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 "gmxpre.h"
  40
  41 #include "config.h"
  42
  43 #include <math.h>
  44
  45 #include "../nb_kernel.h"
  46 #include "gromacs/legacyheaders/types/simple.h"
  47 #include "gromacs/math/vec.h"
  48 #include "gromacs/legacyheaders/nrnb.h"
  49
  50 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  51 #include "kernelutil_x86_avx_128_fma_single.h"
  52 /* #endif */
  53
  54 /* ## List of variables set by the generating script:                                    */
  55 /* ##                                                                                    */
  56 /* ## Setttings that apply to the entire kernel:                                         */
  57 /* ## KERNEL_ELEC:           String, choice for electrostatic interactions               */
  58 /* ## KERNEL_VDW:            String, choice for van der Waals interactions               */
  59 /* ## KERNEL_NAME:           String, name of this kernel                                 */
  60 /* ## KERNEL_VF:             String telling if we calculate potential, force, or both    */
  61 /* ## GEOMETRY_I/GEOMETRY_J: String, name of each geometry, e.g. 'Water3' or '1Particle' */
  62 /* ##                                                                                    */
  63 /* ## Setttings that apply to particles in the outer (I) or inner (J) loops:             */
  64 /* ## PARTICLES_I[]/         Arrays with lists of i/j particles to use in kernel. It is  */
  65 /* ## PARTICLES_J[]:         just [0] for particle geometry, but can be longer for water */
  66 /* ## PARTICLES_ELEC_I[]/    Arrays with lists of i/j particle that have electrostatics  */
  67 /* ## PARTICLES_ELEC_J[]:    interactions that should be calculated in this kernel.      */
  68 /* ## PARTICLES_VDW_I[]/     Arrays with the list of i/j particle that have VdW          */
  69 /* ## PARTICLES_VDW_J[]:     interactions that should be calculated in this kernel.      */
  70 /* ##                                                                                    */
  71 /* ## Setttings for pairs of interactions (e.g. 2nd i particle against 1st j particle)   */
  72 /* ## PAIRS_IJ[]:            Array with (i,j) tuples of pairs for which interactions     */
  73 /* ##                        should be calculated in this kernel. Zero-charge particles  */
  74 /* ##                        do not have interactions with particles without vdw, and    */
  75 /* ##                        Vdw-only interactions are not evaluated in a no-vdw-kernel. */
  76 /* ## INTERACTION_FLAGS[][]: 2D matrix, dimension e.g. 3*3 for water-water interactions. */
  77 /* ##                        For each i-j pair, the element [I][J] is a list of strings  */
  78 /* ##                        defining properties/flags of this interaction. Examples     */
  79 /* ##                        include 'electrostatics'/'vdw' if that type of interaction  */
  80 /* ##                        should be evaluated, 'rsq'/'rinv'/'rinvsq' if those values  */
  81 /* ##                        are needed, and 'exactcutoff' or 'shift','switch' to        */
  82 /* ##                        decide if the force/potential should be modified. This way  */
  83 /* ##                        we only calculate values absolutely needed for each case.   */
  84
  85 /* ## Calculate the size and offset for (merged/interleaved) table data */
  86
  87 /*
  88  * Gromacs nonbonded kernel:   {KERNEL_NAME}
  89  * Electrostatics interaction: {KERNEL_ELEC}
  90  * VdW interaction:            {KERNEL_VDW}
  91  * Geometry:                   {GEOMETRY_I}-{GEOMETRY_J}
  92  * Calculate force/pot:        {KERNEL_VF}
  93  */
  94 void
  95 {KERNEL_NAME}
  96                     (t_nblist                    * gmx_restrict       nlist,
  97                      rvec                        * gmx_restrict          xx,
  98                      rvec                        * gmx_restrict          ff,
  99                      t_forcerec                  * gmx_restrict          fr,
 100                      t_mdatoms                   * gmx_restrict     mdatoms,
 101                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 102                      t_nrnb                      * gmx_restrict        nrnb)
 103 {
 104     /* ## Not all variables are used for all kernels, but any optimizing compiler fixes that, */
 105     /* ## so there is no point in going to extremes to exclude variables that are not needed. */
 106     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 107      * just 0 for non-waters.
 108      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 109      * jnr indices corresponding to data put in the four positions in the SIMD register.
 110      */
 111     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 112     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 113     int              jnrA,jnrB,jnrC,jnrD;
 114     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 115     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 116     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 117     real             rcutoff_scalar;
 118     real             *shiftvec,*fshift,*x,*f;
 119     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 120     real             scratch[4*DIM];
 121     __m128           fscal,rcutoff,rcutoff2,jidxall;
 122     /* #for I in PARTICLES_I */
 123     int              vdwioffset{I};
 124     __m128           ix{I},iy{I},iz{I},fix{I},fiy{I},fiz{I},iq{I},isai{I};
 125     /* #endfor */
 126     /* #for J in PARTICLES_J */
 127     int              vdwjidx{J}A,vdwjidx{J}B,vdwjidx{J}C,vdwjidx{J}D;
 128     __m128           jx{J},jy{J},jz{J},fjx{J},fjy{J},fjz{J},jq{J},isaj{J};
 129     /* #endfor */
 130     /* #for I,J in PAIRS_IJ */
 131     __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};
 132     /* #endfor */
 133     /* #if KERNEL_ELEC != 'None' */
 134     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 135     real             *charge;
 136     /* #endif */
 137     /* #if 'GeneralizedBorn' in KERNEL_ELEC */
 138     __m128i          gbitab;
 139     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
 140     __m128           minushalf = _mm_set1_ps(-0.5);
 141     real             *invsqrta,*dvda,*gbtab;
 142     /* #endif */
 143     /* #if KERNEL_VDW != 'None' */
 144     int              nvdwtype;
