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, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_256_single kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_256_single.h"
  48 #include "kernelutil_x86_avx_256_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_256_single
  52  * Electrostatics interaction: GeneralizedBorn
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_256_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrE,jnrF,jnrG,jnrH;
  76     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  77     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  80     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  81     real             rcutoff_scalar;
  82     real             *shiftvec,*fshift,*x,*f;
  83     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  84     real             scratch[4*DIM];
  85     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  86     real *           vdwioffsetptr0;
  87     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  89     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  91     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
  92     real             *charge;
  93     __m256i          gbitab;
  94     __m128i          gbitab_lo,gbitab_hi;
  95     __m256           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
  96     __m256           minushalf = _mm256_set1_ps(-0.5);
  97     real             *invsqrta,*dvda,*gbtab;
  98     int              nvdwtype;
  99     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 100     int              *vdwtype;
 101     real             *vdwparam;
 102     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 103     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 104     __m256i          vfitab;
 105     __m128i          vfitab_lo,vfitab_hi;
 106     __m128i          ifour       = _mm_set1_epi32(4);
 107     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 108     real             *vftab;
 109     __m256           dummy_mask,cutoff_mask;
 110     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 111     __m256           one     = _mm256_set1_ps(1.0);
 112     __m256           two     = _mm256_set1_ps(2.0);
 113     x                = xx[0];
 114     f                = ff[0];
 115
 116     nri              = nlist->nri;
 117     iinr             = nlist->iinr;
 118     jindex           = nlist->jindex;
 119     jjnr             = nlist->jjnr;
 120     shiftidx         = nlist->shift;
 121     gid              = nlist->gid;
 122     shiftvec         = fr->shift_vec[0];
 123     fshift           = fr->fshift[0];
 124     facel            = _mm256_set1_ps(fr->epsfac);
 125     charge           = mdatoms->chargeA;
 126     nvdwtype         = fr->ntype;
 127     vdwparam         = fr->nbfp;
 128     vdwtype          = mdatoms->typeA;
 129
 130     invsqrta         = fr->invsqrta;
 131     dvda             = fr->dvda;
 132     gbtabscale       = _mm256_set1_ps(fr->gbtab.scale);
 133     gbtab            = fr->gbtab.data;
 134     gbinvepsdiff     = _mm256_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 135
 136     /* Avoid stupid compiler warnings */
 137     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 138     j_coord_offsetA = 0;
 139     j_coord_offsetB = 0;
 140     j_coord_offsetC = 0;
 141     j_coord_offsetD = 0;
 142     j_coord_offsetE = 0;
 143     j_coord_offsetF = 0;
 144     j_coord_offsetG = 0;
 145     j_coord_offsetH = 0;
 146
 147     outeriter        = 0;
 148     inneriter        = 0;
 149
 150     for(iidx=0;iidx<4*DIM;iidx++)
 151     {
 152         scratch[iidx] = 0.0;
 153     }
 154
 155     /* Start outer loop over neighborlists */
 156     for(iidx=0; iidx<nri; iidx++)
 157     {
 158         /* Load shift vector for this list */
 159         i_shift_offset   = DIM*shiftidx[iidx];
 160
 161         /* Load limits for loop over neighbors */
 162         j_index_start    = jindex[iidx];
 163         j_index_end      = jindex[iidx+1];
 164
 165         /* Get outer coordinate index */
 166         inr              = iinr[iidx];
 167         i_coord_offset   = DIM*inr;
 168
 169         /* Load i particle coords and add shift vector */
 170         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 171
 172         fix0             = _mm256_setzero_ps();
 173         fiy0             = _mm256_setzero_ps();
 174         fiz0             = _mm256_setzero_ps();
 175
 176         /* Load parameters for i particles */
 177         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 178         isai0            = _mm256_set1_ps(invsqrta[inr+0]);
 179         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 180
 181         /* Reset potential sums */
 182         velecsum         = _mm256_setzero_ps();
