src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_template_avx_128_fma_single.pre

   1 /* #if 0 */
   2 /*
   3  * This file is part of the GROMACS molecular simulation package.
   4  *
   5  * Copyright (c) 2012,2013,2014,2015,2017, by the GROMACS development team, led by
   6  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   7  * and including many others, as listed in the AUTHORS file in the
   8  * top-level source directory and at http://www.gromacs.org.
   9  *
  10  * GROMACS is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public License
  12  * as published by the Free Software Foundation; either version 2.1
  13  * of the License, or (at your option) any later version.
  14  *
  15  * GROMACS is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with GROMACS; if not, see
  22  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  23  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  24  *
  25  * If you want to redistribute modifications to GROMACS, please
  26  * consider that scientific software is very special. Version
  27  * control is crucial - bugs must be traceable. We will be happy to
  28  * consider code for inclusion in the official distribution, but
  29  * derived work must not be called official GROMACS. Details are found
  30  * in the README & COPYING files - if they are missing, get the
  31  * official version at http://www.gromacs.org.
  32  *
  33  * To help us fund GROMACS development, we humbly ask that you cite
  34  * the research papers on the package. Check out http://www.gromacs.org.
  35  */
  36 #error This file must be processed with the Gromacs pre-preprocessor
  37 /* #endif */
  38 /* #if INCLUDE_HEADER */
  39 #include "gmxpre.h"
  40
  41 #include "config.h"
  42
  43 #include <math.h>
  44
  45 #include "../nb_kernel.h"
  46 #include "gromacs/gmxlib/nrnb.h"
  47
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49 /* #endif */
  50
  51 /* ## List of variables set by the generating script:                                    */
  52 /* ##                                                                                    */
  53 /* ## Setttings that apply to the entire kernel:                                         */
  54 /* ## KERNEL_ELEC:           String, choice for electrostatic interactions               */
  55 /* ## KERNEL_VDW:            String, choice for van der Waals interactions               */
  56 /* ## KERNEL_NAME:           String, name of this kernel                                 */
  57 /* ## KERNEL_VF:             String telling if we calculate potential, force, or both    */
  58 /* ## GEOMETRY_I/GEOMETRY_J: String, name of each geometry, e.g. 'Water3' or '1Particle' */
  59 /* ##                                                                                    */
  60 /* ## Setttings that apply to particles in the outer (I) or inner (J) loops:             */
  61 /* ## PARTICLES_I[]/         Arrays with lists of i/j particles to use in kernel. It is  */
  62 /* ## PARTICLES_J[]:         just [0] for particle geometry, but can be longer for water */
  63 /* ## PARTICLES_ELEC_I[]/    Arrays with lists of i/j particle that have electrostatics  */
  64 /* ## PARTICLES_ELEC_J[]:    interactions that should be calculated in this kernel.      */
  65 /* ## PARTICLES_VDW_I[]/     Arrays with the list of i/j particle that have VdW          */
  66 /* ## PARTICLES_VDW_J[]:     interactions that should be calculated in this kernel.      */
  67 /* ##                                                                                    */
  68 /* ## Setttings for pairs of interactions (e.g. 2nd i particle against 1st j particle)   */
  69 /* ## PAIRS_IJ[]:            Array with (i,j) tuples of pairs for which interactions     */
  70 /* ##                        should be calculated in this kernel. Zero-charge particles  */
  71 /* ##                        do not have interactions with particles without vdw, and    */
  72 /* ##                        Vdw-only interactions are not evaluated in a no-vdw-kernel. */
  73 /* ## INTERACTION_FLAGS[][]: 2D matrix, dimension e.g. 3*3 for water-water interactions. */
  74 /* ##                        For each i-j pair, the element [I][J] is a list of strings  */
  75 /* ##                        defining properties/flags of this interaction. Examples     */
  76 /* ##                        include 'electrostatics'/'vdw' if that type of interaction  */
  77 /* ##                        should be evaluated, 'rsq'/'rinv'/'rinvsq' if those values  */
  78 /* ##                        are needed, and 'exactcutoff' or 'shift','switch' to        */
  79 /* ##                        decide if the force/potential should be modified. This way  */
  80 /* ##                        we only calculate values absolutely needed for each case.   */
  81
  82 /* ## Calculate the size and offset for (merged/interleaved) table data */
  83
  84 /*
  85  * Gromacs nonbonded kernel:   {KERNEL_NAME}
  86  * Electrostatics interaction: {KERNEL_ELEC}
  87  * VdW interaction:            {KERNEL_VDW}
  88  * Geometry:                   {GEOMETRY_I}-{GEOMETRY_J}
  89  * Calculate force/pot:        {KERNEL_VF}
  90  */
  91 void
  92 {KERNEL_NAME}
  93                     (t_nblist                    * gmx_restrict       nlist,
  94                      rvec                        * gmx_restrict          xx,
  95                      rvec                        * gmx_restrict          ff,
  96                      struct t_forcerec           * gmx_restrict          fr,
  97                      t_mdatoms                   * gmx_restrict     mdatoms,
  98                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  99                      t_nrnb                      * gmx_restrict        nrnb)
 100 {
 101     /* ## Not all variables are used for all kernels, but any optimizing compiler fixes that, */
 102     /* ## so there is no point in going to extremes to exclude variables that are not needed. */
 103     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 104      * just 0 for non-waters.
 105      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 106      * jnr indices corresponding to data put in the four positions in the SIMD register.
