src/gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2016,2017,2018,2019, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35
  36
  37 {
  38     using namespace gmx;
  39
  40     /* Unpack pointers for output */
  41     real* f      = out->f.data();
  42     real* fshift = out->fshift.data();
  43 #ifdef CALC_ENERGIES
  44 #    ifdef ENERGY_GROUPS
  45     real* Vvdw = out->VSvdw.data();
  46     real* Vc   = out->VSc.data();
  47 #    else
  48     real* Vvdw = out->Vvdw.data();
  49     real* Vc   = out->Vc.data();
  50 #    endif
  51 #endif
  52
  53     const nbnxn_cj_t* l_cj;
  54     int               ci, ci_sh;
  55     int               ish, ish3;
  56     gmx_bool          do_LJ, half_LJ, do_coul;
  57     int               cjind0, cjind1, cjind;
  58
  59 #ifdef ENERGY_GROUPS
  60     int   Vstride_i;
  61     int   egps_ishift, egps_imask;
  62     int   egps_jshift, egps_jmask, egps_jstride;
  63     int   egps_i;
  64     real* vvdwtp[UNROLLI];
  65     real* vctp[UNROLLI];
  66 #endif
  67
  68     SimdReal shX_S;
  69     SimdReal shY_S;
  70     SimdReal shZ_S;
  71     SimdReal ix_S0, iy_S0, iz_S0;
  72     SimdReal ix_S2, iy_S2, iz_S2;
  73     SimdReal fix_S0, fiy_S0, fiz_S0;
  74     SimdReal fix_S2, fiy_S2, fiz_S2;
  75
  76     SimdReal diagonal_jmi_S;
  77 #if UNROLLI == UNROLLJ
  78     SimdBool diagonal_mask_S0, diagonal_mask_S2;
  79 #else
  80     SimdBool                         diagonal_mask0_S0, diagonal_mask0_S2;
  81     SimdBool                         diagonal_mask1_S0, diagonal_mask1_S2;
  82 #endif
  83
  84     SimdBitMask filter_S0, filter_S2;
  85
  86     SimdReal zero_S(0.0);
  87
  88     SimdReal one_S(1.0);
  89     SimdReal iq_S0 = setZero();
  90     SimdReal iq_S2 = setZero();
  91
  92 #ifdef CALC_COUL_RF
  93     SimdReal mrc_3_S;
  94 #    ifdef CALC_ENERGIES
  95     SimdReal hrc_3_S, moh_rc_S;
  96 #    endif
  97 #endif
  98
  99 #ifdef CALC_COUL_TAB
 100     /* Coulomb table variables */
 101     SimdReal    invtsp_S;
 102     const real* tab_coul_F;
 103 #    if defined CALC_ENERGIES && !defined TAB_FDV0
 104     const real*                           tab_coul_V;
 105 #    endif
 106
 107 #    ifdef CALC_ENERGIES
 108     SimdReal mhalfsp_S;
 109 #    endif
 110 #endif
 111
 112 #ifdef CALC_COUL_EWALD
 113     SimdReal beta2_S, beta_S;
 114 #endif
 115
 116 #if defined CALC_ENERGIES && (defined CALC_COUL_EWALD || defined CALC_COUL_TAB)
 117     SimdReal sh_ewald_S;
 118 #endif
 119
 120 #if defined LJ_CUT && defined CALC_ENERGIES
 121     SimdReal p6_cpot_S, p12_cpot_S;
 122 #endif
 123 #ifdef LJ_POT_SWITCH
 124     SimdReal rswitch_S;
 125     SimdReal swV3_S, swV4_S, swV5_S;
 126     SimdReal swF2_S, swF3_S, swF4_S;
 127 #endif
 128 #ifdef LJ_FORCE_SWITCH
 129     SimdReal rswitch_S;
 130     SimdReal p6_fc2_S, p6_fc3_S;
 131     SimdReal p12_fc2_S, p12_fc3_S;
 132 #    ifdef CALC_ENERGIES
 133     SimdReal p6_vc3_S, p6_vc4_S;
 134     SimdReal p12_vc3_S, p12_vc4_S;
 135     SimdReal p6_6cpot_S, p12_12cpot_S;
 136 #    endif
 137 #endif
 138 #ifdef LJ_EWALD_GEOM
 139     real     lj_ewaldcoeff2, lj_ewaldcoeff6_6;
 140     SimdReal mone_S, half_S, lje_c2_S, lje_c6_6_S;
