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, 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 "config.h"
  40
  41 #include <math.h>
  42
  43 #include "../nb_kernel.h"
  44 #include "types/simple.h"
  45 #include "gromacs/math/vec.h"
  46 #include "nrnb.h"
  47
  48 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  49 #include "kernelutil_x86_avx_128_fma_single.h"
  50 /* #endif */
  51
  52 /* ## List of variables set by the generating script:                                    */
  53 /* ##                                                                                    */
  54 /* ## Setttings that apply to the entire kernel:                                         */
  55 /* ## KERNEL_ELEC:           String, choice for electrostatic interactions               */
  56 /* ## KERNEL_VDW:            String, choice for van der Waals interactions               */
  57 /* ## KERNEL_NAME:           String, name of this kernel                                 */
  58 /* ## KERNEL_VF:             String telling if we calculate potential, force, or both    */
  59 /* ## GEOMETRY_I/GEOMETRY_J: String, name of each geometry, e.g. 'Water3' or '1Particle' */
  60 /* ##                                                                                    */
  61 /* ## Setttings that apply to particles in the outer (I) or inner (J) loops:             */
  62 /* ## PARTICLES_I[]/         Arrays with lists of i/j particles to use in kernel. It is  */
  63 /* ## PARTICLES_J[]:         just [0] for particle geometry, but can be longer for water */
  64 /* ## PARTICLES_ELEC_I[]/    Arrays with lists of i/j particle that have electrostatics  */
  65 /* ## PARTICLES_ELEC_J[]:    interactions that should be calculated in this kernel.      */
  66 /* ## PARTICLES_VDW_I[]/     Arrays with the list of i/j particle that have VdW          */
  67 /* ## PARTICLES_VDW_J[]:     interactions that should be calculated in this kernel.      */
  68 /* ##                                                                                    */
  69 /* ## Setttings for pairs of interactions (e.g. 2nd i particle against 1st j particle)   */
  70 /* ## PAIRS_IJ[]:            Array with (i,j) tuples of pairs for which interactions     */
  71 /* ##                        should be calculated in this kernel. Zero-charge particles  */
  72 /* ##                        do not have interactions with particles without vdw, and    */
  73 /* ##                        Vdw-only interactions are not evaluated in a no-vdw-kernel. */
  74 /* ## INTERACTION_FLAGS[][]: 2D matrix, dimension e.g. 3*3 for water-water interactions. */
  75 /* ##                        For each i-j pair, the element [I][J] is a list of strings  */
  76 /* ##                        defining properties/flags of this interaction. Examples     */
  77 /* ##                        include 'electrostatics'/'vdw' if that type of interaction  */
  78 /* ##                        should be evaluated, 'rsq'/'rinv'/'rinvsq' if those values  */
  79 /* ##                        are needed, and 'exactcutoff' or 'shift','switch' to        */
  80 /* ##                        decide if the force/potential should be modified. This way  */
  81 /* ##                        we only calculate values absolutely needed for each case.   */
  82
  83 /* ## Calculate the size and offset for (merged/interleaved) table data */
  84
  85 /*
  86  * Gromacs nonbonded kernel:   {KERNEL_NAME}
  87  * Electrostatics interaction: {KERNEL_ELEC}
  88  * VdW interaction:            {KERNEL_VDW}
  89  * Geometry:                   {GEOMETRY_I}-{GEOMETRY_J}
  90  * Calculate force/pot:        {KERNEL_VF}
  91  */
  92 void
  93 {KERNEL_NAME}
  94                     (t_nblist                    * gmx_restrict       nlist,
  95                      rvec                        * gmx_restrict          xx,
  96                      rvec                        * gmx_restrict          ff,
  97                      t_forcerec                  * gmx_restrict          fr,
  98                      t_mdatoms                   * gmx_restrict     mdatoms,
  99                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 100                      t_nrnb                      * gmx_restrict        nrnb)
 101 {
 102     /* ## Not all variables are used for all kernels, but any optimizing compiler fixes that, */
 103     /* ## so there is no point in going to extremes to exclude variables that are not needed. */
 104     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 105      * just 0 for non-waters.
 106      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 107      * jnr indices corresponding to data put in the four positions in the SIMD register.
 108      */
 109     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 110     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 111     int              jnrA,jnrB,jnrC,jnrD;
 112     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 113     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 114     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 115     real             rcutoff_scalar;
 116     real             *shiftvec,*fshift,*x,*f;
 117     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 118     real             scratch[4*DIM];
 119     __m128           fscal,rcutoff,rcutoff2,jidxall;
 120     /* #for I in PARTICLES_I */
 121     int              vdwioffset{I};
 122     __m128           ix{I},iy{I},iz{I},fix{I},fiy{I},fiz{I},iq{I},isai{I};
 123     /* #endfor */
 124     /* #for J in PARTICLES_J */
 125     int              vdwjidx{J}A,vdwjidx{J}B,vdwjidx{J}C,vdwjidx{J}D;
 126     __m128           jx{J},jy{J},jz{J},fjx{J},fjy{J},fjz{J},jq{J},isaj{J};
 127     /* #endfor */
 128     /* #for I,J in PAIRS_IJ */
 129     __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};
 130     /* #endfor */
 131     /* #if KERNEL_ELEC != 'None' */
 132     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 133     real             *charge;
 134     /* #endif */
 135     /* #if 'GeneralizedBorn' in KERNEL_ELEC */
 136     __m128i          gbitab;
 137     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
 138     __m128           minushalf = _mm_set1_ps(-0.5);
 139     real             *invsqrta,*dvda,*gbtab;
 140     /* #endif */
 141     /* #if KERNEL_VDW != 'None' */
 142     int              nvdwtype;