 107      */
 108     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 109     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 110     int              jnrA,jnrB,jnrC,jnrD;
 111     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 112     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 113     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 114     real             rcutoff_scalar;
 115     real             *shiftvec,*fshift,*x,*f;
 116     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 117     real             scratch[4*DIM];
 118     __m128           fscal,rcutoff,rcutoff2,jidxall;
 119     /* #for I in PARTICLES_I */
 120     int              vdwioffset{I};
 121     __m128           ix{I},iy{I},iz{I},fix{I},fiy{I},fiz{I},iq{I},isai{I};
 122     /* #endfor */
 123     /* #for J in PARTICLES_J */
 124     int              vdwjidx{J}A,vdwjidx{J}B,vdwjidx{J}C,vdwjidx{J}D;
 125     __m128           jx{J},jy{J},jz{J},fjx{J},fjy{J},fjz{J},jq{J},isaj{J};
 126     /* #endfor */
 127     /* #for I,J in PAIRS_IJ */
 128     __m128           dx{I}{J},dy{I}{J},dz{I}{J},rsq{I}{J},rinv{I}{J},rinvsq{I}{J},r{I}{J},qq{I}{J},c6_{I}{J},c12_{I}{J};
 129     /* #endfor */
 130     /* #if KERNEL_ELEC != 'None' */
 131     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 132     real             *charge;
 133     /* #endif */
 134     /* #if 'GeneralizedBorn' in KERNEL_ELEC */
 135     __m128i          gbitab;
 136     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
 137     __m128           minushalf = _mm_set1_ps(-0.5);
 138     real             *invsqrta,*dvda,*gbtab;
 139     /* #endif */
 140     /* #if KERNEL_VDW != 'None' */
 141     int              nvdwtype;
 142     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 143     int              *vdwtype;
 144     real             *vdwparam;
 145     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 146     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 147     /* #endif */
 148     /* #if 'Table' in KERNEL_ELEC or 'GeneralizedBorn' in KERNEL_ELEC or 'Table' in KERNEL_VDW */
 149     __m128i          vfitab;
 150     __m128i          ifour       = _mm_set1_epi32(4);
 151     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 152     real             *vftab;
 153     /* #endif */
 154     /* #if 'LJEwald' in KERNEL_VDW */
 155     /* #for I,J in PAIRS_IJ */
 156     __m128           c6grid_{I}{J};
 157     /* #endfor */
 158     real             *vdwgridparam;
 159     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 160     __m128           one_half = _mm_set1_ps(0.5);
 161     __m128           minus_one = _mm_set1_ps(-1.0);
 162     /* #endif */
 163     /* #if 'Ewald' in KERNEL_ELEC */
 164     __m128i          ewitab;
 165     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 166     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 167     real             *ewtab;
 168     /* #endif */
 169     /* #if 'PotentialSwitch' in [KERNEL_MOD_ELEC,KERNEL_MOD_VDW] */
 170     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 171     real             rswitch_scalar,d_scalar;
 172     /* #endif */
 173     __m128           dummy_mask,cutoff_mask;
 174     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 175     __m128           one     = _mm_set1_ps(1.0);
 176     __m128           two     = _mm_set1_ps(2.0);
 177     x                = xx[0];
 178     f                = ff[0];
 179
 180     nri              = nlist->nri;
 181     iinr             = nlist->iinr;
 182     jindex           = nlist->jindex;
 183     jjnr             = nlist->jjnr;
 184     shiftidx         = nlist->shift;
 185     gid              = nlist->gid;
 186     shiftvec         = fr->shift_vec[0];
 187     fshift           = fr->fshift[0];
 188     /* #if KERNEL_ELEC != 'None' */
 189     facel            = _mm_set1_ps(fr->ic->epsfac);
 190     charge           = mdatoms->chargeA;
 191     /*     #if 'ReactionField' in KERNEL_ELEC */
 192     krf              = _mm_set1_ps(fr->ic->k_rf);
 193     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 194     crf              = _mm_set1_ps(fr->ic->c_rf);
 195     /*     #endif */
 196     /* #endif */
 197     /* #if KERNEL_VDW != 'None' */
 198     nvdwtype         = fr->ntype;
 199     vdwparam         = fr->nbfp;
 200     vdwtype          = mdatoms->typeA;
 201     /* #endif */
 202     /* #if 'LJEwald' in KERNEL_VDW */
 203     vdwgridparam     = fr->ljpme_c6grid;
 204     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 205     ewclj            = _mm_set1_ps(fr->ic->ewaldcoeff_lj);
 206     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 207     /* #endif */
 208
 209     /* #if 'Table' in KERNEL_ELEC and 'Table' in KERNEL_VDW */
 210     vftab            = kernel_data->table_elec_vdw->data;
 211     vftabscale       = _mm_set1_ps(kernel_data->table_elec_vdw->scale);
 212     /* #elif 'Table' in KERNEL_ELEC */
 213     vftab            = kernel_data->table_elec->data;
 214     vftabscale       = _mm_set1_ps(kernel_data->table_elec->scale);
 215     /* #elif 'Table' in KERNEL_VDW */
 216     vftab            = kernel_data->table_vdw->data;
 217     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 218     /* #endif */
 219
 220     /* #if 'Ewald' in KERNEL_ELEC */
 221     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 222     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 223     beta2            = _mm_mul_ps(beta,beta);
 224     beta3            = _mm_mul_ps(beta,beta2);
 225     /*     #if KERNEL_VF=='Force' and KERNEL_MOD_ELEC!='PotentialSwitch' */
 226     ewtab            = fr->ic->tabq_coul_F;
 227     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 228     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 229     /*     #else */
 230     ewtab            = fr->ic->tabq_coul_FDV0;
 231     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 232     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 233      /*     #endif */
 234     /* #endif */
 235
 236     /* #if KERNEL_ELEC=='GeneralizedBorn' */
 237     invsqrta         = fr->invsqrta;
 238     dvda             = fr->dvda;
 239     gbtabscale       = _mm_set1_ps(fr->gbtab->scale);
 240     gbtab            = fr->gbtab->data;
 241     gbinvepsdiff     = _mm_set1_ps((1.0/fr->ic->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 242     /* #endif */
 243
 244     /* #if 'Water' in GEOMETRY_I */
 245     /* Setup water-specific parameters */
 246     inr              = nlist->iinr[0];
 247     /*     #for I in PARTICLES_ELEC_I */
 248     iq{I}              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+{I}]));
 249     /*     #endfor */
 250     /*     #for I in PARTICLES_VDW_I */
 251     vdwioffset{I}      = 2*nvdwtype*vdwtype[inr+{I}];
 252     /*     #endfor */
 253     /* #endif */
 254
 255     /* #if 'Water' in GEOMETRY_J */
 256     /*     #for J in PARTICLES_ELEC_J */
 257     jq{J}              = _mm_set1_ps(charge[inr+{J}]);
 258     /*     #endfor */
 259     /*     #for J in PARTICLES_VDW_J */
 260     vdwjidx{J}A        = 2*vdwtype[inr+{J}];
 261     /*     #endfor */
 262     /*     #for I,J in PAIRS_IJ */
 263     /*         #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