 183         vgbsum           = _mm256_setzero_ps();
 184         vvdwsum          = _mm256_setzero_ps();
 185         dvdasum          = _mm256_setzero_ps();
 186
 187         /* Start inner kernel loop */
 188         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 189         {
 190
 191             /* Get j neighbor index, and coordinate index */
 192             jnrA             = jjnr[jidx];
 193             jnrB             = jjnr[jidx+1];
 194             jnrC             = jjnr[jidx+2];
 195             jnrD             = jjnr[jidx+3];
 196             jnrE             = jjnr[jidx+4];
 197             jnrF             = jjnr[jidx+5];
 198             jnrG             = jjnr[jidx+6];
 199             jnrH             = jjnr[jidx+7];
 200             j_coord_offsetA  = DIM*jnrA;
 201             j_coord_offsetB  = DIM*jnrB;
 202             j_coord_offsetC  = DIM*jnrC;
 203             j_coord_offsetD  = DIM*jnrD;
 204             j_coord_offsetE  = DIM*jnrE;
 205             j_coord_offsetF  = DIM*jnrF;
 206             j_coord_offsetG  = DIM*jnrG;
 207             j_coord_offsetH  = DIM*jnrH;
 208
 209             /* load j atom coordinates */
 210             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 211                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 212                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 213                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 214                                                  &jx0,&jy0,&jz0);
 215
 216             /* Calculate displacement vector */
 217             dx00             = _mm256_sub_ps(ix0,jx0);
 218             dy00             = _mm256_sub_ps(iy0,jy0);
 219             dz00             = _mm256_sub_ps(iz0,jz0);
 220
 221             /* Calculate squared distance and things based on it */
 222             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 223
 224             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 225
 226             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 227
 228             /* Load parameters for j particles */
 229             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 230                                                                  charge+jnrC+0,charge+jnrD+0,
 231                                                                  charge+jnrE+0,charge+jnrF+0,
 232                                                                  charge+jnrG+0,charge+jnrH+0);
 233             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 234                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 235                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 236                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 237             vdwjidx0A        = 2*vdwtype[jnrA+0];
 238             vdwjidx0B        = 2*vdwtype[jnrB+0];
 239             vdwjidx0C        = 2*vdwtype[jnrC+0];
 240             vdwjidx0D        = 2*vdwtype[jnrD+0];
 241             vdwjidx0E        = 2*vdwtype[jnrE+0];
 242             vdwjidx0F        = 2*vdwtype[jnrF+0];
 243             vdwjidx0G        = 2*vdwtype[jnrG+0];
 244             vdwjidx0H        = 2*vdwtype[jnrH+0];
 245
 246             /**************************
 247              * CALCULATE INTERACTIONS *
 248              **************************/
 249
 250             r00              = _mm256_mul_ps(rsq00,rinv00);
 251
 252             /* Compute parameters for interactions between i and j atoms */
 253             qq00             = _mm256_mul_ps(iq0,jq0);
 254             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 255                                             vdwioffsetptr0+vdwjidx0B,
 256                                             vdwioffsetptr0+vdwjidx0C,
 257                                             vdwioffsetptr0+vdwjidx0D,
 258                                             vdwioffsetptr0+vdwjidx0E,
 259                                             vdwioffsetptr0+vdwjidx0F,
 260                                             vdwioffsetptr0+vdwjidx0G,
 261                                             vdwioffsetptr0+vdwjidx0H,
 262                                             &c6_00,&c12_00);
 263
 264             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 265             isaprod          = _mm256_mul_ps(isai0,isaj0);
 266             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 267             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 268
 269             /* Calculate generalized born table index - this is a separate table from the normal one,
 270              * but we use the same procedure by multiplying r with scale and truncating to integer.