 143     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 144     int              *vdwtype;
 145     real             *vdwparam;
 146     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 147     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 148     /* #endif */
 149     /* #if 'Table' in KERNEL_ELEC or 'GeneralizedBorn' in KERNEL_ELEC or 'Table' in KERNEL_VDW */
 150     __m128i          vfitab;
 151     __m128i          ifour       = _mm_set1_epi32(4);
 152     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 153     real             *vftab;
 154     /* #endif */
 155     /* #if 'LJEwald' in KERNEL_VDW */
 156     /* #for I,J in PAIRS_IJ */
 157     __m128           c6grid_{I}{J};
 158     /* #endfor */
 159     real             *vdwgridparam;
 160     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 161     __m128           one_half = _mm_set1_ps(0.5);
 162     __m128           minus_one = _mm_set1_ps(-1.0);
 163     /* #endif */
 164     /* #if 'Ewald' in KERNEL_ELEC */
 165     __m128i          ewitab;
 166     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 167     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 168     real             *ewtab;
 169     /* #endif */
 170     /* #if 'PotentialSwitch' in [KERNEL_MOD_ELEC,KERNEL_MOD_VDW] */
 171     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 172     real             rswitch_scalar,d_scalar;
 173     /* #endif */
 174     __m128           dummy_mask,cutoff_mask;
 175     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 176     __m128           one     = _mm_set1_ps(1.0);
 177     __m128           two     = _mm_set1_ps(2.0);
 178     x                = xx[0];
 179     f                = ff[0];
 180
 181     nri              = nlist->nri;
 182     iinr             = nlist->iinr;
 183     jindex           = nlist->jindex;
 184     jjnr             = nlist->jjnr;
 185     shiftidx         = nlist->shift;
 186     gid              = nlist->gid;
 187     shiftvec         = fr->shift_vec[0];
 188     fshift           = fr->fshift[0];
 189     /* #if KERNEL_ELEC != 'None' */
 190     facel            = _mm_set1_ps(fr->epsfac);
 191     charge           = mdatoms->chargeA;
 192     /*     #if 'ReactionField' in KERNEL_ELEC */
 193     krf              = _mm_set1_ps(fr->ic->k_rf);
 194     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 195     crf              = _mm_set1_ps(fr->ic->c_rf);
 196     /*     #endif */
 197     /* #endif */
 198     /* #if KERNEL_VDW != 'None' */
 199     nvdwtype         = fr->ntype;
 200     vdwparam         = fr->nbfp;
 201     vdwtype          = mdatoms->typeA;
 202     /* #endif */
 203     /* #if 'LJEwald' in KERNEL_VDW */
 204     vdwgridparam     = fr->ljpme_c6grid;
 205     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 206     ewclj            = _mm_set1_ps(fr->ewaldcoeff_lj);
 207     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 208     /* #endif */
 209
 210     /* #if 'Table' in KERNEL_ELEC and 'Table' in KERNEL_VDW */
 211     vftab            = kernel_data->table_elec_vdw->data;
 212     vftabscale       = _mm_set1_ps(kernel_data->table_elec_vdw->scale);
 213     /* #elif 'Table' in KERNEL_ELEC */
 214     vftab            = kernel_data->table_elec->data;
 215     vftabscale       = _mm_set1_ps(kernel_data->table_elec->scale);
 216     /* #elif 'Table' in KERNEL_VDW */
 217     vftab            = kernel_data->table_vdw->data;
 218     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 219     /* #endif */
 220
 221     /* #if 'Ewald' in KERNEL_ELEC */
 222     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 223     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 224     beta2            = _mm_mul_ps(beta,beta);
 225     beta3            = _mm_mul_ps(beta,beta2);
 226     /*     #if KERNEL_VF=='Force' and KERNEL_MOD_ELEC!='PotentialSwitch' */
 227     ewtab            = fr->ic->tabq_coul_F;
 228     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 229     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 230     /*     #else */
 231     ewtab            = fr->ic->tabq_coul_FDV0;
 232     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 233     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 234      /*     #endif */
 235     /* #endif */
 236
 237     /* #if KERNEL_ELEC=='GeneralizedBorn' */
 238     invsqrta         = fr->invsqrta;
 239     dvda             = fr->dvda;
 240     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 241     gbtab            = fr->gbtab.data;
 242     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 243     /* #endif */
 244
 245     /* #if 'Water' in GEOMETRY_I */
 246     /* Setup water-specific parameters */
 247     inr              = nlist->iinr[0];
 248     /*     #for I in PARTICLES_ELEC_I */
 249     iq{I}              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+{I}]));
 250     /*     #endfor */
 251     /*     #for I in PARTICLES_VDW_I */
 252     vdwioffset{I}      = 2*nvdwtype*vdwtype[inr+{I}];
 253     /*     #endfor */
 254     /* #endif */
 255
 256     /* #if 'Water' in GEOMETRY_J */
 257     /*     #for J in PARTICLES_ELEC_J */
 258     jq{J}              = _mm_set1_ps(charge[inr+{J}]);
 259     /*     #endfor */
 260     /*     #for J in PARTICLES_VDW_J */
 261     vdwjidx{J}A        = 2*vdwtype[inr+{J}];
 262     /*     #endfor */
 263     /*     #for I,J in PAIRS_IJ */
 264     /*         #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
 265     qq{I}{J}             = _mm_mul_ps(iq{I},jq{J});
 266     /*         #endif */
 267     /*         #if 'vdw' in INTERACTION_FLAGS[I][J] */
 268     /*             #if 'LJEwald' in KERNEL_VDW */
 269     c6_{I}{J}            = _mm_set1_ps(vdwparam[vdwioffset{I}+vdwjidx{J}A]);
 270     c12_{I}{J}           = _mm_set1_ps(vdwparam[vdwioffset{I}+vdwjidx{J}A+1]);
 271     c6grid_{I}{J}        = _mm_set1_ps(vdwgridparam[vdwioffset{I}+vdwjidx{J}A]);
 272     /*             #else */
 273     c6_{I}{J}            = _mm_set1_ps(vdwparam[vdwioffset{I}+vdwjidx{J}A]);
 274     c12_{I}{J}           = _mm_set1_ps(vdwparam[vdwioffset{I}+vdwjidx{J}A+1]);
 275     /*             #endif */
 276     /*         #endif */
 277     /*     #endfor */
 278     /* #endif */
 279
 280     /* #if KERNEL_MOD_ELEC!='None' or KERNEL_MOD_VDW!='None' */
 281     /*     #if KERNEL_ELEC!='None' */