 141 #endif
 142
 143 #ifdef LJ_COMB_LB
 144     SimdReal hsig_i_S0, seps_i_S0;
 145     SimdReal hsig_i_S2, seps_i_S2;
 146 #else
 147 #    ifdef FIX_LJ_C
 148     alignas(GMX_SIMD_ALIGNMENT) real pvdw_c6[2 * UNROLLI * UNROLLJ];
 149     real*                            pvdw_c12          = pvdw_c6 + UNROLLI * UNROLLJ;
 150 #    endif
 151 #endif /* LJ_COMB_LB */
 152
 153     SimdReal minRsq_S;
 154     SimdReal rc2_S;
 155 #ifdef VDW_CUTOFF_CHECK
 156     SimdReal rcvdw2_S;
 157 #endif
 158
 159     int ninner;
 160
 161 #ifdef COUNT_PAIRS
 162     int npair = 0;
 163 #endif
 164
 165     const nbnxn_atomdata_t::Params& nbatParams = nbat->params();
 166
 167 #if defined LJ_COMB_GEOM || defined LJ_COMB_LB || defined LJ_EWALD_GEOM
 168     const real* gmx_restrict ljc = nbatParams.lj_comb.data();
 169 #endif
 170 #if !(defined LJ_COMB_GEOM || defined LJ_COMB_LB || defined FIX_LJ_C)
 171     /* No combination rule used */
 172     const real* gmx_restrict nbfp_ptr = nbatParams.nbfp_aligned.data();
 173     const int* gmx_restrict type      = nbatParams.type.data();
 174 #endif
 175
 176     /* Load j-i for the first i */
 177     diagonal_jmi_S = load<SimdReal>(nbat->simdMasks.diagonal_2xnn_j_minus_i.data());
 178     /* Generate all the diagonal masks as comparison results */
 179 #if UNROLLI == UNROLLJ
 180     diagonal_mask_S0 = (zero_S < diagonal_jmi_S);
 181     diagonal_jmi_S   = diagonal_jmi_S - one_S;
 182     diagonal_jmi_S   = diagonal_jmi_S - one_S;
 183     diagonal_mask_S2 = (zero_S < diagonal_jmi_S);
 184 #else
 185 #    if 2 * UNROLLI == UNROLLJ
 186     diagonal_mask0_S0                                  = (zero_S < diagonal_jmi_S);
 187     diagonal_jmi_S                                     = diagonal_jmi_S - one_S;
 188     diagonal_jmi_S                                     = diagonal_jmi_S - one_S;
 189     diagonal_mask0_S2                                  = (zero_S < diagonal_jmi_S);
 190     diagonal_jmi_S                                     = diagonal_jmi_S - one_S;
 191     diagonal_jmi_S                                     = diagonal_jmi_S - one_S;
 192     diagonal_mask1_S0                                  = (zero_S < diagonal_jmi_S);
 193     diagonal_jmi_S                                     = diagonal_jmi_S - one_S;
 194     diagonal_jmi_S                                     = diagonal_jmi_S - one_S;
 195     diagonal_mask1_S2                                  = (zero_S < diagonal_jmi_S);
 196 #    endif
 197 #endif
 198
 199     /* Load masks for topology exclusion masking. filter_stride is
 200        static const, so the conditional will be optimized away. */
 201 #if GMX_DOUBLE && !GMX_SIMD_HAVE_INT32_LOGICAL
 202     const std::uint64_t* gmx_restrict exclusion_filter = nbat->simdMasks.exclusion_filter64.data();
 203 #else
 204     const std::uint32_t* gmx_restrict exclusion_filter = nbat->simdMasks.exclusion_filter.data();
 205 #endif
 206
 207     /* Here we cast the exclusion filters from unsigned * to int * or real *.
 208      * Since we only check bits, the actual value they represent does not
 209      * matter, as long as both filter and mask data are treated the same way.
