidx0B        = 2*vdwtype[jnrB+0];
 911             vdwjidx0C        = 2*vdwtype[jnrC+0];
 912             vdwjidx0D        = 2*vdwtype[jnrD+0];
 913             vdwjidx0E        = 2*vdwtype[jnrE+0];
 914             vdwjidx0F        = 2*vdwtype[jnrF+0];
 915             vdwjidx0G        = 2*vdwtype[jnrG+0];
 916             vdwjidx0H        = 2*vdwtype[jnrH+0];
 917
 918             /**************************
 919              * CALCULATE INTERACTIONS *
 920              **************************/
 921
 922             r00              = _mm256_mul_ps(rsq00,rinv00);
 923             r00              = _mm256_andnot_ps(dummy_mask,r00);
 924
 925             /* Compute parameters for interactions between i and j atoms */
 926             qq00             = _mm256_mul_ps(iq0,jq0);
 927             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 928                                             vdwioffsetptr0+vdwjidx0B,
 929                                             vdwioffsetptr0+vdwjidx0C,
 930                                             vdwioffsetptr0+vdwjidx0D,
 931                                             vdwioffsetptr0+vdwjidx0E,
 932                                             vdwioffsetptr0+vdwjidx0F,
 933                                             vdwioffsetptr0+vdwjidx0G,
 934                                             vdwioffsetptr0+vdwjidx0H,
 935                                             &c6_00,&c12_00);
 936
 937             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 938             isaprod          = _mm256_mul_ps(isai0,isaj0);
 939             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 940             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 941
 942             /* Calculate generalized born table index - this is a separate table from the normal one,
 943              * but we use the same procedure by multiplying r with scale and truncating to integer.
 944              */
 945             rt               = _mm256_mul_ps(r00,gbscale);
 946             gbitab           = _mm256_cvttps_epi32(rt);
 947             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 948             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 949             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 950             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 951             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 952             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 953             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 954                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 955             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 956                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 957             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 958                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 959             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 960                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 961             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 962             Heps             = _mm256_mul_ps(gbeps,H);
 963             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 964             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 965             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 966
 967             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 968             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 969             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 970             dvdatmp          = _mm256_andnot_ps(dummy_mask,dvdatmp);
 971             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 972             /* 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. */
 973             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 974             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 975             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 976             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 977             fjptrE             = (jnrlistE>=0) ? dvda+jnrE : scratch;
 978             fjptrF             = (jnrlistF>=0) ? dvda+jnrF : scratch;
 979             fjptrG             = (jnrlistG>=0) ? dvda+jnrG : scratch;
 980             fjptrH             = (jnrlistH>=0) ? dvda+jnrH : scratch;
 981             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 982                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 983             velec            = _mm256_mul_ps(qq00,rinv00);
 984             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 985
 986             /* LENNARD-JONES DISPERSION/REPULSION */
 987
 988             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 989             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
 990
 991             fscal            = _mm256_add_ps(felec,fvdw);
 992
 993             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 994
 995             /* Calculate temporary vectorial force */
 996             tx               = _mm256_mul_ps(fscal,dx00);
 997             ty               = _mm256_mul_ps(fscal,dy00);
 998             tz               = _mm256_mul_ps(fscal,dz00);
 999
1000             /* Update vectorial force */
1001             fix0             = _mm256_add_ps(fix0,tx);
1002             fiy0             = _mm256_add_ps(fiy0,ty);
1003             fiz0             = _mm256_add_ps(fiz0,tz);
1004
1005             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1006             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1007             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1008             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1009             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
1010             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
1011             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
1012             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
1013             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
1014
1015             /* Inner loop uses 64 flops */
1016         }
1017
1018         /* End of innermost loop */
1019
1020         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
1021                                                  f+i_coord_offset,fshift+i_shift_offset);
1022
1023         dvdasum = _mm256_mul_ps(dvdasum, _mm256_mul_ps(isai0,isai0));
1024         gmx_mm256_update_1pot_ps(dvdasum,dvda+inr);
1025
1026         /* Increment number of inner iterations */
1027         inneriter                  += j_index_end - j_index_start;
1028
1029         /* Outer loop uses 7 flops */
1030     }
1031
1032     /* Increment number of outer iterations */
1033     outeriter        += nri;
1034
1035     /* Update outer/inner flops */
1036
1037     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*64);
1038 }