 282     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 283     rcutoff_scalar   = fr->rcoulomb;
 284     /*     #else */
 285     rcutoff_scalar   = fr->rvdw;
 286     /*     #endif */
 287     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 288     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 289     /* #endif */
 290
 291     /* #if KERNEL_MOD_VDW=='PotentialShift' */
 292     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 293     rvdw             = _mm_set1_ps(fr->rvdw);
 294     /* #endif */
 295
 296     /* #if 'PotentialSwitch' in [KERNEL_MOD_ELEC,KERNEL_MOD_VDW] */
 297     /*     #if KERNEL_MOD_ELEC=='PotentialSwitch'  */
 298     rswitch_scalar   = fr->rcoulomb_switch;
 299     rswitch          = _mm_set1_ps(rswitch_scalar);
 300     /*     #else */
 301     rswitch_scalar   = fr->rvdw_switch;
 302     rswitch          = _mm_set1_ps(rswitch_scalar);
 303     /*     #endif */
 304     /* Setup switch parameters */
 305     d_scalar         = rcutoff_scalar-rswitch_scalar;
 306     d                = _mm_set1_ps(d_scalar);
 307     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 308     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 309     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 310     /*     #if 'Force' in KERNEL_VF */
 311     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 312     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 313     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 314     /*     #endif */
 315     /* #endif */
 316
 317     /* Avoid stupid compiler warnings */
 318     jnrA = jnrB = jnrC = jnrD = 0;
 319     j_coord_offsetA = 0;
 320     j_coord_offsetB = 0;
 321     j_coord_offsetC = 0;
 322     j_coord_offsetD = 0;
 323
 324     /* ## Keep track of the floating point operations we issue for reporting! */
 325     /* #define OUTERFLOPS 0 */
 326     outeriter        = 0;
 327     inneriter        = 0;
 328
 329     for(iidx=0;iidx<4*DIM;iidx++)
 330     {
 331         scratch[iidx] = 0.0;
 332     }
 333
 334     /* Start outer loop over neighborlists */
 335     for(iidx=0; iidx<nri; iidx++)
 336     {
 337         /* Load shift vector for this list */
 338         i_shift_offset   = DIM*shiftidx[iidx];
 339
 340         /* Load limits for loop over neighbors */
 341         j_index_start    = jindex[iidx];
 342         j_index_end      = jindex[iidx+1];
 343
 344         /* Get outer coordinate index */
 345         inr              = iinr[iidx];
 346         i_coord_offset   = DIM*inr;
 347
 348         /* Load i particle coords and add shift vector */
 349         /* #if GEOMETRY_I == 'Particle' */
 350         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 351         /* #elif GEOMETRY_I == 'Water3' */
 352         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 353                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 354         /* #elif GEOMETRY_I == 'Water4' */
 355         /*     #if 0 in PARTICLES_I                 */
 356         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 357                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 358         /*     #else                                */
 359         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 360                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 361         /*     #endif                               */
 362         /* #endif                                   */
 363
 364         /* #if 'Force' in KERNEL_VF */
 365         /*     #for I in PARTICLES_I */
 366         fix{I}             = _mm_setzero_ps();
 367         fiy{I}             = _mm_setzero_ps();
 368         fiz{I}             = _mm_setzero_ps();
 369         /*     #endfor */
 370         /* #endif */
 371
 372         /* ## For water we already preloaded parameters at the start of the kernel */
 373         /* #if not 'Water' in GEOMETRY_I */
 374         /* Load parameters for i particles */
 375         /*     #for I in PARTICLES_ELEC_I */
 376         iq{I}              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+{I}));
 377         /*         #define OUTERFLOPS OUTERFLOPS+1 */
 378         /*         #if KERNEL_ELEC=='GeneralizedBorn' */
 379         isai{I}            = _mm_load1_ps(invsqrta+inr+{I});
 380         /*         #endif */
 381         /*     #endfor */
 382         /*     #for I in PARTICLES_VDW_I */
 383         vdwioffset{I}      = 2*nvdwtype*vdwtype[inr+{I}];
 384         /*     #endfor */
 385         /* #endif */
 386
 387         /* #if 'Potential' in KERNEL_VF */
 388         /* Reset potential sums */
 389         /*     #if KERNEL_ELEC != 'None' */
 390         velecsum         = _mm_setzero_ps();
 391         /*     #endif */
 392         /*     #if 'GeneralizedBorn' in KERNEL_ELEC */
 393         vgbsum           = _mm_setzero_ps();
 394         /*     #endif */
 395         /*     #if KERNEL_VDW != 'None' */
 396         vvdwsum          = _mm_setzero_ps();
 397         /*     #endif */
 398         /* #endif */
 399         /*     #if 'GeneralizedBorn' in KERNEL_ELEC and 'Force' in KERNEL_VF */
 400         dvdasum          = _mm_setzero_ps();
 401         /*     #endif */
 402
 403         /* #for ROUND in ['Loop','Epilogue'] */
 404
 405         /* #if ROUND =='Loop' */
 406         /* Start inner kernel loop */
 407         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 408         {
 409         /* ## First round is normal loop (next statement resets indentation) */
 410         /*     #if 0 */
 411         }
 412         /*     #endif */
 413         /* #else */
 414         if(jidx<j_index_end)
 415         {
 416         /* ## Second round is epilogue */
 417         /* #endif */
 418         /* #define INNERFLOPS 0 */
 419
 420             /* Get j neighbor index, and coordinate index */
 421             /* #if ROUND =='Loop' */
 422             jnrA             = jjnr[jidx];
 423             jnrB             = jjnr[jidx+1];
 424             jnrC             = jjnr[jidx+2];
 425             jnrD             = jjnr[jidx+3];
 426             /* #else */
 427             jnrlistA         = jjnr[jidx];
 428             jnrlistB         = jjnr[jidx+1];
 429             jnrlistC         = jjnr[jidx+2];
 430             jnrlistD         = jjnr[jidx+3];
 431             /* Sign of each element will be negative for non-real atoms.
