src/gromacs/nbnxm/kernels_simd_4xm/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_S1, iy_S1, iz_S1;
  73     SimdReal ix_S2, iy_S2, iz_S2;
  74     SimdReal ix_S3, iy_S3, iz_S3;
  75     SimdReal fix_S0, fiy_S0, fiz_S0;
  76     SimdReal fix_S1, fiy_S1, fiz_S1;
  77     SimdReal fix_S2, fiy_S2, fiz_S2;
  78     SimdReal fix_S3, fiy_S3, fiz_S3;
  79
  80     SimdReal diagonal_jmi_S;
  81 #if UNROLLI == UNROLLJ
  82     SimdBool diagonal_mask_S0, diagonal_mask_S1, diagonal_mask_S2, diagonal_mask_S3;
  83 #else
  84     SimdBool diagonal_mask0_S0, diagonal_mask0_S1, diagonal_mask0_S2, diagonal_mask0_S3;
  85     SimdBool diagonal_mask1_S0, diagonal_mask1_S1, diagonal_mask1_S2, diagonal_mask1_S3;
  86 #endif
  87
  88     SimdBitMask filter_S0, filter_S1, filter_S2, filter_S3;
  89
  90     SimdReal zero_S(0.0);
  91
  92     SimdReal one_S(1.0);
  93     SimdReal iq_S0 = setZero();
  94     SimdReal iq_S1 = setZero();
  95     SimdReal iq_S2 = setZero();
  96     SimdReal iq_S3 = setZero();
  97
  98 #ifdef CALC_COUL_RF
  99     SimdReal mrc_3_S;
 100 #    ifdef CALC_ENERGIES
 101     SimdReal hrc_3_S, moh_rc_S;
 102 #    endif
 103 #endif
 104
 105 #ifdef CALC_COUL_TAB
 106     /* Coulomb table variables */
 107     SimdReal    invtsp_S;
 108     const real* tab_coul_F;
 109 #    if defined CALC_ENERGIES && !defined TAB_FDV0
 110     const real gmx_unused* tab_coul_V;
 111 #    endif
 112 #    ifdef CALC_ENERGIES
 113     SimdReal mhalfsp_S;
 114 #    endif
 115 #endif
 116
 117 #ifdef CALC_COUL_EWALD
 118     SimdReal beta2_S, beta_S;
 119 #endif
 120
 121 #if defined CALC_ENERGIES && (defined CALC_COUL_EWALD || defined CALC_COUL_TAB)
 122     SimdReal sh_ewald_S;
 123 #endif
 124
 125 #if defined LJ_CUT && defined CALC_ENERGIES
 126     SimdReal p6_cpot_S, p12_cpot_S;
 127 #endif
 128 #ifdef LJ_POT_SWITCH
 129     SimdReal rswitch_S;
 130     SimdReal swV3_S, swV4_S, swV5_S;
 131     SimdReal swF2_S, swF3_S, swF4_S;
 132 #endif
 133 #ifdef LJ_FORCE_SWITCH
 134     SimdReal rswitch_S;
 135     SimdReal p6_fc2_S, p6_fc3_S;
 136     SimdReal p12_fc2_S, p12_fc3_S;
 137 #    ifdef CALC_ENERGIES
 138     SimdReal p6_vc3_S, p6_vc4_S;
 139     SimdReal p12_vc3_S, p12_vc4_S;
 140     SimdReal p6_6cpot_S, p12_12cpot_S;
 141 #    endif
 142 #endif
 143 #ifdef LJ_EWALD_GEOM
 144     real     lj_ewaldcoeff2, lj_ewaldcoeff6_6;
 145     SimdReal mone_S, half_S, lje_c2_S, lje_c6_6_S;
 146 #endif
 147
 148 #ifdef LJ_COMB_LB
 149     SimdReal hsig_i_S0, seps_i_S0;
 150     SimdReal hsig_i_S1, seps_i_S1;
 151     SimdReal hsig_i_S2, seps_i_S2;
 152     SimdReal hsig_i_S3, seps_i_S3;
 153 #endif /* LJ_COMB_LB */
 154
 155     SimdReal minRsq_S;
 156     SimdReal rc2_S;
 157 #ifdef VDW_CUTOFF_CHECK
 158     SimdReal rcvdw2_S;
 159 #endif
 160
 161     int ninner;
 162
 163 #ifdef COUNT_PAIRS
 164     int npair = 0;
 165 #endif
 166
 167     const nbnxn_atomdata_t::Params& nbatParams = nbat->params();
 168
 169 #if defined LJ_COMB_GEOM || defined LJ_COMB_LB || defined LJ_EWALD_GEOM
 170     const real* gmx_restrict ljc = nbatParams.lj_comb.data();
 171 #endif
 172 #if !(defined LJ_COMB_GEOM || defined LJ_COMB_LB || defined FIX_LJ_C)
 173     /* No combination rule used */
 174     const real* gmx_restrict nbfp_ptr = nbatParams.nbfp_aligned.data();