 210      */
 211 #if GMX_SIMD_HAVE_INT32_LOGICAL
 212     filter_S0 = load<SimdBitMask>(reinterpret_cast<const int*>(exclusion_filter + 0 * UNROLLJ));
 213     filter_S2 = load<SimdBitMask>(reinterpret_cast<const int*>(exclusion_filter + 2 * UNROLLJ));
 214 #else
 215     filter_S0 = load<SimdBitMask>(reinterpret_cast<const real*>(exclusion_filter + 0 * UNROLLJ));
 216     filter_S2 = load<SimdBitMask>(reinterpret_cast<const real*>(exclusion_filter + 2 * UNROLLJ));
 217 #endif
 218
 219 #ifdef CALC_COUL_RF
 220     /* Reaction-field constants */
 221     mrc_3_S = SimdReal(-2 * ic->k_rf);
 222 #    ifdef CALC_ENERGIES
 223     hrc_3_S  = SimdReal(ic->k_rf);
 224     moh_rc_S = SimdReal(-ic->c_rf);
 225 #    endif
 226 #endif
 227
 228 #ifdef CALC_COUL_TAB
 229
 230     invtsp_S = SimdReal(ic->coulombEwaldTables->scale);
 231 #    ifdef CALC_ENERGIES
 232     mhalfsp_S = SimdReal(-0.5_real / ic->coulombEwaldTables->scale);
 233 #    endif
 234
 235 #    ifdef TAB_FDV0
 236     tab_coul_F = ic->coulombEwaldTables->tableFDV0.data();
 237 #    else
 238     tab_coul_F = ic->coulombEwaldTables->tableF.data();
 239 #        ifdef CALC_ENERGIES
 240     tab_coul_V = ic->coulombEwaldTables->tableV.data();
 241 #        endif
 242 #    endif
 243 #endif /* CALC_COUL_TAB */
 244
 245 #ifdef CALC_COUL_EWALD
 246     beta2_S = SimdReal(ic->ewaldcoeff_q * ic->ewaldcoeff_q);
 247     beta_S  = SimdReal(ic->ewaldcoeff_q);
 248 #endif
 249
 250 #if (defined CALC_COUL_TAB || defined CALC_COUL_EWALD) && defined CALC_ENERGIES
 251     sh_ewald_S = SimdReal(ic->sh_ewald);
 252 #endif
 253
 254     /* LJ function constants */
 255 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
 256     SimdReal sixth_S    = SimdReal(1.0 / 6.0);
 257     SimdReal twelveth_S = SimdReal(1.0 / 12.0);
 258 #endif
 259
 260 #if defined LJ_CUT && defined CALC_ENERGIES
 261     /* We shift the potential by cpot, which can be zero */
 262     p6_cpot_S  = SimdReal(ic->dispersion_shift.cpot);
 263     p12_cpot_S = SimdReal(ic->repulsion_shift.cpot);
 264 #endif
 265 #ifdef LJ_POT_SWITCH
 266     rswitch_S = SimdReal(ic->rvdw_switch);
 267     swV3_S    = SimdReal(ic->vdw_switch.c3);
 268     swV4_S    = SimdReal(ic->vdw_switch.c4);
 269     swV5_S    = SimdReal(ic->vdw_switch.c5);
 270     swF2_S    = SimdReal(3 * ic->vdw_switch.c3);
 271     swF3_S    = SimdReal(4 * ic->vdw_switch.c4);
 272     swF4_S    = SimdReal(5 * ic->vdw_switch.c5);
 273 #endif
 274 #ifdef LJ_FORCE_SWITCH
 275     rswitch_S = SimdReal(ic->rvdw_switch);
 276     p6_fc2_S  = SimdReal(ic->dispersion_shift.c2);
 277     p6_fc3_S  = SimdReal(ic->dispersion_shift.c3);
 278     p12_fc2_S = SimdReal(ic->repulsion_shift.c2);
 279     p12_fc3_S = SimdReal(ic->repulsion_shift.c3);
 280 #    ifdef CALC_ENERGIES
 281     {
 282         SimdReal mthird_S  = SimdReal(-1.0 / 3.0);
 283         SimdReal mfourth_S = SimdReal(-1.0 / 4.0);
 284
 285         p6_vc3_S     = mthird_S * p6_fc2_S;
 286         p6_vc4_S     = mfourth_S * p6_fc3_S;
 287         p6_6cpot_S   = SimdReal(ic->dispersion_shift.cpot / 6);
 288         p12_vc3_S    = mthird_S * p12_fc2_S;
 289         p12_vc4_S    = mfourth_S * p12_fc3_S;
 290         p12_12cpot_S = SimdReal(ic->repulsion_shift.cpot / 12);