 432              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 433              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 434              */
 435             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 436             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 437             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 438             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 439             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 440             /* #endif */
 441             j_coord_offsetA  = DIM*jnrA;
 442             j_coord_offsetB  = DIM*jnrB;
 443             j_coord_offsetC  = DIM*jnrC;
 444             j_coord_offsetD  = DIM*jnrD;
 445
 446             /* load j atom coordinates */
 447             /* #if GEOMETRY_J == 'Particle'             */
 448             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 449                                               x+j_coord_offsetC,x+j_coord_offsetD,
 450                                               &jx0,&jy0,&jz0);
 451             /* #elif GEOMETRY_J == 'Water3'             */
 452             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 453                                               x+j_coord_offsetC,x+j_coord_offsetD,
 454                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 455             /* #elif GEOMETRY_J == 'Water4'             */
 456             /*     #if 0 in PARTICLES_J                 */
 457             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 458                                               x+j_coord_offsetC,x+j_coord_offsetD,
 459                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 460                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 461             /*     #else                                */
 462             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 463                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 464                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 465             /*     #endif                               */
 466             /* #endif                                   */
 467
 468             /* Calculate displacement vector */
 469             /* #for I,J in PAIRS_IJ */
 470             dx{I}{J}             = _mm_sub_ps(ix{I},jx{J});
 471             dy{I}{J}             = _mm_sub_ps(iy{I},jy{J});
 472             dz{I}{J}             = _mm_sub_ps(iz{I},jz{J});
 473             /*     #define INNERFLOPS INNERFLOPS+3 */
 474             /* #endfor */
 475
 476             /* Calculate squared distance and things based on it */
 477             /* #for I,J in PAIRS_IJ */
 478             rsq{I}{J}            = gmx_mm_calc_rsq_ps(dx{I}{J},dy{I}{J},dz{I}{J});
 479             /*     #define INNERFLOPS INNERFLOPS+5 */
 480             /* #endfor */
 481
 482             /* #for I,J in PAIRS_IJ */
 483             /*     #if 'rinv' in INTERACTION_FLAGS[I][J] */
 484             rinv{I}{J}           = gmx_mm_invsqrt_ps(rsq{I}{J});
 485             /*         #define INNERFLOPS INNERFLOPS+5 */
 486             /*     #endif */
 487             /* #endfor */
 488
 489             /* #for I,J in PAIRS_IJ */
 490             /*     #if 'rinvsq' in INTERACTION_FLAGS[I][J] */
 491             /*         # if 'rinv' not in INTERACTION_FLAGS[I][J] */
 492             rinvsq{I}{J}         = gmx_mm_inv_ps(rsq{I}{J});
 493             /*             #define INNERFLOPS INNERFLOPS+4 */
 494             /*         #else */
 495             rinvsq{I}{J}         = _mm_mul_ps(rinv{I}{J},rinv{I}{J});
 496             /*             #define INNERFLOPS INNERFLOPS+1 */
 497             /*         #endif */
 498             /*     #endif */
 499             /* #endfor */
 500
 501             /* #if not 'Water' in GEOMETRY_J */
 502             /* Load parameters for j particles */
 503             /*     #for J in PARTICLES_ELEC_J */
 504             jq{J}              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+{J},charge+jnrB+{J},
 505                                                               charge+jnrC+{J},charge+jnrD+{J});
 506             /*         #if KERNEL_ELEC=='GeneralizedBorn' */
 507             isaj{J}            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+{J},invsqrta+jnrB+{J},
 508                                                               invsqrta+jnrC+{J},invsqrta+jnrD+{J});
 509             /*         #endif */
 510             /*     #endfor */
 511             /*     #for J in PARTICLES_VDW_J */
 512             vdwjidx{J}A        = 2*vdwtype[jnrA+{J}];
 513             vdwjidx{J}B        = 2*vdwtype[jnrB+{J}];
 514             vdwjidx{J}C        = 2*vdwtype[jnrC+{J}];
 515             vdwjidx{J}D        = 2*vdwtype[jnrD+{J}];
 516             /*     #endfor */
 517             /* #endif */
 518
 519             /* #if 'Force' in KERNEL_VF and not 'Particle' in GEOMETRY_I */
 520             /*     #for J in PARTICLES_J */
 521             fjx{J}             = _mm_setzero_ps();
 522             fjy{J}             = _mm_setzero_ps();
 523             fjz{J}             = _mm_setzero_ps();
 524             /*     #endfor */
 525             /* #endif */
 526
 527             /* #for I,J in PAIRS_IJ */
 528
 529             /**************************
 530              * CALCULATE INTERACTIONS *
 531              **************************/
 532
 533             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 534             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 535             /*         ## We always calculate rinv/rinvsq above to enable pipelineing in compilers (performance tested on x86) */
 536             if (gmx_mm_any_lt(rsq{I}{J},rcutoff2))
 537             {
 538                 /*     #if 0    ## this and the next two lines is a hack to maintain auto-indentation in template file */
 539             }
 540             /*         #endif */
 541             /*         #define INNERFLOPS INNERFLOPS+1 */
 542             /*     #endif */
 543
 544             /*     #if 'r' in INTERACTION_FLAGS[I][J] */
 545             r{I}{J}              = _mm_mul_ps(rsq{I}{J},rinv{I}{J});
 546             /*         #if ROUND == 'Epilogue' */
 547             r{I}{J}              = _mm_andnot_ps(dummy_mask,r{I}{J});
 548             /*             #define INNERFLOPS INNERFLOPS+1 */
 549             /*         #endif */
 550             /*         #define INNERFLOPS INNERFLOPS+1 */
 551             /*     #endif */
 552
 553             /*     ## For water geometries we already loaded parameters at the start of the kernel */
 554             /*     #if not 'Water' in GEOMETRY_J */
 555             /* Compute parameters for interactions between i and j atoms */
 556             /*         #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
 557             qq{I}{J}             = _mm_mul_ps(iq{I},jq{J});
 558             /*             #define INNERFLOPS INNERFLOPS+1 */
 559             /*         #endif */
 560             /*         #if 'vdw' in INTERACTION_FLAGS[I][J] */
 561             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset{I}+vdwjidx{J}A,
 562                                          vdwparam+vdwioffset{I}+vdwjidx{J}B,
 563                                          vdwparam+vdwioffset{I}+vdwjidx{J}C,
 564                                          vdwparam+vdwioffset{I}+vdwjidx{J}D,
 565                                          &c6_{I}{J},&c12_{I}{J});
 566
 567             /*           #if 'LJEwald' in KERNEL_VDW */
 568             c6grid_{I}{J}       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset{I}+vdwjidx{J}A,
 569                                                                vdwgridparam+vdwioffset{I}+vdwjidx{J}B,
 570                                                                vdwgridparam+vdwioffset{I}+vdwjidx{J}C,
 571                                                                vdwgridparam+vdwioffset{I}+vdwjidx{J}D);
 572             /*           #endif */
 573
 574             /*         #endif */
 575             /*     #endif */
 576
 577             /*     #if 'table' in INTERACTION_FLAGS[I][J] */