 145     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 146     int              *vdwtype;
 147     real             *vdwparam;
 148     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 149     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 150     /* #endif */
 151     /* #if 'Table' in KERNEL_ELEC or 'GeneralizedBorn' in KERNEL_ELEC or 'Table' in KERNEL_VDW */
 152     __m128i          vfitab;
 153     __m128i          ifour       = _mm_set1_epi32(4);
 154     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 155     real             *vftab;
 156     /* #endif */
 157     /* #if 'LJEwald' in KERNEL_VDW */
 158     /* #for I,J in PAIRS_IJ */
 159     __m128           c6grid_{I}{J};
 160     /* #endfor */
 161     real             *vdwgridparam;
 162     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 163     __m128           one_half = _mm_set1_ps(0.5);
 164     __m128           minus_one = _mm_set1_ps(-1.0);
 165     /* #endif */
 166     /* #if 'Ewald' in KERNEL_ELEC */
 167     __m128i          ewitab;
 168     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 169     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 170     real             *ewtab;
 171     /* #endif */
 172     /* #if 'PotentialSwitch' in [KERNEL_MOD_ELEC,KERNEL_MOD_VDW] */
 173     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 174     real             rswitch_scalar,d_scalar;
 175     /* #endif */
 176     __m128           dummy_mask,cutoff_mask;
 177     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 178     __m128           one     = _mm_set1_ps(1.0);
 179     __m128           two     = _mm_set1_ps(2.0);
 180     x                = xx[0];
 181     f                = ff[0];
 182
 183     nri              = nlist->nri;
 184     iinr             = nlist->iinr;
 185     jindex           = nlist->jindex;
 186     jjnr             = nlist->jjnr;
 187     shiftidx         = nlist->shift;
 188     gid              = nlist->gid;
 189     shiftvec         = fr->shift_vec[0];
 190     fshift           = fr->fshift[0];
 191     /* #if KERNEL_ELEC != 'None' */
 192     facel            = _mm_set1_ps(fr->epsfac);
 193     charge           = mdatoms->chargeA;
 194     /*     #if 'ReactionField' in KERNEL_ELEC */
 195     krf              = _mm_set1_ps(fr->ic->k_rf);
 196     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 197     crf              = _mm_set1_ps(fr->ic->c_rf);
 198     /*     #endif */
 199     /* #endif */
 200     /* #if KERNEL_VDW != 'None' */
 201     nvdwtype         = fr->ntype;
 202     vdwparam         = fr->nbfp;
 203     vdwtype          = mdatoms->typeA;
 204     /* #endif */
 205     /* #if 'LJEwald' in KERNEL_VDW */
 206     vdwgridparam     = fr->ljpme_c6grid;
 207     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 208     ewclj            = _mm_set1_ps(fr->ewaldcoeff_lj);
 209     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 210     /* #endif */
 211
 212     /* #if 'Table' in KERNEL_ELEC and 'Table' in KERNEL_VDW */
 213     vftab            = kernel_data->table_elec_vdw->data;
 214     vftabscale       = _mm_set1_ps(kernel_data->table_elec_vdw->scale);
 215     /* #elif 'Table' in KERNEL_ELEC */
 216     vftab            = kernel_data->table_elec->data;
 217     vftabscale       = _mm_set1_ps(kernel_data->table_elec->scale);
 218     /* #elif 'Table' in KERNEL_VDW */
 219     vftab            = kernel_data->table_vdw->data;
 220     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 221     /* #endif */
 222
 223     /* #if 'Ewald' in KERNEL_ELEC */
 224     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 225     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 226     beta2            = _mm_mul_ps(beta,beta);
 227     beta3            = _mm_mul_ps(beta,beta2);
 228     /*     #if KERNEL_VF=='Force' and KERNEL_MOD_ELEC!='PotentialSwitch' */
 229     ewtab            = fr->ic->tabq_coul_F;
 230     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 231     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 232     /*     #else */
 233     ewtab            = fr->ic->tabq_coul_FDV0;
 234     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 235     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 236      /*     #endif */
 237     /* #endif */
 238
 239     /* #if KERNEL_ELEC=='GeneralizedBorn' */
 240     invsqrta         = fr->invsqrta;
 241     dvda             = fr->dvda;
 242     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 243     gbtab            = fr->gbtab.data;
 244     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 245     /* #endif */
 246
 247     /* #if 'Water' in GEOMETRY_I */
 248     /* Setup water-specific parameters */
 249     inr              = nlist->iinr[0];
 250     /*     #for I in PARTICLES_ELEC_I */
 251     iq{I}              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+{I}]));
 252     /*     #endfor */
 253     /*     #for I in PARTICLES_VDW_I */
 254     vdwioffset{I}      = 2*nvdwtype*vdwtype[inr+{I}];
 255     /*     #endfor */
 256     /* #endif */
 257
 258     /* #if 'Water' in GEOMETRY_J */
 259     /*     #for J in PARTICLES_ELEC_J */
 260     jq{J}              = _mm_set1_ps(charge[inr+{J}]);
 261     /*     #endfor */
 262     /*     #for J in PARTICLES_VDW_J */
 263     vdwjidx{J}A        = 2*vdwtype[inr+{J}];
 264     /*     #endfor */
 265     /*     #for I,J in PAIRS_IJ */
 266     /*         #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
 267     qq{I}{J}             = _mm_mul_ps(iq{I},jq{J});
 268     /*         #endif */
 269     /*         #if 'vdw' in INTERACTION_FLAGS[I][J] */
 270     /*             #if 'LJEwald' in KERNEL_VDW */
 271     c6_{I}{J}            = _mm_set1_ps(vdwparam[vdwioffset{I}+vdwjidx{J}A]);
 272     c12_{I}{J}           = _mm_set1_ps(vdwparam[vdwioffset{I}+vdwjidx{J}A+1]);
 273     c6grid_{I}{J}        = _mm_set1_ps(vdwgridparam[vdwioffset{I}+vdwjidx{J}A]);
 274     /*             #else */
 275     c6_{I}{J}            = _mm_set1_ps(vdwparam[vdwioffset{I}+vdwjidx{J}A]);
 276     c12_{I}{J}           = _mm_set1_ps(vdwparam[vdwioffset{I}+vdwjidx{J}A+1]);
 277     /*             #endif */
 278     /*         #endif */
 279     /*     #endfor */
 280     /* #endif */
 281
 282     /* #if KERNEL_MOD_ELEC!='None' or KERNEL_MOD_VDW!='None' */
 283     /*     #if KERNEL_ELEC!='None' */
