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) 2012,2013,2014,2015,2017,2018, 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 KERNEL_VDW != 'None' */
 135     int              nvdwtype;
 136     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 137     int              *vdwtype;
 138     real             *vdwparam;
 139     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 140     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 141     /* #endif */
 142     /* #if 'Table' in KERNEL_ELEC or 'Table' in KERNEL_VDW */
 143     __m128i          vfitab;
 144     __m128i          ifour       = _mm_set1_epi32(4);
 145     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 146     real             *vftab;
 147     /* #endif */
 148     /* #if 'LJEwald' in KERNEL_VDW */
 149     /* #for I,J in PAIRS_IJ */
 150     __m128           c6grid_{I}{J};
 151     /* #endfor */
 152     real             *vdwgridparam;
 153     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 154     __m128           one_half = _mm_set1_ps(0.5);
 155     __m128           minus_one = _mm_set1_ps(-1.0);
 156     /* #endif */
 157     /* #if 'Ewald' in KERNEL_ELEC */
 158     __m128i          ewitab;
 159     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 160     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 161     real             *ewtab;
 162     /* #endif */
 163     /* #if 'PotentialSwitch' in [KERNEL_MOD_ELEC,KERNEL_MOD_VDW] */
 164     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 165     real             rswitch_scalar,d_scalar;
 166     /* #endif */
 167     __m128           dummy_mask,cutoff_mask;
 168     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 169     __m128           one     = _mm_set1_ps(1.0);
 170     __m128           two     = _mm_set1_ps(2.0);
 171     x                = xx[0];
 172     f                = ff[0];
 173
 174     nri              = nlist->nri;
 175     iinr             = nlist->iinr;
 176     jindex           = nlist->jindex;
 177     jjnr             = nlist->jjnr;
 178     shiftidx         = nlist->shift;
 179     gid              = nlist->gid;
 180     shiftvec         = fr->shift_vec[0];
 181     fshift           = fr->fshift[0];
 182     /* #if KERNEL_ELEC != 'None' */
 183     facel            = _mm_set1_ps(fr->ic->epsfac);
 184     charge           = mdatoms->chargeA;
 185     /*     #if 'ReactionField' in KERNEL_ELEC */
 186     krf              = _mm_set1_ps(fr->ic->k_rf);
 187     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 188     crf              = _mm_set1_ps(fr->ic->c_rf);
 189     /*     #endif */
 190     /* #endif */
 191     /* #if KERNEL_VDW != 'None' */
 192     nvdwtype         = fr->ntype;
 193     vdwparam         = fr->nbfp;
 194     vdwtype          = mdatoms->typeA;
 195     /* #endif */
 196     /* #if 'LJEwald' in KERNEL_VDW */
 197     vdwgridparam     = fr->ljpme_c6grid;
 198     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 199     ewclj            = _mm_set1_ps(fr->ic->ewaldcoeff_lj);
 200     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 201     /* #endif */
 202
 203     /* #if 'Table' in KERNEL_ELEC and 'Table' in KERNEL_VDW */
 204     vftab            = kernel_data->table_elec_vdw->data;
 205     vftabscale       = _mm_set1_ps(kernel_data->table_elec_vdw->scale);
 206     /* #elif 'Table' in KERNEL_ELEC */
 207     vftab            = kernel_data->table_elec->data;
 208     vftabscale       = _mm_set1_ps(kernel_data->table_elec->scale);
 209     /* #elif 'Table' in KERNEL_VDW */
 210     vftab            = kernel_data->table_vdw->data;
 211     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 212     /* #endif */
 213
 214     /* #if 'Ewald' in KERNEL_ELEC */
 215     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 216     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 217     beta2            = _mm_mul_ps(beta,beta);
 218     beta3            = _mm_mul_ps(beta,beta2);
 219     /*     #if KERNEL_VF=='Force' and KERNEL_MOD_ELEC!='PotentialSwitch' */
 220     ewtab            = fr->ic->tabq_coul_F;
 221     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 222     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 223     /*     #else */
 224     ewtab            = fr->ic->tabq_coul_FDV0;
 225     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 226     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 227      /*     #endif */
 228     /* #endif */
 229
 230     /* #if 'Water' in GEOMETRY_I */
 231     /* Setup water-specific parameters */
 232     inr              = nlist->iinr[0];
 233     /*     #for I in PARTICLES_ELEC_I */
 234     iq{I}              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+{I}]));
 235     /*     #endfor */
 236     /*     #for I in PARTICLES_VDW_I */
 237     vdwioffset{I}      = 2*nvdwtype*vdwtype[inr+{I}];
 238     /*     #endfor */
 239     /* #endif */
 240
 241     /* #if 'Water' in GEOMETRY_J */
 242     /*     #for J in PARTICLES_ELEC_J */
 243     jq{J}              = _mm_set1_ps(charge[inr+{J}]);
 244     /*     #endfor */
 245     /*     #for J in PARTICLES_VDW_J */
 246     vdwjidx{J}A        = 2*vdwtype[inr+{J}];
 247     /*     #endfor */
 248     /*     #for I,J in PAIRS_IJ */
 249     /*         #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
 250     qq{I}{J}             = _mm_mul_ps(iq{I},jq{J});
 251     /*         #endif */
 252     /*         #if 'vdw' in INTERACTION_FLAGS[I][J] */
 253     /*             #if 'LJEwald' in KERNEL_VDW */
 254     c6_{I}{J}            = _mm_set1_ps(vdwparam[vdwioffset{I}+vdwjidx{J}A]);
 255     c12_{I}{J}           = _mm_set1_ps(vdwparam[vdwioffset{I}+vdwjidx{J}A+1]);
 256     c6grid_{I}{J}        = _mm_set1_ps(vdwgridparam[vdwioffset{I}+vdwjidx{J}A]);
 257     /*             #else */
 258     c6_{I}{J}            = _mm_set1_ps(vdwparam[vdwioffset{I}+vdwjidx{J}A]);
 259     c12_{I}{J}           = _mm_set1_ps(vdwparam[vdwioffset{I}+vdwjidx{J}A+1]);
 260     /*             #endif */
 261     /*         #endif */
 262     /*     #endfor */