 264     qq{I}{J}             = _mm_mul_ps(iq{I},jq{J});
 265     /*         #endif */
 266     /*         #if 'vdw' in INTERACTION_FLAGS[I][J] */
 267     /*             #if 'LJEwald' in KERNEL_VDW */
 268     c6_{I}{J}            = _mm_set1_ps(vdwparam[vdwioffset{I}+vdwjidx{J}A]);
 269     c12_{I}{J}           = _mm_set1_ps(vdwparam[vdwioffset{I}+vdwjidx{J}A+1]);
 270     c6grid_{I}{J}        = _mm_set1_ps(vdwgridparam[vdwioffset{I}+vdwjidx{J}A]);
 271     /*             #else */
 272     c6_{I}{J}            = _mm_set1_ps(vdwparam[vdwioffset{I}+vdwjidx{J}A]);
 273     c12_{I}{J}           = _mm_set1_ps(vdwparam[vdwioffset{I}+vdwjidx{J}A+1]);
 274     /*             #endif */
 275     /*         #endif */
 276     /*     #endfor */
 277     /* #endif */
 278
 279     /* #if KERNEL_MOD_ELEC!='None' or KERNEL_MOD_VDW!='None' */
 280     /*     #if KERNEL_ELEC!='None' */
 281     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 282     rcutoff_scalar   = fr->ic->rcoulomb;
 283     /*     #else */
 284     rcutoff_scalar   = fr->ic->rvdw;
 285     /*     #endif */
 286     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 287     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 288     /* #endif */
 289
 290     /* #if KERNEL_MOD_VDW=='PotentialShift' */
 291     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 292     rvdw             = _mm_set1_ps(fr->ic->rvdw);
 293     /* #endif */
 294
 295     /* #if 'PotentialSwitch' in [KERNEL_MOD_ELEC,KERNEL_MOD_VDW] */
 296     /*     #if KERNEL_MOD_ELEC=='PotentialSwitch'  */
 297     rswitch_scalar   = fr->ic->rcoulomb_switch;
 298     rswitch          = _mm_set1_ps(rswitch_scalar);
 299     /*     #else */
 300     rswitch_scalar   = fr->ic->rvdw_switch;
 301     rswitch          = _mm_set1_ps(rswitch_scalar);
 302     /*     #endif */
 303     /* Setup switch parameters */
 304     d_scalar         = rcutoff_scalar-rswitch_scalar;
 305     d                = _mm_set1_ps(d_scalar);
 306     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 307     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 308     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 309     /*     #if 'Force' in KERNEL_VF */
 310     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 311     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 312     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 313     /*     #endif */
 314     /* #endif */
 315
 316     /* Avoid stupid compiler warnings */
 317     jnrA = jnrB = jnrC = jnrD = 0;
 318     j_coord_offsetA = 0;
 319     j_coord_offsetB = 0;
 320     j_coord_offsetC = 0;
 321     j_coord_offsetD = 0;
 322
 323     /* ## Keep track of the floating point operations we issue for reporting! */
 324     /* #define OUTERFLOPS 0 */
 325     outeriter        = 0;
 326     inneriter        = 0;
 327
 328     for(iidx=0;iidx<4*DIM;iidx++)
 329     {
 330         scratch[iidx] = 0.0;
 331     }
 332
 333     /* Start outer loop over neighborlists */
 334     for(iidx=0; iidx<nri; iidx++)
 335     {
 336         /* Load shift vector for this list */
 337         i_shift_offset   = DIM*shiftidx[iidx];
 338
 339         /* Load limits for loop over neighbors */
 340         j_index_start    = jindex[iidx];
 341         j_index_end      = jindex[iidx+1];
 342
 343         /* Get outer coordinate index */
 344         inr              = iinr[iidx];
 345         i_coord_offset   = DIM*inr;
 346
 347         /* Load i particle coords and add shift vector */
 348         /* #if GEOMETRY_I == 'Particle' */
 349         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 350         /* #elif GEOMETRY_I == 'Water3' */
 351         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 352                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 353         /* #elif GEOMETRY_I == 'Water4' */
 354         /*     #if 0 in PARTICLES_I                 */
 355         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 356                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 357         /*     #else                                */
 358         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 359                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 360         /*     #endif                               */
 361         /* #endif                                   */
 362
 363         /* #if 'Force' in KERNEL_VF */
 364         /*     #for I in PARTICLES_I */
 365         fix{I}             = _mm_setzero_ps();
 366         fiy{I}             = _mm_setzero_ps();
 367         fiz{I}             = _mm_setzero_ps();
 368         /*     #endfor */
 369         /* #endif */
 370
 371         /* ## For water we already preloaded parameters at the start of the kernel */
 372         /* #if not 'Water' in GEOMETRY_I */
 373         /* Load parameters for i particles */
 374         /*     #for I in PARTICLES_ELEC_I */
 375         iq{I}              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+{I}));
 376         /*         #define OUTERFLOPS OUTERFLOPS+1 */
 377         /*         #if KERNEL_ELEC=='GeneralizedBorn' */
 378         isai{I}            = _mm_load1_ps(invsqrta+inr+{I});
 379         /*         #endif */
 380         /*     #endfor */
 381         /*     #for I in PARTICLES_VDW_I */
 382         vdwioffset{I}      = 2*nvdwtype*vdwtype[inr+{I}];
 383         /*     #endfor */
 384         /* #endif */
 385
 386         /* #if 'Potential' in KERNEL_VF */
 387         /* Reset potential sums */
 388         /*     #if KERNEL_ELEC != 'None' */
 389         velecsum         = _mm_setzero_ps();
 390         /*     #endif */
 391         /*     #if 'GeneralizedBorn' in KERNEL_ELEC */
 392         vgbsum           = _mm_setzero_ps();
 393         /*     #endif */
 394         /*     #if KERNEL_VDW != 'None' */
 395         vvdwsum          = _mm_setzero_ps();
 396         /*     #endif */
 397         /* #endif */
 398         /*     #if 'GeneralizedBorn' in KERNEL_ELEC and 'Force' in KERNEL_VF */
 399         dvdasum          = _mm_setzero_ps();
 400         /*     #endif */
 401
 402         /* #for ROUND in ['Loop','Epilogue'] */
 403
 404         /* #if ROUND =='Loop' */
 405         /* Start inner kernel loop */
 406         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 407         {
 408         /* ## First round is normal loop (next statement resets indentation) */
 409         /*     #if 0 */
 410         }
 411         /*     #endif */
 412         /* #else */
 413         if(jidx<j_index_end)
 414         {
 415         /* ## Second round is epilogue */
 416         /* #endif */
 417         /* #define INNERFLOPS 0 */
 418
 419             /* Get j neighbor index, and coordinate index */
 420             /* #if ROUND =='Loop' */
 421             jnrA             = jjnr[jidx];
 422             jnrB             = jjnr[jidx+1];
 423             jnrC             = jjnr[jidx+2];
 424             jnrD             = jjnr[jidx+3];
 425             /* #else */
 426             jnrlistA         = jjnr[jidx];
 427             jnrlistB         = jjnr[jidx+1];
 428             jnrlistC         = jjnr[jidx+2];
 429             jnrlistD         = jjnr[jidx+3];
 430             /* Sign of each element will be negative for non-real atoms.