 578             /* Calculate table index by multiplying r with table scale and truncate to integer */
 579             rt               = _mm_mul_ps(r{I}{J},vftabscale);
 580             vfitab           = _mm_cvttps_epi32(rt);
 581 #ifdef __XOP__
 582             vfeps            = _mm_frcz_ps(rt);
 583 #else
 584             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 585 #endif
 586             twovfeps         = _mm_add_ps(vfeps,vfeps);
 587             /*         #define INNERFLOPS INNERFLOPS+4                          */
 588             /*         #if 'Table' in KERNEL_ELEC and 'Table' in KERNEL_VDW     */
 589             /*             ## 3 tables, 4 bytes per point: multiply index by 12 */
 590             vfitab           = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
 591             /*         #elif 'Table' in KERNEL_ELEC                             */
 592             /*             ## 1 table, 4 bytes per point: multiply index by 4   */
 593             vfitab           = _mm_slli_epi32(vfitab,2);
 594             /*         #elif 'Table' in KERNEL_VDW                              */
 595             /*             ## 2 tables, 4 bytes per point: multiply index by 8  */
 596             vfitab           = _mm_slli_epi32(vfitab,3);
 597             /*         #endif                                                   */
 598             /*     #endif */
 599
 600             /*     ## ELECTROSTATIC INTERACTIONS */
 601             /*     #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
 602
 603             /*         #if KERNEL_ELEC=='Coulomb' */
 604
 605             /* COULOMB ELECTROSTATICS */
 606             velec            = _mm_mul_ps(qq{I}{J},rinv{I}{J});
 607             /*             #define INNERFLOPS INNERFLOPS+1 */
 608             /*             #if 'Force' in KERNEL_VF */
 609             felec            = _mm_mul_ps(velec,rinvsq{I}{J});
 610             /*                 #define INNERFLOPS INNERFLOPS+2 */
 611             /*             #endif */
 612
 613             /*         #elif KERNEL_ELEC=='ReactionField' */
 614
 615             /* REACTION-FIELD ELECTROSTATICS */
 616             /*             #if 'Potential' in KERNEL_VF */
 617             velec            = _mm_mul_ps(qq{I}{J},_mm_sub_ps(_mm_macc_ps(krf,rsq{I}{J},rinv{I}{J}),crf));
 618             /*                 #define INNERFLOPS INNERFLOPS+4 */
 619             /*             #endif */
 620             /*             #if 'Force' in KERNEL_VF */
 621             felec            = _mm_mul_ps(qq{I}{J},_mm_msub_ps(rinv{I}{J},rinvsq{I}{J},krf2));
 622             /*                 #define INNERFLOPS INNERFLOPS+3 */
 623             /*             #endif */
 624
 625             /*         #elif KERNEL_ELEC=='GeneralizedBorn' */
 626
 627             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 628             isaprod          = _mm_mul_ps(isai{I},isaj{J});
 629             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq{I}{J},_mm_mul_ps(isaprod,gbinvepsdiff)));
 630             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 631             /*             #define INNERFLOPS INNERFLOPS+5 */
 632
 633             /* Calculate generalized born table index - this is a separate table from the normal one,
 634              * but we use the same procedure by multiplying r with scale and truncating to integer.
 635              */
 636             rt               = _mm_mul_ps(r{I}{J},gbscale);
 637             gbitab           = _mm_cvttps_epi32(rt);
 638 #ifdef __XOP__
 639             gbeps            = _mm_frcz_ps(rt);
 640 #else
 641             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 642 #endif
 643             gbitab           = _mm_slli_epi32(gbitab,2);
 644
 645             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 646             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 647             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 648             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 649             _MM_TRANSPOSE4_PS(Y,F,G,H);
 650             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 651             VV               = _mm_macc_ps(gbeps,Fp,Y);
 652             vgb              = _mm_mul_ps(gbqqfactor,VV);
 653             /*             #define INNERFLOPS INNERFLOPS+10 */
 654
 655             /*             #if 'Force' in KERNEL_VF */
 656             twogbeps         = _mm_add_ps(gbeps,gbeps);
 657             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 658             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 659             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r{I}{J},vgb));
 660             /*                 #if ROUND == 'Epilogue' */
 661             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 662             /*                 #endif */
 663             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 664             /*                 #if ROUND == 'Loop' */
 665             fjptrA           = dvda+jnrA;
 666             fjptrB           = dvda+jnrB;
 667             fjptrC           = dvda+jnrC;
 668             fjptrD           = dvda+jnrD;
 669             /*                 #else */
 670             /* 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. */
 671             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 672             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 673             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 674             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 675             /*                 #endif */
 676             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj{J},isaj{J})));
 677             /*                 #define INNERFLOPS INNERFLOPS+13 */
 678             /*             #endif */
 679             velec            = _mm_mul_ps(qq{I}{J},rinv{I}{J});
 680             /*                 #define INNERFLOPS INNERFLOPS+1 */
 681             /*             #if 'Force' in KERNEL_VF */
 682             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv{I}{J},fgb),rinv{I}{J});
 683             /*                 #define INNERFLOPS INNERFLOPS+3 */
 684             /*             #endif */
 685
 686             /*         #elif KERNEL_ELEC=='Ewald' */
 687             /* EWALD ELECTROSTATICS */
 688
 689             /* Analytical PME correction */
 690             zeta2            = _mm_mul_ps(beta2,rsq{I}{J});
 691             /*             #if 'Force' in KERNEL_VF */
 692             rinv3            = _mm_mul_ps(rinvsq{I}{J},rinv{I}{J});
 693             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 694             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 695             felec            = _mm_mul_ps(qq{I}{J},felec);
 696             /*             #endif */
 697             /*             #if 'Potential' in KERNEL_VF or KERNEL_MOD_ELEC=='PotentialSwitch' */
 698             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 699             /*                 #if KERNEL_MOD_ELEC=='PotentialShift' */
 700             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv{I}{J},sh_ewald));
 701             /*                 #else */
 702             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv{I}{J});
 703             /*                 #endif */
 704             velec            = _mm_mul_ps(qq{I}{J},velec);
 705             /*             #endif */
 706
 707             /*         #elif KERNEL_ELEC=='CubicSplineTable' */
 708
 709             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 710             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 711             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 712             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 713             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 714             _MM_TRANSPOSE4_PS(Y,F,G,H);
 715             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 716             /*             #define INNERFLOPS INNERFLOPS+4 */
 717             /*             #if 'Potential' in KERNEL_VF */