 263     /* #endif */
 264
 265     /* #if KERNEL_MOD_ELEC!='None' or KERNEL_MOD_VDW!='None' */
 266     /*     #if KERNEL_ELEC!='None' */
 267     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 268     rcutoff_scalar   = fr->ic->rcoulomb;
 269     /*     #else */
 270     rcutoff_scalar   = fr->ic->rvdw;
 271     /*     #endif */
 272     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 273     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 274     /* #endif */
 275
 276     /* #if KERNEL_MOD_VDW=='PotentialShift' */
 277     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 278     rvdw             = _mm_set1_ps(fr->ic->rvdw);
 279     /* #endif */
 280
 281     /* #if 'PotentialSwitch' in [KERNEL_MOD_ELEC,KERNEL_MOD_VDW] */
 282     /*     #if KERNEL_MOD_ELEC=='PotentialSwitch'  */
 283     rswitch_scalar   = fr->ic->rcoulomb_switch;
 284     rswitch          = _mm_set1_ps(rswitch_scalar);
 285     /*     #else */
 286     rswitch_scalar   = fr->ic->rvdw_switch;
 287     rswitch          = _mm_set1_ps(rswitch_scalar);
 288     /*     #endif */
 289     /* Setup switch parameters */
 290     d_scalar         = rcutoff_scalar-rswitch_scalar;
 291     d                = _mm_set1_ps(d_scalar);
 292     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 293     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 294     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 295     /*     #if 'Force' in KERNEL_VF */
 296     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 297     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 298     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 299     /*     #endif */
 300     /* #endif */
 301
 302     /* Avoid stupid compiler warnings */
 303     jnrA = jnrB = jnrC = jnrD = 0;
 304     j_coord_offsetA = 0;
 305     j_coord_offsetB = 0;
 306     j_coord_offsetC = 0;
 307     j_coord_offsetD = 0;
 308
 309     /* ## Keep track of the floating point operations we issue for reporting! */
 310     /* #define OUTERFLOPS 0 */
 311     outeriter        = 0;
 312     inneriter        = 0;
 313
 314     for(iidx=0;iidx<4*DIM;iidx++)
 315     {
 316         scratch[iidx] = 0.0;
 317     }
 318
 319     /* Start outer loop over neighborlists */
 320     for(iidx=0; iidx<nri; iidx++)
 321     {
 322         /* Load shift vector for this list */
 323         i_shift_offset   = DIM*shiftidx[iidx];
 324
 325         /* Load limits for loop over neighbors */
 326         j_index_start    = jindex[iidx];
 327         j_index_end      = jindex[iidx+1];
 328
 329         /* Get outer coordinate index */
 330         inr              = iinr[iidx];
 331         i_coord_offset   = DIM*inr;
 332
 333         /* Load i particle coords and add shift vector */
 334         /* #if GEOMETRY_I == 'Particle' */
 335         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 336         /* #elif GEOMETRY_I == 'Water3' */
 337         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 338                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 339         /* #elif GEOMETRY_I == 'Water4' */
 340         /*     #if 0 in PARTICLES_I                 */
 341         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 342                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 343         /*     #else                                */
 344         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 345                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 346         /*     #endif                               */
 347         /* #endif                                   */
 348
 349         /* #if 'Force' in KERNEL_VF */
 350         /*     #for I in PARTICLES_I */
 351         fix{I}             = _mm_setzero_ps();
 352         fiy{I}             = _mm_setzero_ps();
 353         fiz{I}             = _mm_setzero_ps();
 354         /*     #endfor */
 355         /* #endif */
 356
 357         /* ## For water we already preloaded parameters at the start of the kernel */
 358         /* #if not 'Water' in GEOMETRY_I */
 359         /* Load parameters for i particles */
 360         /*     #for I in PARTICLES_ELEC_I */
 361         iq{I}              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+{I}));
 362         /*         #define OUTERFLOPS OUTERFLOPS+1 */
 363         /*     #endfor */
 364         /*     #for I in PARTICLES_VDW_I */
 365         vdwioffset{I}      = 2*nvdwtype*vdwtype[inr+{I}];
 366         /*     #endfor */
 367         /* #endif */
 368
 369         /* #if 'Potential' in KERNEL_VF */
 370         /* Reset potential sums */
 371         /*     #if KERNEL_ELEC != 'None' */
 372         velecsum         = _mm_setzero_ps();
 373         /*     #endif */
 374         /*     #if KERNEL_VDW != 'None' */
 375         vvdwsum          = _mm_setzero_ps();
 376         /*     #endif */
 377         /* #endif */
 378
 379         /* #for ROUND in ['Loop','Epilogue'] */
 380
 381         /* #if ROUND =='Loop' */
 382         /* Start inner kernel loop */
 383         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 384         {
 385         /* ## First round is normal loop (next statement resets indentation) */
 386         /*     #if 0 */
 387         }
 388         /*     #endif */
 389         /* #else */
 390         if(jidx<j_index_end)
 391         {
 392         /* ## Second round is epilogue */
 393         /* #endif */
 394         /* #define INNERFLOPS 0 */
 395
 396             /* Get j neighbor index, and coordinate index */
 397             /* #if ROUND =='Loop' */
 398             jnrA             = jjnr[jidx];
 399             jnrB             = jjnr[jidx+1];
 400             jnrC             = jjnr[jidx+2];
 401             jnrD             = jjnr[jidx+3];
 402             /* #else */
 403             jnrlistA         = jjnr[jidx];
 404             jnrlistB         = jjnr[jidx+1];
 405             jnrlistC         = jjnr[jidx+2];
 406             jnrlistD         = jjnr[jidx+3];
 407             /* Sign of each element will be negative for non-real atoms.