 431              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 432              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 433              */
 434             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 435             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 436             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 437             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 438             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 439             /* #endif */
 440             j_coord_offsetA  = DIM*jnrA;
 441             j_coord_offsetB  = DIM*jnrB;
 442             j_coord_offsetC  = DIM*jnrC;
 443             j_coord_offsetD  = DIM*jnrD;
 444
 445             /* load j atom coordinates */
 446             /* #if GEOMETRY_J == 'Particle'             */
 447             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 448                                               x+j_coord_offsetC,x+j_coord_offsetD,
 449                                               &jx0,&jy0,&jz0);
 450             /* #elif GEOMETRY_J == 'Water3'             */
 451             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 452                                               x+j_coord_offsetC,x+j_coord_offsetD,
 453                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 454             /* #elif GEOMETRY_J == 'Water4'             */
 455             /*     #if 0 in PARTICLES_J                 */
 456             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 457                                               x+j_coord_offsetC,x+j_coord_offsetD,
 458                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 459                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 460             /*     #else                                */
 461             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 462                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 463                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 464             /*     #endif                               */
 465             /* #endif                                   */
 466
 467             /* Calculate displacement vector */
 468             /* #for I,J in PAIRS_IJ */
 469             dx{I}{J}             = _mm_sub_ps(ix{I},jx{J});
 470             dy{I}{J}             = _mm_sub_ps(iy{I},jy{J});
 471             dz{I}{J}             = _mm_sub_ps(iz{I},jz{J});
 472             /*     #define INNERFLOPS INNERFLOPS+3 */
 473             /* #endfor */
 474
 475             /* Calculate squared distance and things based on it */
 476             /* #for I,J in PAIRS_IJ */
 477             rsq{I}{J}            = gmx_mm_calc_rsq_ps(dx{I}{J},dy{I}{J},dz{I}{J});
 478             /*     #define INNERFLOPS INNERFLOPS+5 */
 479             /* #endfor */
 480
 481             /* #for I,J in PAIRS_IJ */
 482             /*     #if 'rinv' in INTERACTION_FLAGS[I][J] */
 483             rinv{I}{J}           = avx128fma_invsqrt_f(rsq{I}{J});
 484             /*         #define INNERFLOPS INNERFLOPS+5 */
 485             /*     #endif */
 486             /* #endfor */
 487
 488             /* #for I,J in PAIRS_IJ */
 489             /*     #if 'rinvsq' in INTERACTION_FLAGS[I][J] */
 490             /*         # if 'rinv' not in INTERACTION_FLAGS[I][J] */
 491             rinvsq{I}{J}         = avx128fma_inv_f(rsq{I}{J});
 492             /*             #define INNERFLOPS INNERFLOPS+4 */
 493             /*         #else */
 494             rinvsq{I}{J}         = _mm_mul_ps(rinv{I}{J},rinv{I}{J});
 495             /*             #define INNERFLOPS INNERFLOPS+1 */
 496             /*         #endif */
 497             /*     #endif */
 498             /* #endfor */
 499
 500             /* #if not 'Water' in GEOMETRY_J */
 501             /* Load parameters for j particles */
 502             /*     #for J in PARTICLES_ELEC_J */
 503             jq{J}              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+{J},charge+jnrB+{J},
 504                                                               charge+jnrC+{J},charge+jnrD+{J});
 505             /*         #if KERNEL_ELEC=='GeneralizedBorn' */
 506             isaj{J}            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+{J},invsqrta+jnrB+{J},
 507                                                               invsqrta+jnrC+{J},invsqrta+jnrD+{J});
 508             /*         #endif */
 509             /*     #endfor */
 510             /*     #for J in PARTICLES_VDW_J */
 511             vdwjidx{J}A        = 2*vdwtype[jnrA+{J}];
 512             vdwjidx{J}B        = 2*vdwtype[jnrB+{J}];
 513             vdwjidx{J}C        = 2*vdwtype[jnrC+{J}];
 514             vdwjidx{J}D        = 2*vdwtype[jnrD+{J}];
 515             /*     #endfor */
 516             /* #endif */
 517
 518             /* #if 'Force' in KERNEL_VF and not 'Particle' in GEOMETRY_I */
 519             /*     #for J in PARTICLES_J */
 520             fjx{J}             = _mm_setzero_ps();
 521             fjy{J}             = _mm_setzero_ps();
 522             fjz{J}             = _mm_setzero_ps();
 523             /*     #endfor */
 524             /* #endif */
 525
 526             /* #for I,J in PAIRS_IJ */
 527
 528             /**************************
 529              * CALCULATE INTERACTIONS *
 530              **************************/
 531
 532             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 533             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 534             /*         ## We always calculate rinv/rinvsq above to enable pipelineing in compilers (performance tested on x86) */
 535             if (gmx_mm_any_lt(rsq{I}{J},rcutoff2))
 536             {
 537                 /*     #if 0    ## this and the next two lines is a hack to maintain auto-indentation in template file */
 538             }
 539             /*         #endif */
 540             /*         #define INNERFLOPS INNERFLOPS+1 */
 541             /*     #endif */
 542
 543             /*     #if 'r' in INTERACTION_FLAGS[I][J] */
 544             r{I}{J}              = _mm_mul_ps(rsq{I}{J},rinv{I}{J});
 545             /*         #if ROUND == 'Epilogue' */
 546             r{I}{J}              = _mm_andnot_ps(dummy_mask,r{I}{J});
 547             /*             #define INNERFLOPS INNERFLOPS+1 */
 548             /*         #endif */
 549             /*         #define INNERFLOPS INNERFLOPS+1 */
 550             /*     #endif */
 551
 552             /*     ## For water geometries we already loaded parameters at the start of the kernel */
 553             /*     #if not 'Water' in GEOMETRY_J */
 554             /* Compute parameters for interactions between i and j atoms */
 555             /*         #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
 556             qq{I}{J}             = _mm_mul_ps(iq{I},jq{J});
 557             /*             #define INNERFLOPS INNERFLOPS+1 */
 558             /*         #endif */
 559             /*         #if 'vdw' in INTERACTION_FLAGS[I][J] */
 560             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset{I}+vdwjidx{J}A,
 561                                          vdwparam+vdwioffset{I}+vdwjidx{J}B,
 562                                          vdwparam+vdwioffset{I}+vdwjidx{J}C,
 563                                          vdwparam+vdwioffset{I}+vdwjidx{J}D,
 564                                          &c6_{I}{J},&c12_{I}{J});
 565
 566             /*           #if 'LJEwald' in KERNEL_VDW */
 567             c6grid_{I}{J}       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset{I}+vdwjidx{J}A,
 568                                                                vdwgridparam+vdwioffset{I}+vdwjidx{J}B,
 569                                                                vdwgridparam+vdwioffset{I}+vdwjidx{J}C,
 570                                                                vdwgridparam+vdwioffset{I}+vdwjidx{J}D);
 571             /*           #endif */
 572
 573             /*         #endif */
 574             /*     #endif */
 575
 576             /*     #if 'table' in INTERACTION_FLAGS[I][J] */