 284     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 285     rcutoff_scalar   = fr->rcoulomb;
 286     /*     #else */
 287     rcutoff_scalar   = fr->rvdw;
 288     /*     #endif */
 289     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 290     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 291     /* #endif */
 292
 293     /* #if KERNEL_MOD_VDW=='PotentialShift' */
 294     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 295     rvdw             = _mm_set1_ps(fr->rvdw);
 296     /* #endif */
 297
 298     /* #if 'PotentialSwitch' in [KERNEL_MOD_ELEC,KERNEL_MOD_VDW] */
 299     /*     #if KERNEL_MOD_ELEC=='PotentialSwitch'  */
 300     rswitch_scalar   = fr->rcoulomb_switch;
 301     rswitch          = _mm_set1_ps(rswitch_scalar);
 302     /*     #else */
 303     rswitch_scalar   = fr->rvdw_switch;
 304     rswitch          = _mm_set1_ps(rswitch_scalar);
 305     /*     #endif */
 306     /* Setup switch parameters */
 307     d_scalar         = rcutoff_scalar-rswitch_scalar;
 308     d                = _mm_set1_ps(d_scalar);
 309     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 310     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 311     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 312     /*     #if 'Force' in KERNEL_VF */
 313     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 314     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 315     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 316     /*     #endif */
 317     /* #endif */
 318
 319     /* Avoid stupid compiler warnings */
 320     jnrA = jnrB = jnrC = jnrD = 0;
 321     j_coord_offsetA = 0;
 322     j_coord_offsetB = 0;
 323     j_coord_offsetC = 0;
 324     j_coord_offsetD = 0;
 325
 326     /* ## Keep track of the floating point operations we issue for reporting! */
 327     /* #define OUTERFLOPS 0 */
 328     outeriter        = 0;
 329     inneriter        = 0;
 330
 331     for(iidx=0;iidx<4*DIM;iidx++)
 332     {
 333         scratch[iidx] = 0.0;
 334     }
 335
 336     /* Start outer loop over neighborlists */
 337     for(iidx=0; iidx<nri; iidx++)
 338     {
 339         /* Load shift vector for this list */
 340         i_shift_offset   = DIM*shiftidx[iidx];
 341
 342         /* Load limits for loop over neighbors */
 343         j_index_start    = jindex[iidx];
 344         j_index_end      = jindex[iidx+1];
 345
 346         /* Get outer coordinate index */
 347         inr              = iinr[iidx];
 348         i_coord_offset   = DIM*inr;
 349
 350         /* Load i particle coords and add shift vector */
 351         /* #if GEOMETRY_I == 'Particle' */
 352         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 353         /* #elif GEOMETRY_I == 'Water3' */
 354         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 355                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 356         /* #elif GEOMETRY_I == 'Water4' */
 357         /*     #if 0 in PARTICLES_I                 */
 358         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 359                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 360         /*     #else                                */
 361         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 362                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 363         /*     #endif                               */
 364         /* #endif                                   */
 365
 366         /* #if 'Force' in KERNEL_VF */
 367         /*     #for I in PARTICLES_I */
 368         fix{I}             = _mm_setzero_ps();
 369         fiy{I}             = _mm_setzero_ps();
 370         fiz{I}             = _mm_setzero_ps();
 371         /*     #endfor */
 372         /* #endif */
 373
 374         /* ## For water we already preloaded parameters at the start of the kernel */
 375         /* #if not 'Water' in GEOMETRY_I */
 376         /* Load parameters for i particles */
 377         /*     #for I in PARTICLES_ELEC_I */
 378         iq{I}              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+{I}));
 379         /*         #define OUTERFLOPS OUTERFLOPS+1 */
 380         /*         #if KERNEL_ELEC=='GeneralizedBorn' */
 381         isai{I}            = _mm_load1_ps(invsqrta+inr+{I});
 382         /*         #endif */
 383         /*     #endfor */
 384         /*     #for I in PARTICLES_VDW_I */
 385         vdwioffset{I}      = 2*nvdwtype*vdwtype[inr+{I}];
 386         /*     #endfor */
 387         /* #endif */
 388
 389         /* #if 'Potential' in KERNEL_VF */
 390         /* Reset potential sums */
 391         /*     #if KERNEL_ELEC != 'None' */
 392         velecsum         = _mm_setzero_ps();
 393         /*     #endif */
 394         /*     #if 'GeneralizedBorn' in KERNEL_ELEC */
 395         vgbsum           = _mm_setzero_ps();
 396         /*     #endif */
 397         /*     #if KERNEL_VDW != 'None' */
 398         vvdwsum          = _mm_setzero_ps();
 399         /*     #endif */
 400         /* #endif */
 401         /*     #if 'GeneralizedBorn' in KERNEL_ELEC and 'Force' in KERNEL_VF */
 402         dvdasum          = _mm_setzero_ps();
 403         /*     #endif */
 404
 405         /* #for ROUND in ['Loop','Epilogue'] */
 406
 407         /* #if ROUND =='Loop' */
 408         /* Start inner kernel loop */
 409         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 410         {
 411         /* ## First round is normal loop (next statement resets indentation) */
 412         /*     #if 0 */
 413         }
 414         /*     #endif */
 415         /* #else */
 416         if(jidx<j_index_end)
 417         {
 418         /* ## Second round is epilogue */
 419         /* #endif */
 420         /* #define INNERFLOPS 0 */
 421
 422             /* Get j neighbor index, and coordinate index */
 423             /* #if ROUND =='Loop' */
 424             jnrA             = jjnr[jidx];
 425             jnrB             = jjnr[jidx+1];
 426             jnrC             = jjnr[jidx+2];
 427             jnrD             = jjnr[jidx+3];
 428             /* #else */
 429             jnrlistA         = jjnr[jidx];
 430             jnrlistB         = jjnr[jidx+1];
 431             jnrlistC         = jjnr[jidx+2];
 432             jnrlistD         = jjnr[jidx+3];
 433             /* Sign of each element will be negative for non-real atoms.