 271              */
 272             rt               = _mm256_mul_ps(r00,gbscale);
 273             gbitab           = _mm256_cvttps_epi32(rt);
 274             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 275             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 276             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 277             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 278             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 279             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 280             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 281                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 282             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 283                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 284             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 285                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 286             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 287                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 288             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 289             Heps             = _mm256_mul_ps(gbeps,H);
 290             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 291             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 292             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 293
 294             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 295             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 296             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 297             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 298             fjptrA           = dvda+jnrA;
 299             fjptrB           = dvda+jnrB;
 300             fjptrC           = dvda+jnrC;
 301             fjptrD           = dvda+jnrD;
 302             fjptrE           = dvda+jnrE;
 303             fjptrF           = dvda+jnrF;
 304             fjptrG           = dvda+jnrG;
 305             fjptrH           = dvda+jnrH;
 306             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 307                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 308             velec            = _mm256_mul_ps(qq00,rinv00);
 309             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 310
 311             /* LENNARD-JONES DISPERSION/REPULSION */
 312
 313             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 314             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 315             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 316             vvdw             = _mm256_sub_ps( _mm256_mul_ps(vvdw12,one_twelfth) , _mm256_mul_ps(vvdw6,one_sixth) );
 317             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 318
 319             /* Update potential sum for this i atom from the interaction with this j atom. */
 320             velecsum         = _mm256_add_ps(velecsum,velec);
 321             vgbsum           = _mm256_add_ps(vgbsum,vgb);
 322             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 323
 324             fscal            = _mm256_add_ps(felec,fvdw);
 325
 326             /* Calculate temporary vectorial force */
 327             tx               = _mm256_mul_ps(fscal,dx00);
 328             ty               = _mm256_mul_ps(fscal,dy00);
 329             tz               = _mm256_mul_ps(fscal,dz00);
 330
 331             /* Update vectorial force */
 332             fix0             = _mm256_add_ps(fix0,tx);
 333             fiy0             = _mm256_add_ps(fiy0,ty);
 334             fiz0             = _mm256_add_ps(fiz0,tz);
 335
 336             fjptrA             = f+j_coord_offsetA;
 337             fjptrB             = f+j_coord_offsetB;
 338             fjptrC             = f+j_coord_offsetC;
 339             fjptrD             = f+j_coord_offsetD;
 340             fjptrE             = f+j_coord_offsetE;
 341             fjptrF             = f+j_coord_offsetF;
 342             fjptrG             = f+j_coord_offsetG;
 343             fjptrH             = f+j_coord_offsetH;
 344             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 345
 346             /* Inner loop uses 70 flops */
 347         }
 348
 349         if(jidx<j_index_end)
 350         {
 351
 352             /* Get j neighbor index, and coordinate index */
 353             jnrlistA         = jjnr[jidx];
 354             jnrlistB         = jjnr[jidx+1];
 355             jnrlistC         = jjnr[jidx+2];
 356             jnrlistD         = jjnr[jidx+3];
 357             jnrlistE         = jjnr[jidx+4];
 358             jnrlistF         = jjnr[jidx+5];
 359             jnrlistG         = jjnr[jidx+6];
 360             jnrlistH         = jjnr[jidx+7];
 361             /* Sign of each element will be negative for non-real atoms.
 362              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 363              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 364              */
 365             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 366                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 367
 368             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 369             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 370             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 371             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 372             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 373             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 374             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 375             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 376             j_coord_offsetA  = DIM*jnrA;
 377             j_coord_offsetB  = DIM*jnrB;
 378             j_coord_offsetC  = DIM*jnrC;
 379             j_coord_offsetD  = DIM*jnrD;
 380             j_coord_offsetE  = DIM*jnrE;
 381             j_coord_offsetF  = DIM*jnrF;
 382             j_coord_offsetG  = DIM*jnrG;
 383             j_coord_offsetH  = DIM*jnrH;
 384
 385             /* load j atom coordinates */
 386             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 387                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 388                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 389                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 390                                                  &jx0,&jy0,&jz0);
 391
 392             /* Calculate displacement vector */