 577             /* Calculate table index by multiplying r with table scale and truncate to integer */
 578             rt               = _mm_mul_ps(r{I}{J},vftabscale);
 579             vfitab           = _mm_cvttps_epi32(rt);
 580 #ifdef __XOP__
 581             vfeps            = _mm_frcz_ps(rt);
 582 #else
 583             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 584 #endif
 585             twovfeps         = _mm_add_ps(vfeps,vfeps);
 586             /*         #define INNERFLOPS INNERFLOPS+4                          */
 587             /*         #if 'Table' in KERNEL_ELEC and 'Table' in KERNEL_VDW     */
 588             /*             ## 3 tables, 4 bytes per point: multiply index by 12 */
 589             vfitab           = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
 590             /*         #elif 'Table' in KERNEL_ELEC                             */
 591             /*             ## 1 table, 4 bytes per point: multiply index by 4   */
 592             vfitab           = _mm_slli_epi32(vfitab,2);
 593             /*         #elif 'Table' in KERNEL_VDW                              */
 594             /*             ## 2 tables, 4 bytes per point: multiply index by 8  */
 595             vfitab           = _mm_slli_epi32(vfitab,3);
 596             /*         #endif                                                   */
 597             /*     #endif */
 598
 599             /*     ## ELECTROSTATIC INTERACTIONS */
 600             /*     #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
 601
 602             /*         #if KERNEL_ELEC=='Coulomb' */
 603
 604             /* COULOMB ELECTROSTATICS */
 605             velec            = _mm_mul_ps(qq{I}{J},rinv{I}{J});
 606             /*             #define INNERFLOPS INNERFLOPS+1 */
 607             /*             #if 'Force' in KERNEL_VF */
 608             felec            = _mm_mul_ps(velec,rinvsq{I}{J});
 609             /*                 #define INNERFLOPS INNERFLOPS+2 */
 610             /*             #endif */
 611
 612             /*         #elif KERNEL_ELEC=='ReactionField' */
 613
 614             /* REACTION-FIELD ELECTROSTATICS */
 615             /*             #if 'Potential' in KERNEL_VF */
 616             velec            = _mm_mul_ps(qq{I}{J},_mm_sub_ps(_mm_macc_ps(krf,rsq{I}{J},rinv{I}{J}),crf));
 617             /*                 #define INNERFLOPS INNERFLOPS+4 */
 618             /*             #endif */
 619             /*             #if 'Force' in KERNEL_VF */
 620             felec            = _mm_mul_ps(qq{I}{J},_mm_msub_ps(rinv{I}{J},rinvsq{I}{J},krf2));
 621             /*                 #define INNERFLOPS INNERFLOPS+3 */
 622             /*             #endif */
 623
 624             /*         #elif KERNEL_ELEC=='GeneralizedBorn' */
 625
 626             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 627             isaprod          = _mm_mul_ps(isai{I},isaj{J});
 628             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq{I}{J},_mm_mul_ps(isaprod,gbinvepsdiff)));
 629             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 630             /*             #define INNERFLOPS INNERFLOPS+5 */
 631
 632             /* Calculate generalized born table index - this is a separate table from the normal one,
 633              * but we use the same procedure by multiplying r with scale and truncating to integer.
 634              */
 635             rt               = _mm_mul_ps(r{I}{J},gbscale);
 636             gbitab           = _mm_cvttps_epi32(rt);
 637 #ifdef __XOP__
 638             gbeps            = _mm_frcz_ps(rt);
 639 #else
 640             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 641 #endif
 642             gbitab           = _mm_slli_epi32(gbitab,2);
 643
 644             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 645             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 646             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 647             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 648             _MM_TRANSPOSE4_PS(Y,F,G,H);
 649             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 650             VV               = _mm_macc_ps(gbeps,Fp,Y);
 651             vgb              = _mm_mul_ps(gbqqfactor,VV);
 652             /*             #define INNERFLOPS INNERFLOPS+10 */
 653
 654             /*             #if 'Force' in KERNEL_VF */
 655             twogbeps         = _mm_add_ps(gbeps,gbeps);
 656             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 657             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 658             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r{I}{J},vgb));
 659             /*                 #if ROUND == 'Epilogue' */
 660             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 661             /*                 #endif */
 662             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 663             /*                 #if ROUND == 'Loop' */
 664             fjptrA           = dvda+jnrA;
 665             fjptrB           = dvda+jnrB;
 666             fjptrC           = dvda+jnrC;
 667             fjptrD           = dvda+jnrD;
 668             /*                 #else */
 669             /* 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. */
 670             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 671             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 672             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 673             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 674             /*                 #endif */
 675             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj{J},isaj{J})));
 676             /*                 #define INNERFLOPS INNERFLOPS+13 */
 677             /*             #endif */
 678             velec            = _mm_mul_ps(qq{I}{J},rinv{I}{J});
 679             /*                 #define INNERFLOPS INNERFLOPS+1 */
 680             /*             #if 'Force' in KERNEL_VF */
 681             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv{I}{J},fgb),rinv{I}{J});
 682             /*                 #define INNERFLOPS INNERFLOPS+3 */
 683             /*             #endif */
 684
 685             /*         #elif KERNEL_ELEC=='Ewald' */
 686             /* EWALD ELECTROSTATICS */
 687
 688             /* Analytical PME correction */
 689             zeta2            = _mm_mul_ps(beta2,rsq{I}{J});
 690             /*             #if 'Force' in KERNEL_VF */
 691             rinv3            = _mm_mul_ps(rinvsq{I}{J},rinv{I}{J});
 692             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 693             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 694             felec            = _mm_mul_ps(qq{I}{J},felec);
 695             /*             #endif */
 696             /*             #if 'Potential' in KERNEL_VF or KERNEL_MOD_ELEC=='PotentialSwitch' */
 697             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 698             /*                 #if KERNEL_MOD_ELEC=='PotentialShift' */
 699             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv{I}{J},sh_ewald));
 700             /*                 #else */
 701             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv{I}{J});
 702             /*                 #endif */
 703             velec            = _mm_mul_ps(qq{I}{J},velec);
 704             /*             #endif */
 705
 706             /*         #elif KERNEL_ELEC=='CubicSplineTable' */
 707
 708             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 709             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 710             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 711             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 712             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 713             _MM_TRANSPOSE4_PS(Y,F,G,H);
 714             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 715             /*             #define INNERFLOPS INNERFLOPS+4 */
 716             /*             #if 'Potential' in KERNEL_VF */
 717             VV               = _mm_macc_ps(vfeps,Fp,Y);
 718             velec            = _mm_mul_ps(qq{I}{J},VV);