 408              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 409              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 410              */
 411             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 412             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 413             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 414             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 415             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 416             /* #endif */
 417             j_coord_offsetA  = DIM*jnrA;
 418             j_coord_offsetB  = DIM*jnrB;
 419             j_coord_offsetC  = DIM*jnrC;
 420             j_coord_offsetD  = DIM*jnrD;
 421
 422             /* load j atom coordinates */
 423             /* #if GEOMETRY_J == 'Particle'             */
 424             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 425                                               x+j_coord_offsetC,x+j_coord_offsetD,
 426                                               &jx0,&jy0,&jz0);
 427             /* #elif GEOMETRY_J == 'Water3'             */
 428             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 429                                               x+j_coord_offsetC,x+j_coord_offsetD,
 430                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 431             /* #elif GEOMETRY_J == 'Water4'             */
 432             /*     #if 0 in PARTICLES_J                 */
 433             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 434                                               x+j_coord_offsetC,x+j_coord_offsetD,
 435                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 436                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 437             /*     #else                                */
 438             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 439                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 440                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 441             /*     #endif                               */
 442             /* #endif                                   */
 443
 444             /* Calculate displacement vector */
 445             /* #for I,J in PAIRS_IJ */
 446             dx{I}{J}             = _mm_sub_ps(ix{I},jx{J});
 447             dy{I}{J}             = _mm_sub_ps(iy{I},jy{J});
 448             dz{I}{J}             = _mm_sub_ps(iz{I},jz{J});
 449             /*     #define INNERFLOPS INNERFLOPS+3 */
 450             /* #endfor */
 451
 452             /* Calculate squared distance and things based on it */
 453             /* #for I,J in PAIRS_IJ */
 454             rsq{I}{J}            = gmx_mm_calc_rsq_ps(dx{I}{J},dy{I}{J},dz{I}{J});
 455             /*     #define INNERFLOPS INNERFLOPS+5 */
 456             /* #endfor */
 457
 458             /* #for I,J in PAIRS_IJ */
 459             /*     #if 'rinv' in INTERACTION_FLAGS[I][J] */
 460             rinv{I}{J}           = avx128fma_invsqrt_f(rsq{I}{J});
 461             /*         #define INNERFLOPS INNERFLOPS+5 */
 462             /*     #endif */
 463             /* #endfor */
 464
 465             /* #for I,J in PAIRS_IJ */
 466             /*     #if 'rinvsq' in INTERACTION_FLAGS[I][J] */
 467             /*         # if 'rinv' not in INTERACTION_FLAGS[I][J] */
 468             rinvsq{I}{J}         = avx128fma_inv_f(rsq{I}{J});
 469             /*             #define INNERFLOPS INNERFLOPS+4 */
 470             /*         #else */
 471             rinvsq{I}{J}         = _mm_mul_ps(rinv{I}{J},rinv{I}{J});
 472             /*             #define INNERFLOPS INNERFLOPS+1 */
 473             /*         #endif */
 474             /*     #endif */
 475             /* #endfor */
 476
 477             /* #if not 'Water' in GEOMETRY_J */
 478             /* Load parameters for j particles */
 479             /*     #for J in PARTICLES_ELEC_J */
 480             jq{J}              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+{J},charge+jnrB+{J},
 481                                                               charge+jnrC+{J},charge+jnrD+{J});
 482             /*     #endfor */
 483             /*     #for J in PARTICLES_VDW_J */
 484             vdwjidx{J}A        = 2*vdwtype[jnrA+{J}];
 485             vdwjidx{J}B        = 2*vdwtype[jnrB+{J}];
 486             vdwjidx{J}C        = 2*vdwtype[jnrC+{J}];
 487             vdwjidx{J}D        = 2*vdwtype[jnrD+{J}];
 488             /*     #endfor */
 489             /* #endif */
 490
 491             /* #if 'Force' in KERNEL_VF and not 'Particle' in GEOMETRY_I */
 492             /*     #for J in PARTICLES_J */
 493             fjx{J}             = _mm_setzero_ps();
 494             fjy{J}             = _mm_setzero_ps();
 495             fjz{J}             = _mm_setzero_ps();
 496             /*     #endfor */
 497             /* #endif */
 498
 499             /* #for I,J in PAIRS_IJ */
 500
 501             /**************************
 502              * CALCULATE INTERACTIONS *
 503              **************************/
 504
 505             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 506             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 507             /*         ## We always calculate rinv/rinvsq above to enable pipelineing in compilers (performance tested on x86) */
 508             if (gmx_mm_any_lt(rsq{I}{J},rcutoff2))
 509             {
 510                 /*     #if 0    ## this and the next two lines is a hack to maintain auto-indentation in template file */
 511             }
 512             /*         #endif */
 513             /*         #define INNERFLOPS INNERFLOPS+1 */
 514             /*     #endif */
 515
 516             /*     #if 'r' in INTERACTION_FLAGS[I][J] */
 517             r{I}{J}              = _mm_mul_ps(rsq{I}{J},rinv{I}{J});
 518             /*         #if ROUND == 'Epilogue' */
 519             r{I}{J}              = _mm_andnot_ps(dummy_mask,r{I}{J});