 291     }
 292 #    endif
 293 #endif
 294 #ifdef LJ_EWALD_GEOM
 295     mone_S           = SimdReal(-1.0);
 296     half_S           = SimdReal(0.5);
 297     lj_ewaldcoeff2   = ic->ewaldcoeff_lj * ic->ewaldcoeff_lj;
 298     lj_ewaldcoeff6_6 = lj_ewaldcoeff2 * lj_ewaldcoeff2 * lj_ewaldcoeff2 / 6;
 299     lje_c2_S         = SimdReal(lj_ewaldcoeff2);
 300     lje_c6_6_S       = SimdReal(lj_ewaldcoeff6_6);
 301 #    ifdef CALC_ENERGIES
 302     /* Determine the grid potential at the cut-off */
 303     SimdReal lje_vc_S = SimdReal(ic->sh_lj_ewald);
 304 #    endif
 305 #endif
 306
 307     /* The kernel either supports rcoulomb = rvdw or rcoulomb >= rvdw */
 308     rc2_S = SimdReal(ic->rcoulomb * ic->rcoulomb);
 309 #ifdef VDW_CUTOFF_CHECK
 310     rcvdw2_S = SimdReal(ic->rvdw * ic->rvdw);
 311 #endif
 312
 313     minRsq_S = SimdReal(NBNXN_MIN_RSQ);
 314
 315     const real* gmx_restrict q        = nbatParams.q.data();
 316     const real               facel    = ic->epsfac;
 317     const real* gmx_restrict shiftvec = shift_vec[0];
 318     const real* gmx_restrict x        = nbat->x().data();
 319
 320 #ifdef FIX_LJ_C
 321
 322     for (jp = 0; jp < UNROLLJ; jp++)
 323     {
 324         pvdw_c6[0 * UNROLLJ + jp] = nbat->nbfp[0 * 2];
 325         pvdw_c6[1 * UNROLLJ + jp] = nbat->nbfp[0 * 2];
 326         pvdw_c6[2 * UNROLLJ + jp] = nbat->nbfp[0 * 2];
 327         pvdw_c6[3 * UNROLLJ + jp] = nbat->nbfp[0 * 2];
 328
 329         pvdw_c12[0 * UNROLLJ + jp] = nbat->nbfp[0 * 2 + 1];
 330         pvdw_c12[1 * UNROLLJ + jp] = nbat->nbfp[0 * 2 + 1];
 331         pvdw_c12[2 * UNROLLJ + jp] = nbat->nbfp[0 * 2 + 1];
 332         pvdw_c12[3 * UNROLLJ + jp] = nbat->nbfp[0 * 2 + 1];
 333     }
 334     SimdReal c6_S0 = load<SimdReal>(pvdw_c6 + 0 * UNROLLJ);
 335     SimdReal c6_S1 = load<SimdReal>(pvdw_c6 + 1 * UNROLLJ);
 336     SimdReal c6_S2 = load<SimdReal>(pvdw_c6 + 2 * UNROLLJ);
 337     SimdReal c6_S3 = load<SimdReal>(pvdw_c6 + 3 * UNROLLJ);
 338
 339     SimdReal c12_S0 = load<SimdReal>(pvdw_c12 + 0 * UNROLLJ);
 340     SimdReal c12_S1 = load<SimdReal>(pvdw_c12 + 1 * UNROLLJ);
 341     SimdReal c12_S2 = load<SimdReal>(pvdw_c12 + 2 * UNROLLJ);
 342     SimdReal c12_S3 = load<SimdReal>(pvdw_c12 + 3 * UNROLLJ);
 343 #endif /* FIX_LJ_C */
 344
 345 #ifdef ENERGY_GROUPS
 346     egps_ishift  = nbatParams.neg_2log;
 347     egps_imask   = (1 << egps_ishift) - 1;
 348     egps_jshift  = 2 * nbatParams.neg_2log;
 349     egps_jmask   = (1 << egps_jshift) - 1;
 350     egps_jstride = (UNROLLJ >> 1) * UNROLLJ;
 351     /* Major division is over i-particle energy groups, determine the stride */
 352     Vstride_i = nbatParams.nenergrp * (1 << nbatParams.neg_2log) * egps_jstride;
 353 #endif
 354
 355     l_cj = nbl->cj.data();
 356
 357     ninner = 0;
 358     for (const nbnxn_ci_t& ciEntry : nbl->ci)
 359     {
 360         ish    = (ciEntry.shift & NBNXN_CI_SHIFT);
 361         ish3   = ish * 3;
 362         cjind0 = ciEntry.cj_ind_start;
 363         cjind1 = ciEntry.cj_ind_end;
 364         ci     = ciEntry.ci;
 365         ci_sh  = (ish == CENTRAL ? ci : -1);
 366
 367         shX_S = SimdReal(shiftvec[ish3]);
 368         shY_S = SimdReal(shiftvec[ish3 + 1]);