 719             /*                 #define INNERFLOPS INNERFLOPS+3 */
 720             /*             #endif */
 721             /*             #if 'Force' in KERNEL_VF */
 722             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 723             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq{I}{J},FF),_mm_mul_ps(vftabscale,rinv{I}{J})));
 724             /*                 #define INNERFLOPS INNERFLOPS+7 */
 725             /*             #endif */
 726             /*         #endif */
 727             /*         ## End of check for electrostatics interaction forms */
 728             /*     #endif */
 729             /*     ## END OF ELECTROSTATIC INTERACTION CHECK FOR PAIR I-J */
 730
 731             /*     #if 'vdw' in INTERACTION_FLAGS[I][J] */
 732
 733             /*         #if KERNEL_VDW=='LennardJones' */
 734
 735             /* LENNARD-JONES DISPERSION/REPULSION */
 736
 737             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq{I}{J},rinvsq{I}{J}),rinvsq{I}{J});
 738             /*             #define INNERFLOPS INNERFLOPS+2 */
 739             /*             #if 'Potential' in KERNEL_VF or KERNEL_MOD_VDW=='PotentialSwitch' */
 740             vvdw6            = _mm_mul_ps(c6_{I}{J},rinvsix);
 741             vvdw12           = _mm_mul_ps(c12_{I}{J},_mm_mul_ps(rinvsix,rinvsix));
 742             /*                 #define INNERFLOPS INNERFLOPS+3 */
 743             /*                 #if KERNEL_MOD_VDW=='PotentialShift' */
 744             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_{I}{J},_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 745                                           _mm_mul_ps( _mm_nmacc_ps(c6_{I}{J},sh_vdw_invrcut6,vvdw6),one_sixth));
 746             /*                     #define INNERFLOPS INNERFLOPS+8 */
 747             /*                 #else */
 748             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 749             /*                     #define INNERFLOPS INNERFLOPS+3 */
 750             /*                 #endif */
 751             /*                 ## Check for force inside potential check, i.e. this means we already did the potential part */
 752             /*                 #if 'Force' in KERNEL_VF */
 753             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq{I}{J});
 754             /*                     #define INNERFLOPS INNERFLOPS+2 */
 755             /*                 #endif */
 756             /*             #elif KERNEL_VF=='Force' */
 757             /*                 ## Force-only LennardJones makes it possible to save 1 flop (they do add up...) */
 758             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_{I}{J},rinvsix,c6_{I}{J}),_mm_mul_ps(rinvsix,rinvsq{I}{J}));
 759             /*                 #define INNERFLOPS INNERFLOPS+4 */
 760             /*             #endif */
 761
 762             /*         #elif KERNEL_VDW=='CubicSplineTable' */
 763
 764             /* CUBIC SPLINE TABLE DISPERSION */
 765             /*             #if 'Table' in KERNEL_ELEC */
 766             vfitab           = _mm_add_epi32(vfitab,ifour);
 767             /*             #endif                     */
 768             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 769             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 770             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 771             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 772             _MM_TRANSPOSE4_PS(Y,F,G,H);
 773             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 774             /*             #define INNERFLOPS INNERFLOPS+4 */
 775             /*             #if 'Potential' in KERNEL_VF */
 776             VV               = _mm_macc_ps(vfeps,Fp,Y);
 777             vvdw6            = _mm_mul_ps(c6_{I}{J},VV);
 778             /*                 #define INNERFLOPS INNERFLOPS+3 */
 779             /*             #endif */
 780             /*             #if 'Force' in KERNEL_VF */
 781             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 782             fvdw6            = _mm_mul_ps(c6_{I}{J},FF);
 783             /*                 #define INNERFLOPS INNERFLOPS+4 */
 784             /*             #endif */
 785
 786             /* CUBIC SPLINE TABLE REPULSION */
 787             vfitab           = _mm_add_epi32(vfitab,ifour);
 788             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 789             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 790             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 791             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 792             _MM_TRANSPOSE4_PS(Y,F,G,H);
 793             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 794             /*             #define INNERFLOPS INNERFLOPS+4 */
 795             /*             #if 'Potential' in KERNEL_VF */
 796             VV               = _mm_macc_ps(vfeps,Fp,Y);
 797             vvdw12           = _mm_mul_ps(c12_{I}{J},VV);
 798             /*                 #define INNERFLOPS INNERFLOPS+3 */
 799             /*             #endif */
 800             /*             #if 'Force' in KERNEL_VF */
 801             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 802             fvdw12           = _mm_mul_ps(c12_{I}{J},FF);
 803             /*                 #define INNERFLOPS INNERFLOPS+5 */
 804             /*             #endif */
 805             /*             #if 'Potential' in KERNEL_VF */
 806             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 807             /*                 #define INNERFLOPS INNERFLOPS+1 */
 808             /*             #endif */
 809             /*             #if 'Force' in KERNEL_VF */
 810             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv{I}{J})));
 811             /*                 #define INNERFLOPS INNERFLOPS+4 */
 812             /*             #endif */
 813
 814             /*         #elif KERNEL_VDW=='LJEwald' */
 815
 816             /* Analytical LJ-PME */
 817             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq{I}{J},rinvsq{I}{J}),rinvsq{I}{J});
 818             ewcljrsq         = _mm_mul_ps(ewclj2,rsq{I}{J});
 819             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 820             exponent         = avx128fma_exp_f(ewcljrsq);
 821             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 822             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 823             /*                 #define INNERFLOPS INNERFLOPS+10 */
 824             /*             #if 'Potential' in KERNEL_VF or KERNEL_MOD_VDW=='PotentialSwitch' */
 825             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 826             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_{I}{J},_mm_sub_ps(one,poly),c6_{I}{J}),rinvsix);
 827             vvdw12           = _mm_mul_ps(c12_{I}{J},_mm_mul_ps(rinvsix,rinvsix));
 828             /*                     #define INNERFLOPS INNERFLOPS+5 */
 829             /*                 #if KERNEL_MOD_VDW=='PotentialShift' */
 830             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_{I}{J},_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 831                                           _mm_mul_ps(_mm_sub_ps(vvdw6,_mm_macc_ps(c6grid_{I}{J},sh_lj_ewald,_mm_mul_ps(c6_{I}{J},sh_vdw_invrcut6))),one_sixth));
 832             /*                     #define INNERFLOPS INNERFLOPS+7 */
 833             /*                 #else */
 834             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 835             /*                     #define INNERFLOPS INNERFLOPS+2 */
 836             /*                 #endif */
 837             /*                  ## Check for force inside potential check, i.e. this means we already did the potential part */
 838             /*                  #if 'Force' in KERNEL_VF */
 839             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 840             fvdw             = _mm_mul_ps(_mm_add_ps(vvdw12,_mm_msub_ps(_mm_mul_ps(c6grid_{I}{J},one_sixth),_mm_mul_ps(exponent,ewclj6),vvdw6)),rinvsq{I}{J});
 841             /*                 #define INNERFLOPS INNERFLOPS+5 */
 842             /*                  #endif */