 520             /*             #define INNERFLOPS INNERFLOPS+1 */
 521             /*         #endif */
 522             /*         #define INNERFLOPS INNERFLOPS+1 */
 523             /*     #endif */
 524
 525             /*     ## For water geometries we already loaded parameters at the start of the kernel */
 526             /*     #if not 'Water' in GEOMETRY_J */
 527             /* Compute parameters for interactions between i and j atoms */
 528             /*         #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
 529             qq{I}{J}             = _mm_mul_ps(iq{I},jq{J});
 530             /*             #define INNERFLOPS INNERFLOPS+1 */
 531             /*         #endif */
 532             /*         #if 'vdw' in INTERACTION_FLAGS[I][J] */
 533             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset{I}+vdwjidx{J}A,
 534                                          vdwparam+vdwioffset{I}+vdwjidx{J}B,
 535                                          vdwparam+vdwioffset{I}+vdwjidx{J}C,
 536                                          vdwparam+vdwioffset{I}+vdwjidx{J}D,
 537                                          &c6_{I}{J},&c12_{I}{J});
 538
 539             /*           #if 'LJEwald' in KERNEL_VDW */
 540             c6grid_{I}{J}       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset{I}+vdwjidx{J}A,
 541                                                                vdwgridparam+vdwioffset{I}+vdwjidx{J}B,
 542                                                                vdwgridparam+vdwioffset{I}+vdwjidx{J}C,
 543                                                                vdwgridparam+vdwioffset{I}+vdwjidx{J}D);
 544             /*           #endif */
 545
 546             /*         #endif */
 547             /*     #endif */
 548
 549             /*     #if 'table' in INTERACTION_FLAGS[I][J] */
 550             /* Calculate table index by multiplying r with table scale and truncate to integer */
 551             rt               = _mm_mul_ps(r{I}{J},vftabscale);
 552             vfitab           = _mm_cvttps_epi32(rt);
 553 #ifdef __XOP__
 554             vfeps            = _mm_frcz_ps(rt);
 555 #else
 556             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 557 #endif
 558             twovfeps         = _mm_add_ps(vfeps,vfeps);
 559             /*         #define INNERFLOPS INNERFLOPS+4                          */
 560             /*         #if 'Table' in KERNEL_ELEC and 'Table' in KERNEL_VDW     */
 561             /*             ## 3 tables, 4 bytes per point: multiply index by 12 */
 562             vfitab           = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
 563             /*         #elif 'Table' in KERNEL_ELEC                             */
 564             /*             ## 1 table, 4 bytes per point: multiply index by 4   */
 565             vfitab           = _mm_slli_epi32(vfitab,2);
 566             /*         #elif 'Table' in KERNEL_VDW                              */
 567             /*             ## 2 tables, 4 bytes per point: multiply index by 8  */
 568             vfitab           = _mm_slli_epi32(vfitab,3);
 569             /*         #endif                                                   */
 570             /*     #endif */
 571
 572             /*     ## ELECTROSTATIC INTERACTIONS */
 573             /*     #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
 574
 575             /*         #if KERNEL_ELEC=='Coulomb' */
 576
 577             /* COULOMB ELECTROSTATICS */
 578             velec            = _mm_mul_ps(qq{I}{J},rinv{I}{J});
 579             /*             #define INNERFLOPS INNERFLOPS+1 */
 580             /*             #if 'Force' in KERNEL_VF */
 581             felec            = _mm_mul_ps(velec,rinvsq{I}{J});
 582             /*                 #define INNERFLOPS INNERFLOPS+2 */
 583             /*             #endif */
 584
 585             /*         #elif KERNEL_ELEC=='ReactionField' */
 586
 587             /* REACTION-FIELD ELECTROSTATICS */
 588             /*             #if 'Potential' in KERNEL_VF */
 589             velec            = _mm_mul_ps(qq{I}{J},_mm_sub_ps(_mm_macc_ps(krf,rsq{I}{J},rinv{I}{J}),crf));
 590             /*                 #define INNERFLOPS INNERFLOPS+4 */
 591             /*             #endif */
 592             /*             #if 'Force' in KERNEL_VF */
 593             felec            = _mm_mul_ps(qq{I}{J},_mm_msub_ps(rinv{I}{J},rinvsq{I}{J},krf2));
 594             /*                 #define INNERFLOPS INNERFLOPS+3 */
 595             /*             #endif */
 596
 597             /*         #elif KERNEL_ELEC=='Ewald' */
 598             /* EWALD ELECTROSTATICS */
 599
 600             /* Analytical PME correction */
 601             zeta2            = _mm_mul_ps(beta2,rsq{I}{J});
 602             /*             #if 'Force' in KERNEL_VF */
 603             rinv3            = _mm_mul_ps(rinvsq{I}{J},rinv{I}{J});
 604             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 605             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 606             felec            = _mm_mul_ps(qq{I}{J},felec);
 607             /*             #endif */
 608             /*             #if 'Potential' in KERNEL_VF or KERNEL_MOD_ELEC=='PotentialSwitch' */
 609             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 610             /*                 #if KERNEL_MOD_ELEC=='PotentialShift' */
 611             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv{I}{J},sh_ewald));
 612             /*                 #else */
 613             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv{I}{J});
 614             /*                 #endif */
 615             velec            = _mm_mul_ps(qq{I}{J},velec);
 616             /*             #endif */
 617
 618             /*         #elif KERNEL_ELEC=='CubicSplineTable' */
 619