 393             dx00             = _mm256_sub_ps(ix0,jx0);
 394             dy00             = _mm256_sub_ps(iy0,jy0);
 395             dz00             = _mm256_sub_ps(iz0,jz0);
 396
 397             /* Calculate squared distance and things based on it */
 398             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 399
 400             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 401
 402             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 403
 404             /* Load parameters for j particles */
 405             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 406                                                                  charge+jnrC+0,charge+jnrD+0,
 407                                                                  charge+jnrE+0,charge+jnrF+0,
 408                                                                  charge+jnrG+0,charge+jnrH+0);
 409             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 410                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 411                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 412                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 413             vdwjidx0A        = 2*vdwtype[jnrA+0];
 414             vdwjidx0B        = 2*vdwtype[jnrB+0];
 415             vdwjidx0C        = 2*vdwtype[jnrC+0];
 416             vdwjidx0D        = 2*vdwtype[jnrD+0];
 417             vdwjidx0E        = 2*vdwtype[jnrE+0];
 418             vdwjidx0F        = 2*vdwtype[jnrF+0];
 419             vdwjidx0G        = 2*vdwtype[jnrG+0];
 420             vdwjidx0H        = 2*vdwtype[jnrH+0];
 421
 422             /**************************
 423              * CALCULATE INTERACTIONS *
 424              **************************/
 425
 426             r00              = _mm256_mul_ps(rsq00,rinv00);
 427             r00              = _mm256_andnot_ps(dummy_mask,r00);
 428
 429             /* Compute parameters for interactions between i and j atoms */
 430             qq00             = _mm256_mul_ps(iq0,jq0);
 431             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 432                                             vdwioffsetptr0+vdwjidx0B,
 433                                             vdwioffsetptr0+vdwjidx0C,
 434                                             vdwioffsetptr0+vdwjidx0D,
 435                                             vdwioffsetptr0+vdwjidx0E,
 436                                             vdwioffsetptr0+vdwjidx0F,
 437                                             vdwioffsetptr0+vdwjidx0G,
 438                                             vdwioffsetptr0+vdwjidx0H,
 439                                             &c6_00,&c12_00);
 440
 441             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 442             isaprod          = _mm256_mul_ps(isai0,isaj0);
 443             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 444             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 445
 446             /* Calculate generalized born table index - this is a separate table from the normal one,
 447              * but we use the same procedure by multiplying r with scale and truncating to integer.
 448              */
 449             rt               = _mm256_mul_ps(r00,gbscale);
 450             gbitab           = _mm256_cvttps_epi32(rt);
 451             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 452             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 453             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 454             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 455             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 456             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 457             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 458                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 459             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 460                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 461             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 462                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 463             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 464                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 465             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 466             Heps             = _mm256_mul_ps(gbeps,H);
 467             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 468             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 469             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 470
 471             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 472             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 473             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 474             dvdatmp          = _mm256_andnot_ps(dummy_mask,dvdatmp);
 475             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 476             /* 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. */
 477             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 478             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 479             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 480             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 481             fjptrE             = (jnrlistE>=0) ? dvda+jnrE : scratch;
 482             fjptrF             = (jnrlistF>=0) ? dvda+jnrF : scratch;
 483             fjptrG             = (jnrlistG>=0) ? dvda+jnrG : scratch;
 484             fjptrH             = (jnrlistH>=0) ? dvda+jnrH : scratch;
 485             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 486                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 487             velec            = _mm256_mul_ps(qq00,rinv00);
 488             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 489
 490             /* LENNARD-JONES DISPERSION/REPULSION */
 491
 492             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 493             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 494             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 495             vvdw             = _mm256_sub_ps( _mm256_mul_ps(vvdw12,one_twelfth) , _mm256_mul_ps(vvdw6,one_sixth) );
 496             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 497
 498             /* Update potential sum for this i atom from the interaction with this j atom. */
 499             velec            = _mm256_andnot_ps(dummy_mask,velec);