 843             /*              #elif KERNEL_VF=='Force' */
 844             /* f6A = 6 * C6grid * (1 - poly) */
 845             f6A              = _mm_mul_ps(c6grid_{I}{J},_mm_sub_ps(one,poly));
 846             /* f6B = C6grid * exponent * beta^6 */
 847             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_{I}{J},one_sixth),_mm_mul_ps(exponent,ewclj6));
 848             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 849             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_{I}{J},rinvsix,_mm_sub_ps(c6_{I}{J},f6A)),rinvsix,f6B),rinvsq{I}{J});
 850             /*                 #define INNERFLOPS INNERFLOPS+10 */
 851             /*              #endif */
 852             /*         #endif */
 853             /*         ## End of check for vdw interaction forms */
 854             /*     #endif */
 855             /*     ## END OF VDW INTERACTION CHECK FOR PAIR I-J */
 856
 857             /*     #if 'switch' in INTERACTION_FLAGS[I][J] */
 858             d                = _mm_sub_ps(r{I}{J},rswitch);
 859             d                = _mm_max_ps(d,_mm_setzero_ps());
 860             d2               = _mm_mul_ps(d,d);
 861             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 862             /*         #define INNERFLOPS INNERFLOPS+10 */
 863
 864             /*         #if 'Force' in KERNEL_VF */
 865             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 866             /*             #define INNERFLOPS INNERFLOPS+5 */
 867             /*         #endif */
 868
 869             /* Evaluate switch function */
 870             /*         #if 'Force' in KERNEL_VF */
 871             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 872             /*             #if 'electrostatics' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_ELEC=='PotentialSwitch' */
 873             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv{I}{J},_mm_mul_ps(velec,dsw)) );
 874             /*                 #define INNERFLOPS INNERFLOPS+4 */
 875             /*             #endif */
 876             /*             #if 'vdw' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_VDW=='PotentialSwitch' */
 877             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv{I}{J},_mm_mul_ps(vvdw,dsw)) );
 878             /*                 #define INNERFLOPS INNERFLOPS+4 */
 879             /*             #endif */
 880             /*         #endif */
 881             /*         #if 'Potential' in KERNEL_VF */
 882             /*             #if 'electrostatics' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_ELEC=='PotentialSwitch' */
 883             velec            = _mm_mul_ps(velec,sw);
 884             /*                 #define INNERFLOPS INNERFLOPS+1 */
 885             /*             #endif */
 886             /*             #if 'vdw' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_VDW=='PotentialSwitch' */
 887             vvdw             = _mm_mul_ps(vvdw,sw);
 888             /*                 #define INNERFLOPS INNERFLOPS+1 */
 889             /*             #endif */
 890             /*         #endif */
 891             /*     #endif */
 892             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 893             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 894             cutoff_mask      = _mm_cmplt_ps(rsq{I}{J},rcutoff2);
 895             /*         #define INNERFLOPS INNERFLOPS+1 */
 896             /*     #endif */
 897
 898             /*     #if 'Potential' in KERNEL_VF */
 899             /* Update potential sum for this i atom from the interaction with this j atom. */
 900             /*         #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
 901             /*             #if 'exactcutoff' in INTERACTION_FLAGS[I][J] */
 902             velec            = _mm_and_ps(velec,cutoff_mask);
 903             /*                 #define INNERFLOPS INNERFLOPS+1 */
 904             /*             #endif                                       */
 905             /*             #if ROUND == 'Epilogue' */
 906             velec            = _mm_andnot_ps(dummy_mask,velec);
 907             /*             #endif */
 908             velecsum         = _mm_add_ps(velecsum,velec);
 909             /*             #define INNERFLOPS INNERFLOPS+1 */
 910             /*             #if KERNEL_ELEC=='GeneralizedBorn' */
 911             /*             #if 'exactcutoff' in INTERACTION_FLAGS[I][J] */
 912             vgb              = _mm_and_ps(vgb,cutoff_mask);
 913             /*                 #define INNERFLOPS INNERFLOPS+1 */
 914             /*             #endif                                       */
 915             /*             #if ROUND == 'Epilogue' */
 916             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 917             /*             #endif */
 918             vgbsum           = _mm_add_ps(vgbsum,vgb);
 919             /*                 #define INNERFLOPS INNERFLOPS+1 */
 920             /*             #endif */
 921             /*         #endif */
 922             /*         #if 'vdw' in INTERACTION_FLAGS[I][J] */
 923             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 924             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 925             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 926             /*                 #define INNERFLOPS INNERFLOPS+1 */
 927             /*             #endif                                       */
 928             /*             #if ROUND == 'Epilogue' */
 929             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 930             /*             #endif */
 931             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 932             /*             #define INNERFLOPS INNERFLOPS+1 */
 933             /*         #endif */
 934             /*     #endif */
 935
 936             /*     #if 'Force' in KERNEL_VF */
 937
 938             /*         #if 'electrostatics' in INTERACTION_FLAGS[I][J] and 'vdw' in INTERACTION_FLAGS[I][J] */
 939             fscal            = _mm_add_ps(felec,fvdw);
 940             /*             #define INNERFLOPS INNERFLOPS+1 */
 941             /*         #elif 'electrostatics' in INTERACTION_FLAGS[I][J] */
 942             fscal            = felec;
 943             /*         #elif 'vdw' in INTERACTION_FLAGS[I][J] */
 944             fscal            = fvdw;
 945             /*        #endif */
 946
 947             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 948             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 949             fscal            = _mm_and_ps(fscal,cutoff_mask);
 950             /*                 #define INNERFLOPS INNERFLOPS+1 */
 951             /*             #endif                                       */
 952
 953             /*             #if ROUND == 'Epilogue' */
 954             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 955             /*             #endif */
 956
 957             /* ## Construction of vectorial force built into FMA instructions now */
 958             /* #define INNERFLOPS INNERFLOPS+3      */
 959
 960              /* Update vectorial force */
 961             fix{I}             = _mm_macc_ps(dx{I}{J},fscal,fix{I});
 962             fiy{I}             = _mm_macc_ps(dy{I}{J},fscal,fiy{I});
 963             fiz{I}             = _mm_macc_ps(dz{I}{J},fscal,fiz{I});
 964             /*             #define INNERFLOPS INNERFLOPS+6 */
 965
 966             /* #if GEOMETRY_I == 'Particle'             */
 967             /*     #if ROUND == 'Loop' */
 968             fjptrA             = f+j_coord_offsetA;
 969             fjptrB             = f+j_coord_offsetB;
 970             fjptrC             = f+j_coord_offsetC;
 971             fjptrD             = f+j_coord_offsetD;
 972             /*     #else */
 973             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 974             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 975             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 976             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 977             /*     #endif */