 620             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 621             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 622             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 623             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 624             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 625             _MM_TRANSPOSE4_PS(Y,F,G,H);
 626             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 627             /*             #define INNERFLOPS INNERFLOPS+4 */
 628             /*             #if 'Potential' in KERNEL_VF */
 629             VV               = _mm_macc_ps(vfeps,Fp,Y);
 630             velec            = _mm_mul_ps(qq{I}{J},VV);
 631             /*                 #define INNERFLOPS INNERFLOPS+3 */
 632             /*             #endif */
 633             /*             #if 'Force' in KERNEL_VF */
 634             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 635             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq{I}{J},FF),_mm_mul_ps(vftabscale,rinv{I}{J})));
 636             /*                 #define INNERFLOPS INNERFLOPS+7 */
 637             /*             #endif */
 638             /*         #endif */
 639             /*         ## End of check for electrostatics interaction forms */
 640             /*     #endif */
 641             /*     ## END OF ELECTROSTATIC INTERACTION CHECK FOR PAIR I-J */
 642
 643             /*     #if 'vdw' in INTERACTION_FLAGS[I][J] */
 644
 645             /*         #if KERNEL_VDW=='LennardJones' */
 646
 647             /* LENNARD-JONES DISPERSION/REPULSION */
 648
 649             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq{I}{J},rinvsq{I}{J}),rinvsq{I}{J});
 650             /*             #define INNERFLOPS INNERFLOPS+2 */
 651             /*             #if 'Potential' in KERNEL_VF or KERNEL_MOD_VDW=='PotentialSwitch' */
 652             vvdw6            = _mm_mul_ps(c6_{I}{J},rinvsix);
 653             vvdw12           = _mm_mul_ps(c12_{I}{J},_mm_mul_ps(rinvsix,rinvsix));
 654             /*                 #define INNERFLOPS INNERFLOPS+3 */
 655             /*                 #if KERNEL_MOD_VDW=='PotentialShift' */
 656             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_{I}{J},_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 657                                           _mm_mul_ps( _mm_nmacc_ps(c6_{I}{J},sh_vdw_invrcut6,vvdw6),one_sixth));
 658             /*                     #define INNERFLOPS INNERFLOPS+8 */
 659             /*                 #else */
 660             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 661             /*                     #define INNERFLOPS INNERFLOPS+3 */
 662             /*                 #endif */
 663             /*                 ## Check for force inside potential check, i.e. this means we already did the potential part */
 664             /*                 #if 'Force' in KERNEL_VF */
 665             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq{I}{J});
 666             /*                     #define INNERFLOPS INNERFLOPS+2 */
 667             /*                 #endif */
 668             /*             #elif KERNEL_VF=='Force' */
 669             /*                 ## Force-only LennardJones makes it possible to save 1 flop (they do add up...) */
 670             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_{I}{J},rinvsix,c6_{I}{J}),_mm_mul_ps(rinvsix,rinvsq{I}{J}));
 671             /*                 #define INNERFLOPS INNERFLOPS+4 */
 672             /*             #endif */
 673
 674             /*         #elif KERNEL_VDW=='CubicSplineTable' */
 675
 676             /* CUBIC SPLINE TABLE DISPERSION */
 677             /*             #if 'Table' in KERNEL_ELEC */
 678             vfitab           = _mm_add_epi32(vfitab,ifour);
 679             /*             #endif                     */
 680             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 681             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 682             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 683             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 684             _MM_TRANSPOSE4_PS(Y,F,G,H);
 685             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 686             /*             #define INNERFLOPS INNERFLOPS+4 */
 687             /*             #if 'Potential' in KERNEL_VF */
 688             VV               = _mm_macc_ps(vfeps,Fp,Y);
 689             vvdw6            = _mm_mul_ps(c6_{I}{J},VV);
 690             /*                 #define INNERFLOPS INNERFLOPS+3 */
 691             /*             #endif */
 692             /*             #if 'Force' in KERNEL_VF */
 693             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 694             fvdw6            = _mm_mul_ps(c6_{I}{J},FF);
 695             /*                 #define INNERFLOPS INNERFLOPS+4 */
 696             /*             #endif */
 697
 698             /* CUBIC SPLINE TABLE REPULSION */
 699             vfitab           = _mm_add_epi32(vfitab,ifour);
 700             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 701             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 702             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 703             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 704             _MM_TRANSPOSE4_PS(Y,F,G,H);
 705             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 706             /*             #define INNERFLOPS INNERFLOPS+4 */
 707             /*             #if 'Potential' in KERNEL_VF */
 708             VV               = _mm_macc_ps(vfeps,Fp,Y);
 709             vvdw12           = _mm_mul_ps(c12_{I}{J},VV);
 710             /*                 #define INNERFLOPS INNERFLOPS+3 */
 711             /*             #endif */
 712             /*             #if 'Force' in KERNEL_VF */
 713             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 714             fvdw12           = _mm_mul_ps(c12_{I}{J},FF);