 369         shZ_S = SimdReal(shiftvec[ish3 + 2]);
 370
 371 #if UNROLLJ <= 4
 372         int sci  = ci * STRIDE;
 373         int scix = sci * DIM;
 374 #    if defined LJ_COMB_LB || defined LJ_COMB_GEOM || defined LJ_EWALD_GEOM
 375         int sci2 = sci * 2;
 376 #    endif
 377 #else
 378         int sci  = (ci >> 1) * STRIDE;
 379         int scix = sci * DIM + (ci & 1) * (STRIDE >> 1);
 380 #    if defined LJ_COMB_LB || defined LJ_COMB_GEOM || defined LJ_EWALD_GEOM
 381         int sci2 = sci * 2 + (ci & 1) * (STRIDE >> 1);
 382 #    endif
 383         sci += (ci & 1) * (STRIDE >> 1);
 384 #endif
 385
 386         /* We have 5 LJ/C combinations, but use only three inner loops,
 387          * as the other combinations are unlikely and/or not much faster:
 388          * inner half-LJ + C for half-LJ + C / no-LJ + C
 389          * inner LJ + C      for full-LJ + C
 390          * inner LJ          for full-LJ + no-C / half-LJ + no-C
 391          */
 392         do_LJ   = ((ciEntry.shift & NBNXN_CI_DO_LJ(0)) != 0);
 393         do_coul = ((ciEntry.shift & NBNXN_CI_DO_COUL(0)) != 0);
 394         half_LJ = (((ciEntry.shift & NBNXN_CI_HALF_LJ(0)) != 0) || !do_LJ) && do_coul;
 395
 396 #ifdef ENERGY_GROUPS
 397         egps_i = nbatParams.energrp[ci];
 398         {
 399             int ia, egp_ia;
 400
 401             for (ia = 0; ia < UNROLLI; ia++)
 402             {
 403                 egp_ia     = (egps_i >> (ia * egps_ishift)) & egps_imask;
 404                 vvdwtp[ia] = Vvdw + egp_ia * Vstride_i;
 405                 vctp[ia]   = Vc + egp_ia * Vstride_i;
 406             }
 407         }
 408 #endif
 409
 410 #ifdef CALC_ENERGIES
 411 #    ifdef LJ_EWALD_GEOM
 412         gmx_bool do_self = TRUE;
 413 #    else
 414         gmx_bool do_self = do_coul;
 415 #    endif
 416 #    if UNROLLJ == 4
 417         if (do_self && l_cj[ciEntry.cj_ind_start].cj == ci_sh)
 418 #    endif
 419 #    if UNROLLJ == 8
 420             if (do_self && l_cj[ciEntry.cj_ind_start].cj == (ci_sh >> 1))
 421 #    endif
 422             {
 423                 if (do_coul)
 424                 {
 425                     real Vc_sub_self;
 426                     int  ia;
 427
 428 #    ifdef CALC_COUL_RF
 429                     Vc_sub_self = 0.5 * ic->c_rf;
 430 #    endif
 431 #    ifdef CALC_COUL_TAB
 432 #        ifdef TAB_FDV0
 433                     Vc_sub_self = 0.5 * tab_coul_F[2];
 434 #        else
 435                     Vc_sub_self = 0.5 * tab_coul_V[0];
 436 #        endif
 437 #    endif
 438 #    ifdef CALC_COUL_EWALD
 439                     /* beta/sqrt(pi) */
 440                     Vc_sub_self = 0.5 * ic->ewaldcoeff_q * M_2_SQRTPI;
 441 #    endif
 442
 443                     for (ia = 0; ia < UNROLLI; ia++)
 444                     {
 445                         real qi;
 446
 447                         qi = q[sci + ia];
 448 #    ifdef ENERGY_GROUPS
 449                         vctp[ia][((egps_i >> (ia * egps_ishift)) & egps_imask) * egps_jstride]
 450 #    else
 451                     Vc[0]
 452 #    endif
 453                                 -= facel * qi * qi * Vc_sub_self;
 454                     }
 455                 }
 456