 718             VV               = _mm_macc_ps(vfeps,Fp,Y);
 719             velec            = _mm_mul_ps(qq{I}{J},VV);
 720             /*                 #define INNERFLOPS INNERFLOPS+3 */
 721             /*             #endif */
 722             /*             #if 'Force' in KERNEL_VF */
 723             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 724             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq{I}{J},FF),_mm_mul_ps(vftabscale,rinv{I}{J})));
 725             /*                 #define INNERFLOPS INNERFLOPS+7 */
 726             /*             #endif */
 727             /*         #endif */
 728             /*         ## End of check for electrostatics interaction forms */
 729             /*     #endif */
 730             /*     ## END OF ELECTROSTATIC INTERACTION CHECK FOR PAIR I-J */
 731
 732             /*     #if 'vdw' in INTERACTION_FLAGS[I][J] */
 733
 734             /*         #if KERNEL_VDW=='LennardJones' */
 735
 736             /* LENNARD-JONES DISPERSION/REPULSION */
 737
 738             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq{I}{J},rinvsq{I}{J}),rinvsq{I}{J});
 739             /*             #define INNERFLOPS INNERFLOPS+2 */
 740             /*             #if 'Potential' in KERNEL_VF or KERNEL_MOD_VDW=='PotentialSwitch' */
 741             vvdw6            = _mm_mul_ps(c6_{I}{J},rinvsix);
 742             vvdw12           = _mm_mul_ps(c12_{I}{J},_mm_mul_ps(rinvsix,rinvsix));
 743             /*                 #define INNERFLOPS INNERFLOPS+3 */
 744             /*                 #if KERNEL_MOD_VDW=='PotentialShift' */
 745             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_{I}{J},_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 746                                           _mm_mul_ps( _mm_nmacc_ps(c6_{I}{J},sh_vdw_invrcut6,vvdw6),one_sixth));
 747             /*                     #define INNERFLOPS INNERFLOPS+8 */
 748             /*                 #else */
 749             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 750             /*                     #define INNERFLOPS INNERFLOPS+3 */
 751             /*                 #endif */
 752             /*                 ## Check for force inside potential check, i.e. this means we already did the potential part */
 753             /*                 #if 'Force' in KERNEL_VF */
 754             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq{I}{J});
 755             /*                     #define INNERFLOPS INNERFLOPS+2 */
 756             /*                 #endif */
 757             /*             #elif KERNEL_VF=='Force' */
 758             /*                 ## Force-only LennardJones makes it possible to save 1 flop (they do add up...) */
 759             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_{I}{J},rinvsix,c6_{I}{J}),_mm_mul_ps(rinvsix,rinvsq{I}{J}));
 760             /*                 #define INNERFLOPS INNERFLOPS+4 */
 761             /*             #endif */
 762
 763             /*         #elif KERNEL_VDW=='CubicSplineTable' */
 764
 765             /* CUBIC SPLINE TABLE DISPERSION */
 766             /*             #if 'Table' in KERNEL_ELEC */
 767             vfitab           = _mm_add_epi32(vfitab,ifour);
 768             /*             #endif                     */
 769             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 770             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 771             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 772             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 773             _MM_TRANSPOSE4_PS(Y,F,G,H);
 774             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 775             /*             #define INNERFLOPS INNERFLOPS+4 */
 776             /*             #if 'Potential' in KERNEL_VF */
 777             VV               = _mm_macc_ps(vfeps,Fp,Y);
 778             vvdw6            = _mm_mul_ps(c6_{I}{J},VV);
 779             /*                 #define INNERFLOPS INNERFLOPS+3 */
 780             /*             #endif */
 781             /*             #if 'Force' in KERNEL_VF */
 782             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 783             fvdw6            = _mm_mul_ps(c6_{I}{J},FF);
 784             /*                 #define INNERFLOPS INNERFLOPS+4 */
 785             /*             #endif */
 786
 787             /* CUBIC SPLINE TABLE REPULSION */
 788             vfitab           = _mm_add_epi32(vfitab,ifour);
 789             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 790             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 791             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 792             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 793             _MM_TRANSPOSE4_PS(Y,F,G,H);
 794             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 795             /*             #define INNERFLOPS INNERFLOPS+4 */
 796             /*             #if 'Potential' in KERNEL_VF */
 797             VV               = _mm_macc_ps(vfeps,Fp,Y);
 798             vvdw12           = _mm_mul_ps(c12_{I}{J},VV);
 799             /*                 #define INNERFLOPS INNERFLOPS+3 */
 800             /*             #endif */
 801             /*             #if 'Force' in KERNEL_VF */
 802             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 803             fvdw12           = _mm_mul_ps(c12_{I}{J},FF);
 804             /*                 #define INNERFLOPS INNERFLOPS+5 */
 805             /*             #endif */
 806             /*             #if 'Potential' in KERNEL_VF */
 807             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 808             /*                 #define INNERFLOPS INNERFLOPS+1 */
 809             /*             #endif */
 810             /*             #if 'Force' in KERNEL_VF */
 811             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv{I}{J})));
 812             /*                 #define INNERFLOPS INNERFLOPS+4 */
 813             /*             #endif */
 814
 815             /*         #elif KERNEL_VDW=='LJEwald' */
 816
 817             /* Analytical LJ-PME */
 818             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq{I}{J},rinvsq{I}{J}),rinvsq{I}{J});
 819             ewcljrsq         = _mm_mul_ps(ewclj2,rsq{I}{J});
 820             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 821             exponent         = gmx_simd_exp_r(ewcljrsq);
 822             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 823             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 824             /*                 #define INNERFLOPS INNERFLOPS+10 */
 825             /*             #if 'Potential' in KERNEL_VF or KERNEL_MOD_VDW=='PotentialSwitch' */
 826             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 827             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_{I}{J},_mm_sub_ps(one,poly),c6_{I}{J}),rinvsix);
 828             vvdw12           = _mm_mul_ps(c12_{I}{J},_mm_mul_ps(rinvsix,rinvsix));
 829             /*                     #define INNERFLOPS INNERFLOPS+5 */
 830             /*                 #if KERNEL_MOD_VDW=='PotentialShift' */
 831             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_{I}{J},_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 832                                           _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));
 833             /*                     #define INNERFLOPS INNERFLOPS+7 */
 834             /*                 #else */
 835             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 836             /*                     #define INNERFLOPS INNERFLOPS+2 */
 837             /*                 #endif */
 838             /*                  ## Check for force inside potential check, i.e. this means we already did the potential part */
 839             /*                  #if 'Force' in KERNEL_VF */
 840             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 841             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});
 842             /*                 #define INNERFLOPS INNERFLOPS+5 */