 434              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 435              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 436              */
 437             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 438             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 439             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 440             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 441             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 442             /* #endif */
 443             j_coord_offsetA  = DIM*jnrA;
 444             j_coord_offsetB  = DIM*jnrB;
 445             j_coord_offsetC  = DIM*jnrC;
 446             j_coord_offsetD  = DIM*jnrD;
 447
 448             /* load j atom coordinates */
 449             /* #if GEOMETRY_J == 'Particle'             */
 450             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 451                                               x+j_coord_offsetC,x+j_coord_offsetD,
 452                                               &jx0,&jy0,&jz0);
 453             /* #elif GEOMETRY_J == 'Water3'             */
 454             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 455                                               x+j_coord_offsetC,x+j_coord_offsetD,
 456                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 457             /* #elif GEOMETRY_J == 'Water4'             */
 458             /*     #if 0 in PARTICLES_J                 */
 459             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 460                                               x+j_coord_offsetC,x+j_coord_offsetD,
 461                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 462                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 463             /*     #else                                */
 464             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 465                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 466                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 467             /*     #endif                               */
 468             /* #endif                                   */
 469
 470             /* Calculate displacement vector */
 471             /* #for I,J in PAIRS_IJ */
 472             dx{I}{J}             = _mm_sub_ps(ix{I},jx{J});
 473             dy{I}{J}             = _mm_sub_ps(iy{I},jy{J});
 474             dz{I}{J}             = _mm_sub_ps(iz{I},jz{J});
 475             /*     #define INNERFLOPS INNERFLOPS+3 */
 476             /* #endfor */
 477
 478             /* Calculate squared distance and things based on it */
 479             /* #for I,J in PAIRS_IJ */
 480             rsq{I}{J}            = gmx_mm_calc_rsq_ps(dx{I}{J},dy{I}{J},dz{I}{J});
 481             /*     #define INNERFLOPS INNERFLOPS+5 */
 482             /* #endfor */
 483
 484             /* #for I,J in PAIRS_IJ */
 485             /*     #if 'rinv' in INTERACTION_FLAGS[I][J] */
 486             rinv{I}{J}           = gmx_mm_invsqrt_ps(rsq{I}{J});
 487             /*         #define INNERFLOPS INNERFLOPS+5 */
 488             /*     #endif */
 489             /* #endfor */
 490
 491             /* #for I,J in PAIRS_IJ */
 492             /*     #if 'rinvsq' in INTERACTION_FLAGS[I][J] */
 493             /*         # if 'rinv' not in INTERACTION_FLAGS[I][J] */
 494             rinvsq{I}{J}         = gmx_mm_inv_ps(rsq{I}{J});
 495             /*             #define INNERFLOPS INNERFLOPS+4 */
 496             /*         #else */
 497             rinvsq{I}{J}         = _mm_mul_ps(rinv{I}{J},rinv{I}{J});
 498             /*             #define INNERFLOPS INNERFLOPS+1 */
 499             /*         #endif */
 500             /*     #endif */
 501             /* #endfor */
 502
 503             /* #if not 'Water' in GEOMETRY_J */
 504             /* Load parameters for j particles */
 505             /*     #for J in PARTICLES_ELEC_J */
 506             jq{J}              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+{J},charge+jnrB+{J},
 507                                                               charge+jnrC+{J},charge+jnrD+{J});
 508             /*         #if KERNEL_ELEC=='GeneralizedBorn' */
 509             isaj{J}            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+{J},invsqrta+jnrB+{J},
 510                                                               invsqrta+jnrC+{J},invsqrta+jnrD+{J});
 511             /*         #endif */
 512             /*     #endfor */
 513             /*     #for J in PARTICLES_VDW_J */
 514             vdwjidx{J}A        = 2*vdwtype[jnrA+{J}];
 515             vdwjidx{J}B        = 2*vdwtype[jnrB+{J}];
 516             vdwjidx{J}C        = 2*vdwtype[jnrC+{J}];
 517             vdwjidx{J}D        = 2*vdwtype[jnrD+{J}];
 518             /*     #endfor */
 519             /* #endif */
 520
 521             /* #if 'Force' in KERNEL_VF and not 'Particle' in GEOMETRY_I */
 522             /*     #for J in PARTICLES_J */
 523             fjx{J}             = _mm_setzero_ps();
 524             fjy{J}             = _mm_setzero_ps();
 525             fjz{J}             = _mm_setzero_ps();
 526             /*     #endfor */
 527             /* #endif */
 528
 529             /* #for I,J in PAIRS_IJ */
 530
 531             /**************************
 532              * CALCULATE INTERACTIONS *
 533              **************************/
 534
 535             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 536             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 537             /*         ## We always calculate rinv/rinvsq above to enable pipelineing in compilers (performance tested on x86) */
 538             if (gmx_mm_any_lt(rsq{I}{J},rcutoff2))
 539             {
 540                 /*     #if 0    ## this and the next two lines is a hack to maintain auto-indentation in template file */
 541             }
 542             /*         #endif */
 543             /*         #define INNERFLOPS INNERFLOPS+1 */
 544             /*     #endif */
 545
 546             /*     #if 'r' in INTERACTION_FLAGS[I][J] */
 547             r{I}{J}              = _mm_mul_ps(rsq{I}{J},rinv{I}{J});
 548             /*         #if ROUND == 'Epilogue' */
 549             r{I}{J}              = _mm_andnot_ps(dummy_mask,r{I}{J});
 550             /*             #define INNERFLOPS INNERFLOPS+1 */
 551             /*         #endif */
 552             /*         #define INNERFLOPS INNERFLOPS+1 */
 553             /*     #endif */
 554
 555             /*     ## For water geometries we already loaded parameters at the start of the kernel */
 556             /*     #if not 'Water' in GEOMETRY_J */
 557             /* Compute parameters for interactions between i and j atoms */
 558             /*         #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
 559             qq{I}{J}             = _mm_mul_ps(iq{I},jq{J});
 560             /*             #define INNERFLOPS INNERFLOPS+1 */
 561             /*         #endif */
 562             /*         #if 'vdw' in INTERACTION_FLAGS[I][J] */
 563             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset{I}+vdwjidx{J}A,
 564                                          vdwparam+vdwioffset{I}+vdwjidx{J}B,
 565                                          vdwparam+vdwioffset{I}+vdwjidx{J}C,
 566                                          vdwparam+vdwioffset{I}+vdwjidx{J}D,
 567                                          &c6_{I}{J},&c12_{I}{J});
 568
 569             /*           #if 'LJEwald' in KERNEL_VDW */
 570             c6grid_{I}{J}       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset{I}+vdwjidx{J}A,
 571                                                                vdwgridparam+vdwioffset{I}+vdwjidx{J}B,
 572                                                                vdwgridparam+vdwioffset{I}+vdwjidx{J}C,
 573                                                                vdwgridparam+vdwioffset{I}+vdwjidx{J}D);
 574             /*           #endif */
 575
 576             /*         #endif */
 577             /*     #endif */
 578
 579             /*     #if 'table' in INTERACTION_FLAGS[I][J] */