 715             /*                 #define INNERFLOPS INNERFLOPS+5 */
 716             /*             #endif */
 717             /*             #if 'Potential' in KERNEL_VF */
 718             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 719             /*                 #define INNERFLOPS INNERFLOPS+1 */
 720             /*             #endif */
 721             /*             #if 'Force' in KERNEL_VF */
 722             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv{I}{J})));
 723             /*                 #define INNERFLOPS INNERFLOPS+4 */
 724             /*             #endif */
 725
 726             /*         #elif KERNEL_VDW=='LJEwald' */
 727
 728             /* Analytical LJ-PME */
 729             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq{I}{J},rinvsq{I}{J}),rinvsq{I}{J});
 730             ewcljrsq         = _mm_mul_ps(ewclj2,rsq{I}{J});
 731             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 732             exponent         = avx128fma_exp_f(ewcljrsq);
 733             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 734             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 735             /*                 #define INNERFLOPS INNERFLOPS+10 */
 736             /*             #if 'Potential' in KERNEL_VF or KERNEL_MOD_VDW=='PotentialSwitch' */
 737             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 738             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_{I}{J},_mm_sub_ps(one,poly),c6_{I}{J}),rinvsix);
 739             vvdw12           = _mm_mul_ps(c12_{I}{J},_mm_mul_ps(rinvsix,rinvsix));
 740             /*                     #define INNERFLOPS INNERFLOPS+5 */
 741             /*                 #if KERNEL_MOD_VDW=='PotentialShift' */
 742             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_{I}{J},_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 743                                           _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));
 744             /*                     #define INNERFLOPS INNERFLOPS+7 */
 745             /*                 #else */
 746             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 747             /*                     #define INNERFLOPS INNERFLOPS+2 */
 748             /*                 #endif */
 749             /*                  ## Check for force inside potential check, i.e. this means we already did the potential part */
 750             /*                  #if 'Force' in KERNEL_VF */
 751             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 752             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});
 753             /*                 #define INNERFLOPS INNERFLOPS+5 */
 754             /*                  #endif */
 755             /*              #elif KERNEL_VF=='Force' */
 756             /* f6A = 6 * C6grid * (1 - poly) */
 757             f6A              = _mm_mul_ps(c6grid_{I}{J},_mm_sub_ps(one,poly));
 758             /* f6B = C6grid * exponent * beta^6 */
 759             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_{I}{J},one_sixth),_mm_mul_ps(exponent,ewclj6));
 760             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 761             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});
 762             /*                 #define INNERFLOPS INNERFLOPS+10 */
 763             /*              #endif */
 764             /*         #endif */
 765             /*         ## End of check for vdw interaction forms */
 766             /*     #endif */
 767             /*     ## END OF VDW INTERACTION CHECK FOR PAIR I-J */
 768
 769             /*     #if 'switch' in INTERACTION_FLAGS[I][J] */
 770             d                = _mm_sub_ps(r{I}{J},rswitch);
 771             d                = _mm_max_ps(d,_mm_setzero_ps());
 772             d2               = _mm_mul_ps(d,d);
 773             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 774             /*         #define INNERFLOPS INNERFLOPS+10 */
 775
 776             /*         #if 'Force' in KERNEL_VF */
 777             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 778             /*             #define INNERFLOPS INNERFLOPS+5 */
 779             /*         #endif */
 780
 781             /* Evaluate switch function */
 782             /*         #if 'Force' in KERNEL_VF */
 783             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 784             /*             #if 'electrostatics' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_ELEC=='PotentialSwitch' */
 785             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv{I}{J},_mm_mul_ps(velec,dsw)) );
 786             /*                 #define INNERFLOPS INNERFLOPS+4 */
 787             /*             #endif */
 788             /*             #if 'vdw' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_VDW=='PotentialSwitch' */
 789             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv{I}{J},_mm_mul_ps(vvdw,dsw)) );
 790             /*                 #define INNERFLOPS INNERFLOPS+4 */
 791             /*             #endif */
 792             /*         #endif */
 793             /*         #if 'Potential' in KERNEL_VF */
 794             /*             #if 'electrostatics' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_ELEC=='PotentialSwitch' */
 795             velec            = _mm_mul_ps(velec,sw);
 796             /*                 #define INNERFLOPS INNERFLOPS+1 */
 797             /*             #endif */
 798             /*             #if 'vdw' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_VDW=='PotentialSwitch' */
 799             vvdw             = _mm_mul_ps(vvdw,sw);
 800             /*                 #define INNERFLOPS INNERFLOPS+1 */
 801             /*             #endif */
 802             /*         #endif */
 803             /*     #endif */
 804             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 805             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 806             cutoff_mask      = _mm_cmplt_ps(rsq{I}{J},rcutoff2);