 500             velecsum         = _mm256_add_ps(velecsum,velec);
 501             vgb              = _mm256_andnot_ps(dummy_mask,vgb);
 502             vgbsum           = _mm256_add_ps(vgbsum,vgb);
 503             vvdw             = _mm256_andnot_ps(dummy_mask,vvdw);
 504             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 505
 506             fscal            = _mm256_add_ps(felec,fvdw);
 507
 508             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 509
 510             /* Calculate temporary vectorial force */
 511             tx               = _mm256_mul_ps(fscal,dx00);
 512             ty               = _mm256_mul_ps(fscal,dy00);
 513             tz               = _mm256_mul_ps(fscal,dz00);
 514
 515             /* Update vectorial force */
 516             fix0             = _mm256_add_ps(fix0,tx);
 517             fiy0             = _mm256_add_ps(fiy0,ty);
 518             fiz0             = _mm256_add_ps(fiz0,tz);
 519
 520             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 521             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 522             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 523             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 524             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 525             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 526             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 527             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 528             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 529
 530             /* Inner loop uses 71 flops */
 531         }
 532
 533         /* End of innermost loop */
 534
 535         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 536                                                  f+i_coord_offset,fshift+i_shift_offset);
 537
 538         ggid                        = gid[iidx];
 539         /* Update potential energies */
 540         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 541         gmx_mm256_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 542         gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 543         dvdasum = _mm256_mul_ps(dvdasum, _mm256_mul_ps(isai0,isai0));
 544         gmx_mm256_update_1pot_ps(dvdasum,dvda+inr);
 545
 546         /* Increment number of inner iterations */
 547         inneriter                  += j_index_end - j_index_start;
 548
 549         /* Outer loop uses 10 flops */
 550     }
 551
 552     /* Increment number of outer iterations */
 553     outeriter        += nri;
 554
 555     /* Update outer/inner flops */
 556
 557     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*71);
 558 }
 559 /*
 560  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_single
 561  * Electrostatics interaction: GeneralizedBorn
 562  * VdW interaction:            LennardJones
 563  * Geometry:                   Particle-Particle
 564  * Calculate force/pot:        Force
 565  */
 566 void
 567 nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_single
 568                     (t_nblist                    * gmx_restrict       nlist,
 569                      rvec                        * gmx_restrict          xx,
 570                      rvec                        * gmx_restrict          ff,
 571                      t_forcerec                  * gmx_restrict          fr,
 572                      t_mdatoms                   * gmx_restrict     mdatoms,
 573                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 574                      t_nrnb                      * gmx_restrict        nrnb)
 575 {
 576     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 577      * just 0 for non-waters.
 578      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 579      * jnr indices corresponding to data put in the four positions in the SIMD register.
 580      */
 581     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 582     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 583     int              jnrA,jnrB,jnrC,jnrD;
 584     int              jnrE,jnrF,jnrG,jnrH;
 585     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 586     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 587     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 588     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 589     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 590     real             rcutoff_scalar;
 591     real             *shiftvec,*fshift,*x,*f;
 592     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 593     real             scratch[4*DIM];
 594     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 595     real *           vdwioffsetptr0;
 596     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 597     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 598     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 599     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 600     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 601     real             *charge;
 602     __m256i          gbitab;
 603     __m128i          gbitab_lo,gbitab_hi;
 604     __m256           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 605     __m256           minushalf = _mm256_set1_ps(-0.5);
 606     real             *invsqrta,*dvda,*gbtab;
 607     int              nvdwtype;
 608     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 609     int              *vdwtype;
 610     real             *vdwparam;
 611     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 612     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 613     __m256i          vfitab;
 614     __m128i          vfitab_lo,vfitab_hi;
 615     __m128i          ifour       = _mm_set1_epi32(4);
 616     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 617     real             *vftab;
 618     __m256           dummy_mask,cutoff_mask;
 619     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 620     __m256           one     = _mm256_set1_ps(1.0);
 621     __m256           two     = _mm256_set1_ps(2.0);
 622     x                = xx[0];
 623     f                = ff[0];
 624
 625     nri              = nlist->nri;
 626     iinr             = nlist->iinr;
 627     jindex           = nlist->jindex;
 628     jjnr             = nlist->jjnr;
 629     shiftidx         = nlist->shift;
 630     gid              = nlist->gid;
 631     shiftvec         = fr->shift_vec[0];
 632     fshift           = fr->fshift[0];
 633     facel            = _mm256_set1_ps(fr->epsfac);