 580             /* Calculate table index by multiplying r with table scale and truncate to integer */
 581             rt               = _mm_mul_ps(r{I}{J},vftabscale);
 582             vfitab           = _mm_cvttps_epi32(rt);
 583 #ifdef __XOP__
 584             vfeps            = _mm_frcz_ps(rt);
 585 #else
 586             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 587 #endif
 588             twovfeps         = _mm_add_ps(vfeps,vfeps);
 589             /*         #define INNERFLOPS INNERFLOPS+4                          */
 590             /*         #if 'Table' in KERNEL_ELEC and 'Table' in KERNEL_VDW     */
 591             /*             ## 3 tables, 4 bytes per point: multiply index by 12 */
 592             vfitab           = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
 593             /*         #elif 'Table' in KERNEL_ELEC                             */
 594             /*             ## 1 table, 4 bytes per point: multiply index by 4   */
 595             vfitab           = _mm_slli_epi32(vfitab,2);
 596             /*         #elif 'Table' in KERNEL_VDW                              */
 597             /*             ## 2 tables, 4 bytes per point: multiply index by 8  */
 598             vfitab           = _mm_slli_epi32(vfitab,3);
 599             /*         #endif                                                   */
 600             /*     #endif */
 601
 602             /*     ## ELECTROSTATIC INTERACTIONS */
 603             /*     #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
 604
 605             /*         #if KERNEL_ELEC=='Coulomb' */
 606
 607             /* COULOMB ELECTROSTATICS */
 608             velec            = _mm_mul_ps(qq{I}{J},rinv{I}{J});
 609             /*             #define INNERFLOPS INNERFLOPS+1 */
 610             /*             #if 'Force' in KERNEL_VF */
 611             felec            = _mm_mul_ps(velec,rinvsq{I}{J});
 612             /*                 #define INNERFLOPS INNERFLOPS+2 */
 613             /*             #endif */
 614
 615             /*         #elif KERNEL_ELEC=='ReactionField' */
 616
 617             /* REACTION-FIELD ELECTROSTATICS */
 618             /*             #if 'Potential' in KERNEL_VF */
 619             velec            = _mm_mul_ps(qq{I}{J},_mm_sub_ps(_mm_macc_ps(krf,rsq{I}{J},rinv{I}{J}),crf));
 620             /*                 #define INNERFLOPS INNERFLOPS+4 */
 621             /*             #endif */
 622             /*             #if 'Force' in KERNEL_VF */
 623             felec            = _mm_mul_ps(qq{I}{J},_mm_msub_ps(rinv{I}{J},rinvsq{I}{J},krf2));
 624             /*                 #define INNERFLOPS INNERFLOPS+3 */
 625             /*             #endif */
 626
 627             /*         #elif KERNEL_ELEC=='GeneralizedBorn' */
 628
 629             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 630             isaprod          = _mm_mul_ps(isai{I},isaj{J});
 631             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq{I}{J},_mm_mul_ps(isaprod,gbinvepsdiff)));
 632             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 633             /*             #define INNERFLOPS INNERFLOPS+5 */
 634
 635             /* Calculate generalized born table index - this is a separate table from the normal one,
 636              * but we use the same procedure by multiplying r with scale and truncating to integer.
 637              */
 638             rt               = _mm_mul_ps(r{I}{J},gbscale);
 639             gbitab           = _mm_cvttps_epi32(rt);
 640 #ifdef __XOP__
 641             gbeps            = _mm_frcz_ps(rt);
 642 #else
 643             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 644 #endif
 645             gbitab           = _mm_slli_epi32(gbitab,2);
 646
 647             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 648             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 649             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 650             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 651             _MM_TRANSPOSE4_PS(Y,F,G,H);
 652             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 653             VV               = _mm_macc_ps(gbeps,Fp,Y);
 654             vgb              = _mm_mul_ps(gbqqfactor,VV);
 655             /*             #define INNERFLOPS INNERFLOPS+10 */
 656
 657             /*             #if 'Force' in KERNEL_VF */
 658             twogbeps         = _mm_add_ps(gbeps,gbeps);
 659             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 660             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 661             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r{I}{J},vgb));
 662             /*                 #if ROUND == 'Epilogue' */
 663             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 664             /*                 #endif */
 665             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 666             /*                 #if ROUND == 'Loop' */
 667             fjptrA           = dvda+jnrA;
 668             fjptrB           = dvda+jnrB;
 669             fjptrC           = dvda+jnrC;
 670             fjptrD           = dvda+jnrD;
 671             /*                 #else */
 672             /* 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. */
 673             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 674             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 675             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 676             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 677             /*                 #endif */
 678             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj{J},isaj{J})));
 679             /*                 #define INNERFLOPS INNERFLOPS+13 */
 680             /*             #endif */
 681             velec            = _mm_mul_ps(qq{I}{J},rinv{I}{J});
 682             /*                 #define INNERFLOPS INNERFLOPS+1 */
 683             /*             #if 'Force' in KERNEL_VF */
 684             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv{I}{J},fgb),rinv{I}{J});
 685             /*                 #define INNERFLOPS INNERFLOPS+3 */
 686             /*             #endif */
 687
 688             /*         #elif KERNEL_ELEC=='Ewald' */
 689             /* EWALD ELECTROSTATICS */
 690
 691             /* Analytical PME correction */
 692             zeta2            = _mm_mul_ps(beta2,rsq{I}{J});
 693             /*             #if 'Force' in KERNEL_VF */
 694             rinv3            = _mm_mul_ps(rinvsq{I}{J},rinv{I}{J});
 695             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 696             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 697             felec            = _mm_mul_ps(qq{I}{J},felec);
 698             /*             #endif */
 699             /*             #if 'Potential' in KERNEL_VF or KERNEL_MOD_ELEC=='PotentialSwitch' */
 700             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 701             /*                 #if KERNEL_MOD_ELEC=='PotentialShift' */
 702             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv{I}{J},sh_ewald));
 703             /*                 #else */
 704             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv{I}{J});
 705             /*                 #endif */
 706             velec            = _mm_mul_ps(qq{I}{J},velec);
 707             /*             #endif */
 708
 709             /*         #elif KERNEL_ELEC=='CubicSplineTable' */
 710
 711             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 712             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 713             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 714             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 715             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 716             _MM_TRANSPOSE4_PS(Y,F,G,H);
 717             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 718             /*             #define INNERFLOPS INNERFLOPS+4 */
 719             /*             #if 'Potential' in KERNEL_VF */
 720             VV               = _mm_macc_ps(vfeps,Fp,Y);