 843             /*                  #endif */
 844             /*              #elif KERNEL_VF=='Force' */
 845             /* f6A = 6 * C6grid * (1 - poly) */
 846             f6A              = _mm_mul_ps(c6grid_{I}{J},_mm_sub_ps(one,poly));
 847             /* f6B = C6grid * exponent * beta^6 */
 848             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_{I}{J},one_sixth),_mm_mul_ps(exponent,ewclj6));
 849             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 850             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});
 851             /*                 #define INNERFLOPS INNERFLOPS+10 */
 852             /*              #endif */
 853             /*         #endif */
 854             /*         ## End of check for vdw interaction forms */
 855             /*     #endif */
 856             /*     ## END OF VDW INTERACTION CHECK FOR PAIR I-J */
 857
 858             /*     #if 'switch' in INTERACTION_FLAGS[I][J] */
 859             d                = _mm_sub_ps(r{I}{J},rswitch);
 860             d                = _mm_max_ps(d,_mm_setzero_ps());
 861             d2               = _mm_mul_ps(d,d);
 862             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 863             /*         #define INNERFLOPS INNERFLOPS+10 */
 864
 865             /*         #if 'Force' in KERNEL_VF */
 866             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 867             /*             #define INNERFLOPS INNERFLOPS+5 */
 868             /*         #endif */
 869
 870             /* Evaluate switch function */
 871             /*         #if 'Force' in KERNEL_VF */
 872             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 873             /*             #if 'electrostatics' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_ELEC=='PotentialSwitch' */
 874             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv{I}{J},_mm_mul_ps(velec,dsw)) );
 875             /*                 #define INNERFLOPS INNERFLOPS+4 */
 876             /*             #endif */
 877             /*             #if 'vdw' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_VDW=='PotentialSwitch' */
 878             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv{I}{J},_mm_mul_ps(vvdw,dsw)) );
 879             /*                 #define INNERFLOPS INNERFLOPS+4 */
 880             /*             #endif */
 881             /*         #endif */
 882             /*         #if 'Potential' in KERNEL_VF */
 883             /*             #if 'electrostatics' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_ELEC=='PotentialSwitch' */
 884             velec            = _mm_mul_ps(velec,sw);
 885             /*                 #define INNERFLOPS INNERFLOPS+1 */
 886             /*             #endif */
 887             /*             #if 'vdw' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_VDW=='PotentialSwitch' */
 888             vvdw             = _mm_mul_ps(vvdw,sw);
 889             /*                 #define INNERFLOPS INNERFLOPS+1 */
 890             /*             #endif */
 891             /*         #endif */
 892             /*     #endif */
 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             cutoff_mask      = _mm_cmplt_ps(rsq{I}{J},rcutoff2);
 896             /*         #define INNERFLOPS INNERFLOPS+1 */
 897             /*     #endif */
 898
 899             /*     #if 'Potential' in KERNEL_VF */
 900             /* Update potential sum for this i atom from the interaction with this j atom. */
 901             /*         #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
 902             /*             #if 'exactcutoff' in INTERACTION_FLAGS[I][J] */
 903             velec            = _mm_and_ps(velec,cutoff_mask);
 904             /*                 #define INNERFLOPS INNERFLOPS+1 */
 905             /*             #endif                                       */
 906             /*             #if ROUND == 'Epilogue' */
 907             velec            = _mm_andnot_ps(dummy_mask,velec);
 908             /*             #endif */
 909             velecsum         = _mm_add_ps(velecsum,velec);
 910             /*             #define INNERFLOPS INNERFLOPS+1 */
 911             /*             #if KERNEL_ELEC=='GeneralizedBorn' */
 912             /*             #if 'exactcutoff' in INTERACTION_FLAGS[I][J] */
 913             vgb              = _mm_and_ps(vgb,cutoff_mask);
 914             /*                 #define INNERFLOPS INNERFLOPS+1 */
 915             /*             #endif                                       */
 916             /*             #if ROUND == 'Epilogue' */
 917             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 918             /*             #endif */
 919             vgbsum           = _mm_add_ps(vgbsum,vgb);
 920             /*                 #define INNERFLOPS INNERFLOPS+1 */
 921             /*             #endif */
 922             /*         #endif */
 923             /*         #if 'vdw' in INTERACTION_FLAGS[I][J] */
 924             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 925             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 926             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 927             /*                 #define INNERFLOPS INNERFLOPS+1 */
 928             /*             #endif                                       */
 929             /*             #if ROUND == 'Epilogue' */
 930             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 931             /*             #endif */
 932             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 933             /*             #define INNERFLOPS INNERFLOPS+1 */
 934             /*         #endif */
 935             /*     #endif */
 936
 937             /*     #if 'Force' in KERNEL_VF */
 938
 939             /*         #if 'electrostatics' in INTERACTION_FLAGS[I][J] and 'vdw' in INTERACTION_FLAGS[I][J] */
 940             fscal            = _mm_add_ps(felec,fvdw);
 941             /*             #define INNERFLOPS INNERFLOPS+1 */
 942             /*         #elif 'electrostatics' in INTERACTION_FLAGS[I][J] */
 943             fscal            = felec;
 944             /*         #elif 'vdw' in INTERACTION_FLAGS[I][J] */
 945             fscal            = fvdw;
 946             /*        #endif */
 947
 948             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 949             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 950             fscal            = _mm_and_ps(fscal,cutoff_mask);
 951             /*                 #define INNERFLOPS INNERFLOPS+1 */
 952             /*             #endif                                       */
 953
 954             /*             #if ROUND == 'Epilogue' */
 955             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 956             /*             #endif */
 957
 958             /* ## Construction of vectorial force built into FMA instructions now */
 959             /* #define INNERFLOPS INNERFLOPS+3      */
 960
 961              /* Update vectorial force */
 962             fix{I}             = _mm_macc_ps(dx{I}{J},fscal,fix{I});
 963             fiy{I}             = _mm_macc_ps(dy{I}{J},fscal,fiy{I});
 964             fiz{I}             = _mm_macc_ps(dz{I}{J},fscal,fiz{I});
 965             /*             #define INNERFLOPS INNERFLOPS+6 */
 966
 967             /* #if GEOMETRY_I == 'Particle'             */
 968             /*     #if ROUND == 'Loop' */
 969             fjptrA             = f+j_coord_offsetA;
 970             fjptrB             = f+j_coord_offsetB;
 971             fjptrC             = f+j_coord_offsetC;
 972             fjptrD             = f+j_coord_offsetD;
 973             /*     #else */
 974             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 975             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 976             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 977             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 978             /*     #endif */
 979             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 980                                                    _mm_mul_ps(dx{I}{J},fscal),