 634     charge           = mdatoms->chargeA;
 635     nvdwtype         = fr->ntype;
 636     vdwparam         = fr->nbfp;
 637     vdwtype          = mdatoms->typeA;
 638
 639     invsqrta         = fr->invsqrta;
 640     dvda             = fr->dvda;
 641     gbtabscale       = _mm256_set1_ps(fr->gbtab.scale);
 642     gbtab            = fr->gbtab.data;
 643     gbinvepsdiff     = _mm256_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 644
 645     /* Avoid stupid compiler warnings */
 646     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 647     j_coord_offsetA = 0;
 648     j_coord_offsetB = 0;
 649     j_coord_offsetC = 0;
 650     j_coord_offsetD = 0;
 651     j_coord_offsetE = 0;
 652     j_coord_offsetF = 0;
 653     j_coord_offsetG = 0;
 654     j_coord_offsetH = 0;
 655
 656     outeriter        = 0;
 657     inneriter        = 0;
 658
 659     for(iidx=0;iidx<4*DIM;iidx++)
 660     {
 661         scratch[iidx] = 0.0;
 662     }
 663
 664     /* Start outer loop over neighborlists */
 665     for(iidx=0; iidx<nri; iidx++)
 666     {
 667         /* Load shift vector for this list */
 668         i_shift_offset   = DIM*shiftidx[iidx];
 669
 670         /* Load limits for loop over neighbors */
 671         j_index_start    = jindex[iidx];
 672         j_index_end      = jindex[iidx+1];
 673
 674         /* Get outer coordinate index */
 675         inr              = iinr[iidx];
 676         i_coord_offset   = DIM*inr;
 677
 678         /* Load i particle coords and add shift vector */
 679         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 680
 681         fix0             = _mm256_setzero_ps();
 682         fiy0             = _mm256_setzero_ps();
 683         fiz0             = _mm256_setzero_ps();
 684
 685         /* Load parameters for i particles */
 686         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 687         isai0            = _mm256_set1_ps(invsqrta[inr+0]);
 688         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 689
 690         dvdasum          = _mm256_setzero_ps();
 691
 692         /* Start inner kernel loop */
 693         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 694         {
 695
 696             /* Get j neighbor index, and coordinate index */
 697             jnrA             = jjnr[jidx];
 698             jnrB             = jjnr[jidx+1];
 699             jnrC             = jjnr[jidx+2];
 700             jnrD             = jjnr[jidx+3];
 701             jnrE             = jjnr[jidx+4];
 702             jnrF             = jjnr[jidx+5];
 703             jnrG             = jjnr[jidx+6];
 704             jnrH             = jjnr[jidx+7];
 705             j_coord_offsetA  = DIM*jnrA;
 706             j_coord_offsetB  = DIM*jnrB;
 707             j_coord_offsetC  = DIM*jnrC;
 708             j_coord_offsetD  = DIM*jnrD;
 709             j_coord_offsetE  = DIM*jnrE;
 710             j_coord_offsetF  = DIM*jnrF;
 711             j_coord_offsetG  = DIM*jnrG;
 712             j_coord_offsetH  = DIM*jnrH;
 713
 714             /* load j atom coordinates */
 715             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 716                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 717                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 718                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 719                                                  &jx0,&jy0,&jz0);
 720
 721             /* Calculate displacement vector */
 722             dx00             = _mm256_sub_ps(ix0,jx0);
 723             dy00             = _mm256_sub_ps(iy0,jy0);
 724             dz00             = _mm256_sub_ps(iz0,jz0);
 725
 726             /* Calculate squared distance and things based on it */
 727             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 728
 729             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 730
 731             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 732
 733             /* Load parameters for j particles */
 734             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 735                                                                  charge+jnrC+0,charge+jnrD+0,
 736                                                                  charge+jnrE+0,charge+jnrF+0,
 737                                                                  charge+jnrG+0,charge+jnrH+0);
 738             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 739                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 740                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 741                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 742             vdwjidx0A        = 2*vdwtype[jnrA+0];
 743             vdwjidx0B        = 2*vdwtype[jnrB+0];
 744             vdwjidx0C        = 2*vdwtype[jnrC+0];
 745             vdwjidx0D        = 2*vdwtype[jnrD+0];
 746             vdwjidx0E        = 2*vdwtype[jnrE+0];
 747             vdwjidx0F        = 2*vdwtype[jnrF+0];
 748             vdwjidx0G        = 2*vdwtype[jnrG+0];
 749             vdwjidx0H        = 2*vdwtype[jnrH+0];
 750
 751             /**************************
 752              * CALCULATE INTERACTIONS *
 753              **************************/
 754
 755             r00              = _mm256_mul_ps(rsq00,rinv00);
 756
 757             /* Compute parameters for interactions between i and j atoms */
 758             qq00             = _mm256_mul_ps(iq0,jq0);
 759             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 760                                             vdwioffsetptr0+vdwjidx0B,
 761                                             vdwioffsetptr0+vdwjidx0C,
 762                                             vdwioffsetptr0+vdwjidx0D,
 763                                             vdwioffsetptr0+vdwjidx0E,
 764                                             vdwioffsetptr0+vdwjidx0F,
 765                                             vdwioffsetptr0+vdwjidx0G,
 766                                             vdwioffsetptr0+vdwjidx0H,
 767                                             &c6_00,&c12_00);
 768
 769             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 770             isaprod          = _mm256_mul_ps(isai0,isaj0);
 771             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 772             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 773
 774             /* Calculate generalized born table index - this is a separate table from the normal one,
 775              * but we use the same procedure by multiplying r with scale and truncating to integer.