 978             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 979                                                    _mm_mul_ps(dx{I}{J},fscal),
 980                                                    _mm_mul_ps(dy{I}{J},fscal),
 981                                                    _mm_mul_ps(dz{I}{J},fscal));
 982             /*     #define INNERFLOPS INNERFLOPS+3      */
 983             /* #else                                    */
 984             fjx{J}             = _mm_macc_ps(dx{I}{J},fscal,fjx{J});
 985             fjy{J}             = _mm_macc_ps(dy{I}{J},fscal,fjy{J});
 986             fjz{J}             = _mm_macc_ps(dz{I}{J},fscal,fjz{J});
 987             /*     #define INNERFLOPS INNERFLOPS+3      */
 988             /* #endif                                   */
 989
 990             /*     #endif */
 991
 992             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 993             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 994             /*         #if 0    ## This and next two lines is a hack to maintain indentation in template file */
 995             {
 996                 /*     #endif */
 997             }
 998             /*     #endif */
 999             /*    ## End of check for the interaction being outside the cutoff */
1000
1001             /* #endfor */
1002             /* ## End of loop over i-j interaction pairs */
1003
1004             /* #if GEOMETRY_I != 'Particle' */
1005             /*     #if ROUND == 'Loop' */
1006             fjptrA             = f+j_coord_offsetA;
1007             fjptrB             = f+j_coord_offsetB;
1008             fjptrC             = f+j_coord_offsetC;
1009             fjptrD             = f+j_coord_offsetD;
1010             /*     #else */
1011             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1012             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1013             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1014             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1015             /*     #endif */
1016             /* #endif */
1017
1018             /* #if 'Water' in GEOMETRY_I and GEOMETRY_J == 'Particle' */
1019             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1020             /* #elif GEOMETRY_J == 'Water3'               */
1021             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1022                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1023             /*     #define INNERFLOPS INNERFLOPS+9      */
1024             /* #elif GEOMETRY_J == 'Water4'             */
1025             /*     #if 0 in PARTICLES_J                 */
1026             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1027                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1028                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1029             /*     #define INNERFLOPS INNERFLOPS+12     */
1030             /*     #else                                */
1031             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1032                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1033             /*     #define INNERFLOPS INNERFLOPS+9      */
1034             /*     #endif                               */
1035             /* #endif                                   */
1036
1037             /* Inner loop uses {INNERFLOPS} flops */
1038         }
1039
1040         /* #endfor */
1041
1042         /* End of innermost loop */
1043
1044         /* #if 'Force' in KERNEL_VF */
1045         /*     #if GEOMETRY_I == 'Particle'            */
1046         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
1047                                               f+i_coord_offset,fshift+i_shift_offset);
1048         /*         #define OUTERFLOPS OUTERFLOPS+6     */
1049         /*     #elif GEOMETRY_I == 'Water3'            */
1050         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1051                                               f+i_coord_offset,fshift+i_shift_offset);
1052         /*         #define OUTERFLOPS OUTERFLOPS+18    */
1053         /*     #elif GEOMETRY_I == 'Water4'            */
1054         /*         #if 0 in PARTICLES_I                */
1055         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1056                                               f+i_coord_offset,fshift+i_shift_offset);
1057         /*             #define OUTERFLOPS OUTERFLOPS+24    */
1058         /*         #else                               */
1059         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1060                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1061         /*             #define OUTERFLOPS OUTERFLOPS+18    */
1062         /*         #endif                              */
1063         /*     #endif                                  */
1064         /* #endif                                      */
1065
1066         /* #if 'Potential' in KERNEL_VF */
1067         ggid                        = gid[iidx];
1068         /* Update potential energies */
1069         /*     #if KERNEL_ELEC != 'None' */
1070         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1071         /*         #define OUTERFLOPS OUTERFLOPS+1 */
1072         /*     #endif */
1073         /*     #if 'GeneralizedBorn' in KERNEL_ELEC */
1074         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
1075         /*         #define OUTERFLOPS OUTERFLOPS+1 */
1076         /*     #endif */
1077         /*     #if KERNEL_VDW != 'None' */
1078         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1079         /*         #define OUTERFLOPS OUTERFLOPS+1 */
1080         /*     #endif */
1081         /* #endif */
1082         /*     #if 'GeneralizedBorn' in KERNEL_ELEC and 'Force' in KERNEL_VF */
1083         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai{I},isai{I}));
1084         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
1085         /*     #endif */
1086
1087         /* Increment number of inner iterations */
1088         inneriter                  += j_index_end - j_index_start;
1089
1090         /* Outer loop uses {OUTERFLOPS} flops */
1091     }
1092
1093     /* Increment number of outer iterations */
1094     outeriter        += nri;
1095
1096     /* Update outer/inner flops */
1097     /* ## NB: This is not important, it just affects the flopcount. However, since our preprocessor is */
1098     /* ## primitive and replaces aggressively even in strings inside these directives, we need to      */
1099     /* ## assemble the main part of the name (containing KERNEL/ELEC/VDW) directly in the source.      */
1100     /* #if GEOMETRY_I == 'Water3'            */
1101     /*     #define ISUFFIX '_W3'             */
1102     /* #elif GEOMETRY_I == 'Water4'          */
1103     /*     #define ISUFFIX '_W4'             */
1104     /* #else                                 */
1105     /*     #define ISUFFIX ''                */
1106     /* #endif                                */
1107     /* #if GEOMETRY_J == 'Water3'            */
1108     /*     #define JSUFFIX 'W3'              */
1109     /* #elif GEOMETRY_J == 'Water4'          */
1110     /*     #define JSUFFIX 'W4'              */
1111     /* #else                                 */
1112     /*     #define JSUFFIX ''                */
1113     /* #endif                                */
1114     /* #if 'PotentialAndForce' in KERNEL_VF  */
1115     /*     #define VFSUFFIX  '_VF'           */
1116     /* #elif 'Potential' in KERNEL_VF        */
1117     /*     #define VFSUFFIX '_V'             */
1118     /* #else                                 */
1119     /*     #define VFSUFFIX '_F'             */
1120     /* #endif                                */
1121
1122     /* #if KERNEL_ELEC != 'None' and KERNEL_VDW != 'None' */
1123     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW{ISUFFIX}{JSUFFIX}{VFSUFFIX},outeriter*{OUTERFLOPS} + inneriter*{INNERFLOPS});
1124     /* #elif KERNEL_ELEC != 'None' */
1125     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC{ISUFFIX}{JSUFFIX}{VFSUFFIX},outeriter*{OUTERFLOPS} + inneriter*{INNERFLOPS});
1126     /* #else */
1127     inc_nrnb(nrnb,eNR_NBKERNEL_VDW{ISUFFIX}{JSUFFIX}{VFSUFFIX},outeriter*{OUTERFLOPS} + inneriter*{INNERFLOPS});
1128     /* #endif  */
1129 }