 457 #    ifdef LJ_EWALD_GEOM
 458                 {
 459                     int ia;
 460
 461                     for (ia = 0; ia < UNROLLI; ia++)
 462                     {
 463                         real c6_i;
 464
 465                         c6_i = nbatParams.nbfp[nbatParams.type[sci + ia] * (nbatParams.numTypes + 1) * 2]
 466                                / 6;
 467 #        ifdef ENERGY_GROUPS
 468                         vvdwtp[ia][((egps_i >> (ia * egps_ishift)) & egps_imask) * egps_jstride]
 469 #        else
 470                         Vvdw[0]
 471 #        endif
 472                                 += 0.5 * c6_i * lj_ewaldcoeff6_6;
 473                     }
 474                 }
 475 #    endif /* LJ_EWALD */
 476             }
 477 #endif
 478
 479         /* Load i atom data */
 480         int sciy = scix + STRIDE;
 481         int sciz = sciy + STRIDE;
 482         ix_S0    = loadU1DualHsimd(x + scix);
 483         ix_S2    = loadU1DualHsimd(x + scix + 2);
 484         iy_S0    = loadU1DualHsimd(x + sciy);
 485         iy_S2    = loadU1DualHsimd(x + sciy + 2);
 486         iz_S0    = loadU1DualHsimd(x + sciz);
 487         iz_S2    = loadU1DualHsimd(x + sciz + 2);
 488         ix_S0    = ix_S0 + shX_S;
 489         ix_S2    = ix_S2 + shX_S;
 490         iy_S0    = iy_S0 + shY_S;
 491         iy_S2    = iy_S2 + shY_S;
 492         iz_S0    = iz_S0 + shZ_S;
 493         iz_S2    = iz_S2 + shZ_S;
 494
 495         if (do_coul)
 496         {
 497             SimdReal facel_S;
 498
 499             facel_S = SimdReal(facel);
 500
 501             iq_S0 = loadU1DualHsimd(q + sci);
 502             iq_S2 = loadU1DualHsimd(q + sci + 2);
 503             iq_S0 = facel_S * iq_S0;
 504             iq_S2 = facel_S * iq_S2;
 505         }
 506
 507 #ifdef LJ_COMB_LB
 508         hsig_i_S0 = loadU1DualHsimd(ljc + sci2);
 509         hsig_i_S2 = loadU1DualHsimd(ljc + sci2 + 2);
 510         seps_i_S0 = loadU1DualHsimd(ljc + sci2 + STRIDE);
 511         seps_i_S2 = loadU1DualHsimd(ljc + sci2 + STRIDE + 2);
 512 #else
 513 #    ifdef LJ_COMB_GEOM
 514         SimdReal c6s_S0, c12s_S0;
 515         SimdReal c6s_S2, c12s_S2;
 516
 517         c6s_S0 = loadU1DualHsimd(ljc + sci2);
 518
 519         if (!half_LJ)
 520         {
 521             c6s_S2 = loadU1DualHsimd(ljc + sci2 + 2);
 522         }
 523         c12s_S0 = loadU1DualHsimd(ljc + sci2 + STRIDE);
 524         if (!half_LJ)
 525         {
 526             c12s_S2 = loadU1DualHsimd(ljc + sci2 + STRIDE + 2);
 527         }
 528 #    elif !defined LJ_COMB_LB && !defined FIX_LJ_C
 529         const int   numTypes = nbatParams.numTypes;
 530         const real* nbfp0    = nbfp_ptr + type[sci] * numTypes * c_simdBestPairAlignment;
 531         const real* nbfp1    = nbfp_ptr + type[sci + 1] * numTypes * c_simdBestPairAlignment;
 532         const real *nbfp2 = nullptr, *nbfp3 = nullptr;
 533         if (!half_LJ)
 534         {
 535             nbfp2 = nbfp_ptr + type[sci + 2] * numTypes * c_simdBestPairAlignment;
 536             nbfp3 = nbfp_ptr + type[sci + 3] * numTypes * c_simdBestPairAlignment;
 537         }
 538 #    endif
 539 #endif
 540 #ifdef LJ_EWALD_GEOM