 981                                                    _mm_mul_ps(dy{I}{J},fscal),
 982                                                    _mm_mul_ps(dz{I}{J},fscal));
 983             /*     #define INNERFLOPS INNERFLOPS+3      */
 984             /* #else                                    */
 985             fjx{J}             = _mm_macc_ps(dx{I}{J},fscal,fjx{J});
 986             fjy{J}             = _mm_macc_ps(dy{I}{J},fscal,fjy{J});
 987             fjz{J}             = _mm_macc_ps(dz{I}{J},fscal,fjz{J});
 988             /*     #define INNERFLOPS INNERFLOPS+3      */
 989             /* #endif                                   */
 990
 991             /*     #endif */
 992
 993             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 994             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 995             /*         #if 0    ## This and next two lines is a hack to maintain indentation in template file */
 996             {
 997                 /*     #endif */
 998             }
 999             /*     #endif */
1000             /*    ## End of check for the interaction being outside the cutoff */
1001
1002             /* #endfor */
1003             /* ## End of loop over i-j interaction pairs */
1004
1005             /* #if GEOMETRY_I != 'Particle' */
1006             /*     #if ROUND == 'Loop' */
1007             fjptrA             = f+j_coord_offsetA;
1008             fjptrB             = f+j_coord_offsetB;
1009             fjptrC             = f+j_coord_offsetC;
1010             fjptrD             = f+j_coord_offsetD;
1011             /*     #else */
1012             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1013             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1014             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1015             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1016             /*     #endif */
1017             /* #endif */
1018
1019             /* #if 'Water' in GEOMETRY_I and GEOMETRY_J == 'Particle' */
1020             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1021             /* #elif GEOMETRY_J == 'Water3'               */
1022             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1023                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1024             /*     #define INNERFLOPS INNERFLOPS+9      */
1025             /* #elif GEOMETRY_J == 'Water4'             */
1026             /*     #if 0 in PARTICLES_J                 */
1027             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1028                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1029                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1030             /*     #define INNERFLOPS INNERFLOPS+12     */
1031             /*     #else                                */
1032             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1033                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1034             /*     #define INNERFLOPS INNERFLOPS+9      */
1035             /*     #endif                               */
1036             /* #endif                                   */
1037
1038             /* Inner loop uses {INNERFLOPS} flops */
1039         }
1040
1041         /* #endfor */
1042
1043         /* End of innermost loop */
1044
1045         /* #if 'Force' in KERNEL_VF */
1046         /*     #if GEOMETRY_I == 'Particle'            */
1047         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
1048                                               f+i_coord_offset,fshift+i_shift_offset);
1049         /*         #define OUTERFLOPS OUTERFLOPS+6     */
1050         /*     #elif GEOMETRY_I == 'Water3'            */
1051         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1052                                               f+i_coord_offset,fshift+i_shift_offset);
1053         /*         #define OUTERFLOPS OUTERFLOPS+18    */
1054         /*     #elif GEOMETRY_I == 'Water4'            */
1055         /*         #if 0 in PARTICLES_I                */
1056         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1057                                               f+i_coord_offset,fshift+i_shift_offset);
1058         /*             #define OUTERFLOPS OUTERFLOPS+24    */
1059         /*         #else                               */
1060         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1061                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1062         /*             #define OUTERFLOPS OUTERFLOPS+18    */
1063         /*         #endif                              */
1064         /*     #endif                                  */
1065         /* #endif                                      */
1066
1067         /* #if 'Potential' in KERNEL_VF */
1068         ggid                        = gid[iidx];
1069         /* Update potential energies */
1070         /*     #if KERNEL_ELEC != 'None' */
1071         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1072         /*         #define OUTERFLOPS OUTERFLOPS+1 */
1073         /*     #endif */
1074         /*     #if 'GeneralizedBorn' in KERNEL_ELEC */
1075         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
1076         /*         #define OUTERFLOPS OUTERFLOPS+1 */
1077         /*     #endif */
1078         /*     #if KERNEL_VDW != 'None' */
1079         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1080         /*         #define OUTERFLOPS OUTERFLOPS+1 */
1081         /*     #endif */
1082         /* #endif */
1083         /*     #if 'GeneralizedBorn' in KERNEL_ELEC and 'Force' in KERNEL_VF */
1084         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai{I},isai{I}));
1085         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
1086         /*     #endif */
1087
1088         /* Increment number of inner iterations */
1089         inneriter                  += j_index_end - j_index_start;
1090
1091         /* Outer loop uses {OUTERFLOPS} flops */
1092     }
1093
1094     /* Increment number of outer iterations */
1095     outeriter        += nri;
1096
1097     /* Update outer/inner flops */
1098     /* ## NB: This is not important, it just affects the flopcount. However, since our preprocessor is */
1099     /* ## primitive and replaces aggressively even in strings inside these directives, we need to      */
1100     /* ## assemble the main part of the name (containing KERNEL/ELEC/VDW) directly in the source.      */
1101     /* #if GEOMETRY_I == 'Water3'            */
1102     /*     #define ISUFFIX '_W3'             */
1103     /* #elif GEOMETRY_I == 'Water4'          */
1104     /*     #define ISUFFIX '_W4'             */
1105     /* #else                                 */
1106     /*     #define ISUFFIX ''                */
1107     /* #endif                                */
1108     /* #if GEOMETRY_J == 'Water3'            */
1109     /*     #define JSUFFIX 'W3'              */
1110     /* #elif GEOMETRY_J == 'Water4'          */
1111     /*     #define JSUFFIX 'W4'              */
1112     /* #else                                 */
1113     /*     #define JSUFFIX ''                */
1114     /* #endif                                */
1115     /* #if 'PotentialAndForce' in KERNEL_VF  */
1116     /*     #define VFSUFFIX  '_VF'           */
1117     /* #elif 'Potential' in KERNEL_VF        */
1118     /*     #define VFSUFFIX '_V'             */
1119     /* #else                                 */
1120     /*     #define VFSUFFIX '_F'             */
1121     /* #endif                                */
1122
1123     /* #if KERNEL_ELEC != 'None' and KERNEL_VDW != 'None' */
1124     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW{ISUFFIX}{JSUFFIX}{VFSUFFIX},outeriter*{OUTERFLOPS} + inneriter*{INNERFLOPS});
1125     /* #elif KERNEL_ELEC != 'None' */
1126     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC{ISUFFIX}{JSUFFIX}{VFSUFFIX},outeriter*{OUTERFLOPS} + inneriter*{INNERFLOPS});
1127     /* #else */
1128     inc_nrnb(nrnb,eNR_NBKERNEL_VDW{ISUFFIX}{JSUFFIX}{VFSUFFIX},outeriter*{OUTERFLOPS} + inneriter*{INNERFLOPS});
1129     /* #endif  */
1130 }