 776              */
 777             rt               = _mm256_mul_ps(r00,gbscale);
 778             gbitab           = _mm256_cvttps_epi32(rt);
 779             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 780             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 781             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 782             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 783             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 784             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 785             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 786                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 787             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 788                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 789             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 790                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 791             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 792                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 793             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 794             Heps             = _mm256_mul_ps(gbeps,H);
 795             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 796             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 797             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 798
 799             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 800             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 801             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 802             dvdasum          = _mm256_add_ps(dvdasum,dvdatmp);
 803             fjptrA           = dvda+jnrA;
 804             fjptrB           = dvda+jnrB;
 805             fjptrC           = dvda+jnrC;
 806             fjptrD           = dvda+jnrD;
 807             fjptrE           = dvda+jnrE;
 808             fjptrF           = dvda+jnrF;
 809             fjptrG           = dvda+jnrG;
 810             fjptrH           = dvda+jnrH;
 811             gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
 812                                                  _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
 813             velec            = _mm256_mul_ps(qq00,rinv00);
 814             felec            = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
 815
 816             /* LENNARD-JONES DISPERSION/REPULSION */
 817
 818             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 819             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
 820
 821             fscal            = _mm256_add_ps(felec,fvdw);
 822
 823             /* Calculate temporary vectorial force */
 824             tx               = _mm256_mul_ps(fscal,dx00);
 825             ty               = _mm256_mul_ps(fscal,dy00);
 826             tz               = _mm256_mul_ps(fscal,dz00);
 827
 828             /* Update vectorial force */
 829             fix0             = _mm256_add_ps(fix0,tx);
 830             fiy0             = _mm256_add_ps(fiy0,ty);
 831             fiz0             = _mm256_add_ps(fiz0,tz);
 832
 833             fjptrA             = f+j_coord_offsetA;
 834             fjptrB             = f+j_coord_offsetB;
 835             fjptrC             = f+j_coord_offsetC;
 836             fjptrD             = f+j_coord_offsetD;
 837             fjptrE             = f+j_coord_offsetE;
 838             fjptrF             = f+j_coord_offsetF;
 839             fjptrG             = f+j_coord_offsetG;
 840             fjptrH             = f+j_coord_offsetH;
 841             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 842
 843             /* Inner loop uses 63 flops */
 844         }
 845
 846         if(jidx<j_index_end)
 847         {
 848
 849             /* Get j neighbor index, and coordinate index */
 850             jnrlistA         = jjnr[jidx];
 851             jnrlistB         = jjnr[jidx+1];
 852             jnrlistC         = jjnr[jidx+2];
 853             jnrlistD         = jjnr[jidx+3];
 854             jnrlistE         = jjnr[jidx+4];
 855             jnrlistF         = jjnr[jidx+5];
 856             jnrlistG         = jjnr[jidx+6];
 857             jnrlistH         = jjnr[jidx+7];
 858             /* Sign of each element will be negative for non-real atoms.