 807             /*         #define INNERFLOPS INNERFLOPS+1 */
 808             /*     #endif */
 809
 810             /*     #if 'Potential' in KERNEL_VF */
 811             /* Update potential sum for this i atom from the interaction with this j atom. */
 812             /*         #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
 813             /*             #if 'exactcutoff' in INTERACTION_FLAGS[I][J] */
 814             velec            = _mm_and_ps(velec,cutoff_mask);
 815             /*                 #define INNERFLOPS INNERFLOPS+1 */
 816             /*             #endif                                       */
 817             /*             #if ROUND == 'Epilogue' */
 818             velec            = _mm_andnot_ps(dummy_mask,velec);
 819             /*             #endif */
 820             velecsum         = _mm_add_ps(velecsum,velec);
 821             /*             #define INNERFLOPS INNERFLOPS+1 */
 822             /*         #endif */
 823             /*         #if 'vdw' in INTERACTION_FLAGS[I][J] */
 824             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 825             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 826             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 827             /*                 #define INNERFLOPS INNERFLOPS+1 */
 828             /*             #endif                                       */
 829             /*             #if ROUND == 'Epilogue' */
 830             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 831             /*             #endif */
 832             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 833             /*             #define INNERFLOPS INNERFLOPS+1 */
 834             /*         #endif */
 835             /*     #endif */
 836
 837             /*     #if 'Force' in KERNEL_VF */
 838
 839             /*         #if 'electrostatics' in INTERACTION_FLAGS[I][J] and 'vdw' in INTERACTION_FLAGS[I][J] */
 840             fscal            = _mm_add_ps(felec,fvdw);
 841             /*             #define INNERFLOPS INNERFLOPS+1 */
 842             /*         #elif 'electrostatics' in INTERACTION_FLAGS[I][J] */
 843             fscal            = felec;
 844             /*         #elif 'vdw' in INTERACTION_FLAGS[I][J] */
 845             fscal            = fvdw;
 846             /*        #endif */
 847
 848             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 849             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 850             fscal            = _mm_and_ps(fscal,cutoff_mask);
 851             /*                 #define INNERFLOPS INNERFLOPS+1 */
 852             /*             #endif                                       */
 853
 854             /*             #if ROUND == 'Epilogue' */
 855             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 856             /*             #endif */
 857
 858             /* ## Construction of vectorial force built into FMA instructions now */
 859             /* #define INNERFLOPS INNERFLOPS+3      */
 860
 861              /* Update vectorial force */
 862             fix{I}             = _mm_macc_ps(dx{I}{J},fscal,fix{I});
 863             fiy{I}             = _mm_macc_ps(dy{I}{J},fscal,fiy{I});
 864             fiz{I}             = _mm_macc_ps(dz{I}{J},fscal,fiz{I});
 865             /*             #define INNERFLOPS INNERFLOPS+6 */
 866
 867             /* #if GEOMETRY_I == 'Particle'             */
 868             /*     #if ROUND == 'Loop' */
 869             fjptrA             = f+j_coord_offsetA;
 870             fjptrB             = f+j_coord_offsetB;
 871             fjptrC             = f+j_coord_offsetC;
 872             fjptrD             = f+j_coord_offsetD;
 873             /*     #else */
 874             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 875             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 876             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 877             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 878             /*     #endif */
 879             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 880                                                    _mm_mul_ps(dx{I}{J},fscal),
 881                                                    _mm_mul_ps(dy{I}{J},fscal),
 882                                                    _mm_mul_ps(dz{I}{J},fscal));
 883             /*     #define INNERFLOPS INNERFLOPS+3      */
 884             /* #else                                    */
 885             fjx{J}             = _mm_macc_ps(dx{I}{J},fscal,fjx{J});
 886             fjy{J}             = _mm_macc_ps(dy{I}{J},fscal,fjy{J});
 887             fjz{J}             = _mm_macc_ps(dz{I}{J},fscal,fjz{J});
 888             /*     #define INNERFLOPS INNERFLOPS+3      */
 889             /* #endif                                   */
 890
 891             /*     #endif */
 892
 893             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 894             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 895             /*         #if 0    ## This and next two lines is a hack to maintain indentation in template file */
 896             {
 897                 /*     #endif */
 898             }
 899             /*     #endif */
 900             /*    ## End of check for the interaction being outside the cutoff */
 901
 902             /* #endfor */
 903             /* ## End of loop over i-j interaction pairs */
 904
 905             /* #if GEOMETRY_I != 'Particle' */
 906             /*     #if ROUND == 'Loop' */
 907             fjptrA             = f+j_coord_offsetA;
 908             fjptrB             = f+j_coord_offsetB;
 909             fjptrC             = f+j_coord_offsetC;
 910             fjptrD             = f+j_coord_offsetD;
 911             /*     #else */