 721             velec            = _mm_mul_ps(qq{I}{J},VV);
 722             /*                 #define INNERFLOPS INNERFLOPS+3 */
 723             /*             #endif */
 724             /*             #if 'Force' in KERNEL_VF */
 725             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 726             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq{I}{J},FF),_mm_mul_ps(vftabscale,rinv{I}{J})));
 727             /*                 #define INNERFLOPS INNERFLOPS+7 */
 728             /*             #endif */
 729             /*         #endif */
 730             /*         ## End of check for electrostatics interaction forms */
 731             /*     #endif */
 732             /*     ## END OF ELECTROSTATIC INTERACTION CHECK FOR PAIR I-J */
 733
 734             /*     #if 'vdw' in INTERACTION_FLAGS[I][J] */
 735
 736             /*         #if KERNEL_VDW=='LennardJones' */
 737
 738             /* LENNARD-JONES DISPERSION/REPULSION */
 739
 740             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq{I}{J},rinvsq{I}{J}),rinvsq{I}{J});
 741             /*             #define INNERFLOPS INNERFLOPS+2 */
 742             /*             #if 'Potential' in KERNEL_VF or KERNEL_MOD_VDW=='PotentialSwitch' */
 743             vvdw6            = _mm_mul_ps(c6_{I}{J},rinvsix);
 744             vvdw12           = _mm_mul_ps(c12_{I}{J},_mm_mul_ps(rinvsix,rinvsix));
 745             /*                 #define INNERFLOPS INNERFLOPS+3 */
 746             /*                 #if KERNEL_MOD_VDW=='PotentialShift' */
 747             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_{I}{J},_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 748                                           _mm_mul_ps( _mm_nmacc_ps(c6_{I}{J},sh_vdw_invrcut6,vvdw6),one_sixth));
 749             /*                     #define INNERFLOPS INNERFLOPS+8 */
 750             /*                 #else */
 751             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 752             /*                     #define INNERFLOPS INNERFLOPS+3 */
 753             /*                 #endif */
 754             /*                 ## Check for force inside potential check, i.e. this means we already did the potential part */
 755             /*                 #if 'Force' in KERNEL_VF */
 756             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq{I}{J});
 757             /*                     #define INNERFLOPS INNERFLOPS+2 */
 758             /*                 #endif */
 759             /*             #elif KERNEL_VF=='Force' */
 760             /*                 ## Force-only LennardJones makes it possible to save 1 flop (they do add up...) */
 761             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_{I}{J},rinvsix,c6_{I}{J}),_mm_mul_ps(rinvsix,rinvsq{I}{J}));
 762             /*                 #define INNERFLOPS INNERFLOPS+4 */
 763             /*             #endif */
 764
 765             /*         #elif KERNEL_VDW=='CubicSplineTable' */
 766
 767             /* CUBIC SPLINE TABLE DISPERSION */
 768             /*             #if 'Table' in KERNEL_ELEC */
 769             vfitab           = _mm_add_epi32(vfitab,ifour);
 770             /*             #endif                     */
 771             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 772             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 773             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 774             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 775             _MM_TRANSPOSE4_PS(Y,F,G,H);
 776             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 777             /*             #define INNERFLOPS INNERFLOPS+4 */
 778             /*             #if 'Potential' in KERNEL_VF */
 779             VV               = _mm_macc_ps(vfeps,Fp,Y);
 780             vvdw6            = _mm_mul_ps(c6_{I}{J},VV);
 781             /*                 #define INNERFLOPS INNERFLOPS+3 */
 782             /*             #endif */
 783             /*             #if 'Force' in KERNEL_VF */
 784             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 785             fvdw6            = _mm_mul_ps(c6_{I}{J},FF);
 786             /*                 #define INNERFLOPS INNERFLOPS+4 */
 787             /*             #endif */
 788
 789             /* CUBIC SPLINE TABLE REPULSION */
 790             vfitab           = _mm_add_epi32(vfitab,ifour);
 791             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 792             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 793             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 794             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 795             _MM_TRANSPOSE4_PS(Y,F,G,H);
 796             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 797             /*             #define INNERFLOPS INNERFLOPS+4 */
 798             /*             #if 'Potential' in KERNEL_VF */
 799             VV               = _mm_macc_ps(vfeps,Fp,Y);
 800             vvdw12           = _mm_mul_ps(c12_{I}{J},VV);
 801             /*                 #define INNERFLOPS INNERFLOPS+3 */
 802             /*             #endif */
 803             /*             #if 'Force' in KERNEL_VF */
 804             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 805             fvdw12           = _mm_mul_ps(c12_{I}{J},FF);
 806             /*                 #define INNERFLOPS INNERFLOPS+5 */
 807             /*             #endif */
 808             /*             #if 'Potential' in KERNEL_VF */
 809             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 810             /*                 #define INNERFLOPS INNERFLOPS+1 */
 811             /*             #endif */
 812             /*             #if 'Force' in KERNEL_VF */
 813             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv{I}{J})));
 814             /*                 #define INNERFLOPS INNERFLOPS+4 */
 815             /*             #endif */
 816
 817             /*         #elif KERNEL_VDW=='LJEwald' */
 818
 819             /* Analytical LJ-PME */
 820             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq{I}{J},rinvsq{I}{J}),rinvsq{I}{J});
 821             ewcljrsq         = _mm_mul_ps(ewclj2,rsq{I}{J});
 822             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 823             exponent         = gmx_simd_exp_r(ewcljrsq);
 824             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 825             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 826             /*                 #define INNERFLOPS INNERFLOPS+10 */
 827             /*             #if 'Potential' in KERNEL_VF or KERNEL_MOD_VDW=='PotentialSwitch' */
 828             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 829             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_{I}{J},_mm_sub_ps(one,poly),c6_{I}{J}),rinvsix);
 830             vvdw12           = _mm_mul_ps(c12_{I}{J},_mm_mul_ps(rinvsix,rinvsix));
 831             /*                     #define INNERFLOPS INNERFLOPS+5 */
 832             /*                 #if KERNEL_MOD_VDW=='PotentialShift' */
 833             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_{I}{J},_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 834                                           _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));
 835             /*                     #define INNERFLOPS INNERFLOPS+7 */
 836             /*                 #else */
 837             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 838             /*                     #define INNERFLOPS INNERFLOPS+2 */
 839             /*                 #endif */
 840             /*                  ## Check for force inside potential check, i.e. this means we already did the potential part */
 841             /*                  #if 'Force' in KERNEL_VF */
 842             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 843             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});
 844             /*                 #define INNERFLOPS INNERFLOPS+5 */
 845             /*                  #endif */