 859              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 860              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 861              */
 862             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 863                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 864
 865             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 866             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 867             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 868             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 869             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 870             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 871             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 872             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 873             j_coord_offsetA  = DIM*jnrA;
 874             j_coord_offsetB  = DIM*jnrB;
 875             j_coord_offsetC  = DIM*jnrC;
 876             j_coord_offsetD  = DIM*jnrD;
 877             j_coord_offsetE  = DIM*jnrE;
 878             j_coord_offsetF  = DIM*jnrF;
 879             j_coord_offsetG  = DIM*jnrG;
 880             j_coord_offsetH  = DIM*jnrH;
 881
 882             /* load j atom coordinates */
 883             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 884                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 885                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 886                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 887                                                  &jx0,&jy0,&jz0);
 888
 889             /* Calculate displacement vector */
 890             dx00             = _mm256_sub_ps(ix0,jx0);
 891             dy00             = _mm256_sub_ps(iy0,jy0);
 892             dz00             = _mm256_sub_ps(iz0,jz0);
 893
 894             /* Calculate squared distance and things based on it */
 895             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 896
 897             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 898
 899             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 900
 901             /* Load parameters for j particles */
 902             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 903                                                                  charge+jnrC+0,charge+jnrD+0,
 904                                                                  charge+jnrE+0,charge+jnrF+0,
 905                                                                  charge+jnrG+0,charge+jnrH+0);
 906             isaj0            = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 907                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0,
 908                                                                  invsqrta+jnrE+0,invsqrta+jnrF+0,
 909                                                                  invsqrta+jnrG+0,invsqrta+jnrH+0);
 910             vdwjidx0A        = 2*vdwtype[jnrA+0];
 911             vdwjidx0B        = 2*vdwtype[jnrB+0];
 912             vdwjidx0C        = 2*vdwtype[jnrC+0];
 913             vdwjidx0D        = 2*vdwtype[jnrD+0];
 914             vdwjidx0E        = 2*vdwtype[jnrE+0];
 915             vdwjidx0F        = 2*vdwtype[jnrF+0];
 916             vdwjidx0G        = 2*vdwtype[jnrG+0];
 917             vdwjidx0H        = 2*vdwtype[jnrH+0];
 918
 919             /**************************
 920              * CALCULATE INTERACTIONS *
 921              **************************/
 922
 923             r00              = _mm256_mul_ps(rsq00,rinv00);
 924             r00              = _mm256_andnot_ps(dummy_mask,r00);
 925
 926             /* Compute parameters for interactions between i and j atoms */
 927             qq00             = _mm256_mul_ps(iq0,jq0);
 928             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 929                                             vdwioffsetptr0+vdwjidx0B,
 930                                             vdwioffsetptr0+vdwjidx0C,
 931                                             vdwioffsetptr0+vdwjidx0D,
 932                                             vdwioffsetptr0+vdwjidx0E,
 933                                             vdwioffsetptr0+vdwjidx0F,
 934                                             vdwioffsetptr0+vdwjidx0G,
 935                                             vdwioffsetptr0+vdwjidx0H,
 936                                             &c6_00,&c12_00);
 937
 938             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 939             isaprod          = _mm256_mul_ps(isai0,isaj0);
 940             gbqqfactor       = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
 941             gbscale          = _mm256_mul_ps(isaprod,gbtabscale);
 942
 943             /* Calculate generalized born table index - this is a separate table from the normal one,
 944              * but we use the same procedure by multiplying r with scale and truncating to integer.
 945              */
 946             rt               = _mm256_mul_ps(r00,gbscale);
 947             gbitab           = _mm256_cvttps_epi32(rt);
 948             gbeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 949             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 950             gbitab_lo        = _mm256_extractf128_si256(gbitab,0x0);
 951             gbitab_hi        = _mm256_extractf128_si256(gbitab,0x1);
 952             gbitab_lo        = _mm_slli_epi32(gbitab_lo,2);
 953             gbitab_hi        = _mm_slli_epi32(gbitab_hi,2);
 954             Y                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
 955                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
 956             F                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
 957                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
 958             G                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
 959                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
 960             H                = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
 961                                                   _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
 962             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 963             Heps             = _mm256_mul_ps(gbeps,H);
 964             Fp               = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
 965             VV               = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
 966             vgb              = _mm256_mul_ps(gbqqfactor,VV);
 967
 968             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 969             fgb              = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
 970             dvdatmp          = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
 971             dvdatmp          = _mm256_andnot_ps(dummy_mask,dvdatmp);
 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 }