 541         /* We need the geometrically combined C6 for the PME grid correction */
 542         SimdReal c6s_S0, c6s_S2;
 543         c6s_S0 = loadU1DualHsimd(ljc + sci2);
 544         if (!half_LJ)
 545         {
 546             c6s_S2 = loadU1DualHsimd(ljc + sci2 + 2);
 547         }
 548 #endif
 549
 550         /* Zero the potential energy for this list */
 551 #ifdef CALC_ENERGIES
 552         SimdReal Vvdwtot_S = setZero();
 553         SimdReal vctot_S   = setZero();
 554 #endif
 555
 556         /* Clear i atom forces */
 557         fix_S0 = setZero();
 558         fix_S2 = setZero();
 559         fiy_S0 = setZero();
 560         fiy_S2 = setZero();
 561         fiz_S0 = setZero();
 562         fiz_S2 = setZero();
 563
 564         cjind = cjind0;
 565
 566         /* Currently all kernels use (at least half) LJ */
 567 #define CALC_LJ
 568         if (half_LJ)
 569         {
 570             /* Coulomb: all i-atoms, LJ: first half i-atoms */
 571 #define CALC_COULOMB
 572 #define HALF_LJ
 573 #define CHECK_EXCLS
 574             while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
 575             {
 576 #include "kernel_inner.h"
 577                 cjind++;
 578             }
 579 #undef CHECK_EXCLS
 580             for (; (cjind < cjind1); cjind++)
 581             {
 582 #include "kernel_inner.h"
 583             }
 584 #undef HALF_LJ
 585 #undef CALC_COULOMB
 586         }
 587         else if (do_coul)
 588         {
 589             /* Coulomb: all i-atoms, LJ: all i-atoms */
 590 #define CALC_COULOMB
 591 #define CHECK_EXCLS
 592             while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
 593             {
 594 #include "kernel_inner.h"
 595                 cjind++;
 596             }
 597 #undef CHECK_EXCLS
 598             for (; (cjind < cjind1); cjind++)
 599             {
 600 #include "kernel_inner.h"
 601             }
 602 #undef CALC_COULOMB
 603         }
 604         else
 605         {
 606             /* Coulomb: none, LJ: all i-atoms */
 607 #define CHECK_EXCLS
 608             while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
 609             {
 610 #include "kernel_inner.h"
 611                 cjind++;
 612             }
 613 #undef CHECK_EXCLS
 614             for (; (cjind < cjind1); cjind++)
 615             {
 616 #include "kernel_inner.h"
 617             }
 618         }
 619 #undef CALC_LJ
 620         ninner += cjind1 - cjind0;
 621
 622         /* Add accumulated i-forces to the force array */
 623         real fShiftX = reduceIncr4ReturnSumHsimd(f + scix, fix_S0, fix_S2);
 624         real fShiftY = reduceIncr4ReturnSumHsimd(f + sciy, fiy_S0, fiy_S2);
 625         real fShiftZ = reduceIncr4ReturnSumHsimd(f + sciz, fiz_S0, fiz_S2);
 626
 627 #ifdef CALC_SHIFTFORCES
 628         fshift[ish3 + 0] += fShiftX;
 629         fshift[ish3 + 1] += fShiftY;
 630         fshift[ish3 + 2] += fShiftZ;
 631 #endif
 632
 633 #ifdef CALC_ENERGIES
 634         if (do_coul)
 635         {
 636             *Vc += reduce(vctot_S);
 637         }
 638         *Vvdw += reduce(Vvdwtot_S);
 639 #endif
 640
 641         /* Outer loop uses 6 flops/iteration */
 642     }
 643
 644 #ifdef COUNT_PAIRS
 645     printf("atom pairs %d\n", npair);
 646 #endif
 647 }