 846             /*              #elif KERNEL_VF=='Force' */
 847             /* f6A = 6 * C6grid * (1 - poly) */
 848             f6A              = _mm_mul_ps(c6grid_{I}{J},_mm_sub_ps(one,poly));
 849             /* f6B = C6grid * exponent * beta^6 */
 850             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_{I}{J},one_sixth),_mm_mul_ps(exponent,ewclj6));
 851             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 852             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});
 853             /*                 #define INNERFLOPS INNERFLOPS+10 */
 854             /*              #endif */
 855             /*         #endif */
 856             /*         ## End of check for vdw interaction forms */
 857             /*     #endif */
 858             /*     ## END OF VDW INTERACTION CHECK FOR PAIR I-J */
 859
 860             /*     #if 'switch' in INTERACTION_FLAGS[I][J] */
 861             d                = _mm_sub_ps(r{I}{J},rswitch);
 862             d                = _mm_max_ps(d,_mm_setzero_ps());
 863             d2               = _mm_mul_ps(d,d);
 864             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 865             /*         #define INNERFLOPS INNERFLOPS+10 */
 866
 867             /*         #if 'Force' in KERNEL_VF */
 868             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 869             /*             #define INNERFLOPS INNERFLOPS+5 */
 870             /*         #endif */
 871
 872             /* Evaluate switch function */
 873             /*         #if 'Force' in KERNEL_VF */
 874             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 875             /*             #if 'electrostatics' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_ELEC=='PotentialSwitch' */
 876             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv{I}{J},_mm_mul_ps(velec,dsw)) );
 877             /*                 #define INNERFLOPS INNERFLOPS+4 */
 878             /*             #endif */
 879             /*             #if 'vdw' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_VDW=='PotentialSwitch' */
 880             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv{I}{J},_mm_mul_ps(vvdw,dsw)) );
 881             /*                 #define INNERFLOPS INNERFLOPS+4 */
 882             /*             #endif */
 883             /*         #endif */
 884             /*         #if 'Potential' in KERNEL_VF */
 885             /*             #if 'electrostatics' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_ELEC=='PotentialSwitch' */
 886             velec            = _mm_mul_ps(velec,sw);
 887             /*                 #define INNERFLOPS INNERFLOPS+1 */
 888             /*             #endif */
 889             /*             #if 'vdw' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_VDW=='PotentialSwitch' */
 890             vvdw             = _mm_mul_ps(vvdw,sw);
 891             /*                 #define INNERFLOPS INNERFLOPS+1 */
 892             /*             #endif */
 893             /*         #endif */
 894             /*     #endif */
 895             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 896             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 897             cutoff_mask      = _mm_cmplt_ps(rsq{I}{J},rcutoff2);
 898             /*         #define INNERFLOPS INNERFLOPS+1 */
 899             /*     #endif */
 900
 901             /*     #if 'Potential' in KERNEL_VF */
 902             /* Update potential sum for this i atom from the interaction with this j atom. */
 903             /*         #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
 904             /*             #if 'exactcutoff' in INTERACTION_FLAGS[I][J] */
 905             velec            = _mm_and_ps(velec,cutoff_mask);
 906             /*                 #define INNERFLOPS INNERFLOPS+1 */
 907             /*             #endif                                       */
 908             /*             #if ROUND == 'Epilogue' */
 909             velec            = _mm_andnot_ps(dummy_mask,velec);
 910             /*             #endif */
 911             velecsum         = _mm_add_ps(velecsum,velec);
 912             /*             #define INNERFLOPS INNERFLOPS+1 */
 913             /*             #if KERNEL_ELEC=='GeneralizedBorn' */
 914             /*             #if 'exactcutoff' in INTERACTION_FLAGS[I][J] */
 915             vgb              = _mm_and_ps(vgb,cutoff_mask);
 916             /*                 #define INNERFLOPS INNERFLOPS+1 */
 917             /*             #endif                                       */
 918             /*             #if ROUND == 'Epilogue' */
 919             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 920             /*             #endif */
 921             vgbsum           = _mm_add_ps(vgbsum,vgb);
 922             /*                 #define INNERFLOPS INNERFLOPS+1 */
 923             /*             #endif */
 924             /*         #endif */
 925             /*         #if 'vdw' in INTERACTION_FLAGS[I][J] */
 926             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 927             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 928             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 929             /*                 #define INNERFLOPS INNERFLOPS+1 */
 930             /*             #endif                                       */
 931             /*             #if ROUND == 'Epilogue' */
 932             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 933             /*             #endif */
 934             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 935             /*             #define INNERFLOPS INNERFLOPS+1 */
 936             /*         #endif */
 937             /*     #endif */
 938
 939             /*     #if 'Force' in KERNEL_VF */
 940
 941             /*         #if 'electrostatics' in INTERACTION_FLAGS[I][J] and 'vdw' in INTERACTION_FLAGS[I][J] */
 942             fscal            = _mm_add_ps(felec,fvdw);
 943             /*             #define INNERFLOPS INNERFLOPS+1 */
 944             /*         #elif 'electrostatics' in INTERACTION_FLAGS[I][J] */
 945             fscal            = felec;
 946             /*         #elif 'vdw' in INTERACTION_FLAGS[I][J] */
 947             fscal            = fvdw;
 948             /*        #endif */
 949
 950             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 951             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 952             fscal            = _mm_and_ps(fscal,cutoff_mask);
 953             /*                 #define INNERFLOPS INNERFLOPS+1 */
 954             /*             #endif                                       */
 955
 956             /*             #if ROUND == 'Epilogue' */
 957             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 958             /*             #endif */
 959
 960             /* ## Construction of vectorial force built into FMA instructions now */
 961             /* #define INNERFLOPS INNERFLOPS+3      */
 962
 963              /* Update vectorial force */
 964             fix{I}             = _mm_macc_ps(dx{I}{J},fscal,fix{I});
 965             fiy{I}             = _mm_macc_ps(dy{I}{J},fscal,fiy{I});
 966             fiz{I}             = _mm_macc_ps(dz{I}{J},fscal,fiz{I});
 967             /*             #define INNERFLOPS INNERFLOPS+6 */
 968
 969             /* #if GEOMETRY_I == 'Particle'             */
 970             /*     #if ROUND == 'Loop' */
 971             fjptrA             = f+j_coord_offsetA;
 972             fjptrB             = f+j_coord_offsetB;
 973             fjptrC             = f+j_coord_offsetC;
 974             fjptrD             = f+j_coord_offsetD;
 975             /*     #else */
 976             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 977             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 978             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 979             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 980             /*     #endif */
 981             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 982                                                    _mm_mul_ps(dx{I}{J},fscal),