 912             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 913             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 914             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 915             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 916             /*     #endif */
 917             /* #endif */
 918
 919             /* #if 'Water' in GEOMETRY_I and GEOMETRY_J == 'Particle' */
 920             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 921             /* #elif GEOMETRY_J == 'Water3'               */
 922             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 923                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 924             /*     #define INNERFLOPS INNERFLOPS+9      */
 925             /* #elif GEOMETRY_J == 'Water4'             */
 926             /*     #if 0 in PARTICLES_J                 */
 927             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 928                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 929                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 930             /*     #define INNERFLOPS INNERFLOPS+12     */
 931             /*     #else                                */
 932             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 933                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 934             /*     #define INNERFLOPS INNERFLOPS+9      */
 935             /*     #endif                               */
 936             /* #endif                                   */
 937
 938             /* Inner loop uses {INNERFLOPS} flops */
 939         }
 940
 941         /* #endfor */
 942
 943         /* End of innermost loop */
 944
 945         /* #if 'Force' in KERNEL_VF */
 946         /*     #if GEOMETRY_I == 'Particle'            */
 947         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 948                                               f+i_coord_offset,fshift+i_shift_offset);
 949         /*         #define OUTERFLOPS OUTERFLOPS+6     */
 950         /*     #elif GEOMETRY_I == 'Water3'            */
 951         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 952                                               f+i_coord_offset,fshift+i_shift_offset);
 953         /*         #define OUTERFLOPS OUTERFLOPS+18    */
 954         /*     #elif GEOMETRY_I == 'Water4'            */
 955         /*         #if 0 in PARTICLES_I                */
 956         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 957                                               f+i_coord_offset,fshift+i_shift_offset);
 958         /*             #define OUTERFLOPS OUTERFLOPS+24    */
 959         /*         #else                               */
 960         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 961                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 962         /*             #define OUTERFLOPS OUTERFLOPS+18    */
 963         /*         #endif                              */
 964         /*     #endif                                  */
 965         /* #endif                                      */
 966
 967         /* #if 'Potential' in KERNEL_VF */
 968         ggid                        = gid[iidx];
 969         /* Update potential energies */
 970         /*     #if KERNEL_ELEC != 'None' */
 971         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 972         /*         #define OUTERFLOPS OUTERFLOPS+1 */
 973         /*     #endif */
 974         /*     #if KERNEL_VDW != 'None' */
 975         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 976         /*         #define OUTERFLOPS OUTERFLOPS+1 */
 977         /*     #endif */
 978         /* #endif */
 979
 980         /* Increment number of inner iterations */
 981         inneriter                  += j_index_end - j_index_start;
 982
 983         /* Outer loop uses {OUTERFLOPS} flops */
 984     }
 985
 986     /* Increment number of outer iterations */
 987     outeriter        += nri;
 988
 989     /* Update outer/inner flops */
 990     /* ## NB: This is not important, it just affects the flopcount. However, since our preprocessor is */
 991     /* ## primitive and replaces aggressively even in strings inside these directives, we need to      */
 992     /* ## assemble the main part of the name (containing KERNEL/ELEC/VDW) directly in the source.      */
 993     /* #if GEOMETRY_I == 'Water3'            */
 994     /*     #define ISUFFIX '_W3'             */
 995     /* #elif GEOMETRY_I == 'Water4'          */
 996     /*     #define ISUFFIX '_W4'             */
 997     /* #else                                 */
 998     /*     #define ISUFFIX ''                */
 999     /* #endif                                */
1000     /* #if GEOMETRY_J == 'Water3'            */
1001     /*     #define JSUFFIX 'W3'              */
1002     /* #elif GEOMETRY_J == 'Water4'          */
1003     /*     #define JSUFFIX 'W4'              */
1004     /* #else                                 */
1005     /*     #define JSUFFIX ''                */
1006     /* #endif                                */
1007     /* #if 'PotentialAndForce' in KERNEL_VF  */
1008     /*     #define VFSUFFIX  '_VF'           */
1009     /* #elif 'Potential' in KERNEL_VF        */
1010     /*     #define VFSUFFIX '_V'             */
1011     /* #else                                 */
1012     /*     #define VFSUFFIX '_F'             */
1013     /* #endif                                */
1014
1015     /* #if KERNEL_ELEC != 'None' and KERNEL_VDW != 'None' */
1016     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW{ISUFFIX}{JSUFFIX}{VFSUFFIX},outeriter*{OUTERFLOPS} + inneriter*{INNERFLOPS});
1017     /* #elif KERNEL_ELEC != 'None' */
1018     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC{ISUFFIX}{JSUFFIX}{VFSUFFIX},outeriter*{OUTERFLOPS} + inneriter*{INNERFLOPS});
1019     /* #else */
1020     inc_nrnb(nrnb,eNR_NBKERNEL_VDW{ISUFFIX}{JSUFFIX}{VFSUFFIX},outeriter*{OUTERFLOPS} + inneriter*{INNERFLOPS});
1021     /* #endif  */
1022 }