 983                                                    _mm_mul_ps(dy{I}{J},fscal),
 984                                                    _mm_mul_ps(dz{I}{J},fscal));
 985             /*     #define INNERFLOPS INNERFLOPS+3      */
 986             /* #else                                    */
 987             fjx{J}             = _mm_macc_ps(dx{I}{J},fscal,fjx{J});
 988             fjy{J}             = _mm_macc_ps(dy{I}{J},fscal,fjy{J});
 989             fjz{J}             = _mm_macc_ps(dz{I}{J},fscal,fjz{J});
 990             /*     #define INNERFLOPS INNERFLOPS+3      */
 991             /* #endif                                   */
 992
 993             /*     #endif */
 994
 995             /*     ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
 996             /*     #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
 997             /*         #if 0    ## This and next two lines is a hack to maintain indentation in template file */
 998             {
 999                 /*     #endif */
1000             }
1001             /*     #endif */
1002             /*    ## End of check for the interaction being outside the cutoff */
1003
1004             /* #endfor */
1005             /* ## End of loop over i-j interaction pairs */
1006
1007             /* #if GEOMETRY_I != 'Particle' */
1008             /*     #if ROUND == 'Loop' */
1009             fjptrA             = f+j_coord_offsetA;
1010             fjptrB             = f+j_coord_offsetB;
1011             fjptrC             = f+j_coord_offsetC;
1012             fjptrD             = f+j_coord_offsetD;
1013             /*     #else */
1014             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1015             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1016             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1017             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1018             /*     #endif */
1019             /* #endif */
1020
1021             /* #if 'Water' in GEOMETRY_I and GEOMETRY_J == 'Particle' */
1022             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1023             /* #elif GEOMETRY_J == 'Water3'               */
1024             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1025                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1026             /*     #define INNERFLOPS INNERFLOPS+9      */
1027             /* #elif GEOMETRY_J == 'Water4'             */
1028             /*     #if 0 in PARTICLES_J                 */
1029             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1030                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1031                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1032             /*     #define INNERFLOPS INNERFLOPS+12     */
1033             /*     #else                                */
1034             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1035                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1036             /*     #define INNERFLOPS INNERFLOPS+9      */
1037             /*     #endif                               */
1038             /* #endif                                   */
1039
1040             /* Inner loop uses {INNERFLOPS} flops */
1041         }
1042
1043         /* #endfor */
1044
1045         /* End of innermost loop */
1046
1047         /* #if 'Force' in KERNEL_VF */
1048         /*     #if GEOMETRY_I == 'Particle'            */
1049         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
1050                                               f+i_coord_offset,fshift+i_shift_offset);
1051         /*         #define OUTERFLOPS OUTERFLOPS+6     */
1052         /*     #elif GEOMETRY_I == 'Water3'            */
1053         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1054                                               f+i_coord_offset,fshift+i_shift_offset);
1055         /*         #define OUTERFLOPS OUTERFLOPS+18    */
1056         /*     #elif GEOMETRY_I == 'Water4'            */
1057         /*         #if 0 in PARTICLES_I                */
1058         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1059                                               f+i_coord_offset,fshift+i_shift_offset);
1060         /*             #define OUTERFLOPS OUTERFLOPS+24    */
1061         /*         #else                               */
1062         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1063                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1064         /*             #define OUTERFLOPS OUTERFLOPS+18    */
1065         /*         #endif                              */
1066         /*     #endif                                  */
1067         /* #endif                                      */
1068
1069         /* #if 'Potential' in KERNEL_VF */
1070         ggid                        = gid[iidx];
1071         /* Update potential energies */
1072         /*     #if KERNEL_ELEC != 'None' */
1073         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1074         /*         #define OUTERFLOPS OUTERFLOPS+1 */
1075         /*     #endif */
1076         /*     #if 'GeneralizedBorn' in KERNEL_ELEC */
1077         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
1078         /*         #define OUTERFLOPS OUTERFLOPS+1 */
1079         /*     #endif */
1080         /*     #if KERNEL_VDW != 'None' */
1081         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1082         /*         #define OUTERFLOPS OUTERFLOPS+1 */
1083         /*     #endif */
1084         /* #endif */
1085         /*     #if 'GeneralizedBorn' in KERNEL_ELEC and 'Force' in KERNEL_VF */
1086         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai{I},isai{I}));
1087         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
1088         /*     #endif */
1089
1090         /* Increment number of inner iterations */
1091         inneriter                  += j_index_end - j_index_start;
1092
1093         /* Outer loop uses {OUTERFLOPS} flops */
1094     }
1095
1096     /* Increment number of outer iterations */
1097     outeriter        += nri;
1098
1099     /* Update outer/inner flops */
1100     /* ## NB: This is not important, it just affects the flopcount. However, since our preprocessor is */
1101     /* ## primitive and replaces aggressively even in strings inside these directives, we need to      */
1102     /* ## assemble the main part of the name (containing KERNEL/ELEC/VDW) directly in the source.      */
1103     /* #if GEOMETRY_I == 'Water3'            */
1104     /*     #define ISUFFIX '_W3'             */
1105     /* #elif GEOMETRY_I == 'Water4'          */
1106     /*     #define ISUFFIX '_W4'             */
1107     /* #else                                 */
1108     /*     #define ISUFFIX ''                */
1109     /* #endif                                */
1110     /* #if GEOMETRY_J == 'Water3'            */
1111     /*     #define JSUFFIX 'W3'              */
1112     /* #elif GEOMETRY_J == 'Water4'          */
1113     /*     #define JSUFFIX 'W4'              */
1114     /* #else                                 */
1115     /*     #define JSUFFIX ''                */
1116     /* #endif                                */
1117     /* #if 'PotentialAndForce' in KERNEL_VF  */
1118     /*     #define VFSUFFIX  '_VF'           */
1119     /* #elif 'Potential' in KERNEL_VF        */
1120     /*     #define VFSUFFIX '_V'             */
1121     /* #else                                 */
1122     /*     #define VFSUFFIX '_F'             */
1123     /* #endif                                */
1124
1125     /* #if KERNEL_ELEC != 'None' and KERNEL_VDW != 'None' */
1126     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW{ISUFFIX}{JSUFFIX}{VFSUFFIX},outeriter*{OUTERFLOPS} + inneriter*{INNERFLOPS});
1127     /* #elif KERNEL_ELEC != 'None' */
1128     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC{ISUFFIX}{JSUFFIX}{VFSUFFIX},outeriter*{OUTERFLOPS} + inneriter*{INNERFLOPS});
1129     /* #else */
1130     inc_nrnb(nrnb,eNR_NBKERNEL_VDW{ISUFFIX}{JSUFFIX}{VFSUFFIX},outeriter*{OUTERFLOPS} + inneriter*{INNERFLOPS});
1131     /* #endif  */
1132 }