 175     const int* gmx_restrict type      = nbatParams.type.data();
 176 #endif
 177
 178     /* Load j-i for the first i */
 179     diagonal_jmi_S = load<SimdReal>(nbat->simdMasks.diagonal_4xn_j_minus_i.data());
 180     /* Generate all the diagonal masks as comparison results */
 181 #if UNROLLI == UNROLLJ
 182     diagonal_mask_S0 = (zero_S < diagonal_jmi_S);
 183     diagonal_jmi_S   = diagonal_jmi_S - one_S;
 184     diagonal_mask_S1 = (zero_S < diagonal_jmi_S);
 185     diagonal_jmi_S   = diagonal_jmi_S - one_S;
 186     diagonal_mask_S2 = (zero_S < diagonal_jmi_S);
 187     diagonal_jmi_S   = diagonal_jmi_S - one_S;
 188     diagonal_mask_S3 = (zero_S < diagonal_jmi_S);
 189 #else
 190 #    if UNROLLI == 2 * UNROLLJ || 2 * UNROLLI == UNROLLJ
 191     diagonal_mask0_S0 = (zero_S < diagonal_jmi_S);
 192     diagonal_jmi_S    = diagonal_jmi_S - one_S;
 193     diagonal_mask0_S1 = (zero_S < diagonal_jmi_S);
 194     diagonal_jmi_S    = diagonal_jmi_S - one_S;
 195     diagonal_mask0_S2 = (zero_S < diagonal_jmi_S);
 196     diagonal_jmi_S    = diagonal_jmi_S - one_S;
 197     diagonal_mask0_S3 = (zero_S < diagonal_jmi_S);
 198     diagonal_jmi_S    = diagonal_jmi_S - one_S;
 199
 200 #        if UNROLLI == 2 * UNROLLJ
 201     /* Load j-i for the second half of the j-cluster */
 202     diagonal_jmi_S = load<SimdReal>(nbat->simdMasks.diagonal_4xn_j_minus_i.data() + UNROLLJ);
 203 #        endif
 204
 205     diagonal_mask1_S0                                  = (zero_S < diagonal_jmi_S);
 206     diagonal_jmi_S                                     = diagonal_jmi_S - one_S;
 207     diagonal_mask1_S1                                  = (zero_S < diagonal_jmi_S);
 208     diagonal_jmi_S                                     = diagonal_jmi_S - one_S;
 209     diagonal_mask1_S2                                  = (zero_S < diagonal_jmi_S);
 210     diagonal_jmi_S                                     = diagonal_jmi_S - one_S;
 211     diagonal_mask1_S3                                  = (zero_S < diagonal_jmi_S);
 212 #    endif
 213 #endif
 214
 215 #if GMX_DOUBLE && !GMX_SIMD_HAVE_INT32_LOGICAL
 216     const std::uint64_t* gmx_restrict exclusion_filter = nbat->simdMasks.exclusion_filter64.data();
 217 #else
 218     const std::uint32_t* gmx_restrict exclusion_filter = nbat->simdMasks.exclusion_filter.data();
 219 #endif
 220
 221     /* Here we cast the exclusion filters from unsigned * to int * or real *.
 222      * Since we only check bits, the actual value they represent does not
 223      * matter, as long as both filter and mask data are treated the same way.
 224      */
 225 #if GMX_SIMD_HAVE_INT32_LOGICAL
 226     filter_S0 = load<SimdBitMask>(reinterpret_cast<const int*>(exclusion_filter + 0 * UNROLLJ));
 227     filter_S1 = load<SimdBitMask>(reinterpret_cast<const int*>(exclusion_filter + 1 * UNROLLJ));
 228     filter_S2 = load<SimdBitMask>(reinterpret_cast<const int*>(exclusion_filter + 2 * UNROLLJ));
 229     filter_S3 = load<SimdBitMask>(reinterpret_cast<const int*>(exclusion_filter + 3 * UNROLLJ));
 230 #else
 231     filter_S0 = load<SimdBitMask>(reinterpret_cast<const real*>(exclusion_filter + 0 * UNROLLJ));
 232     filter_S1 = load<SimdBitMask>(reinterpret_cast<const real*>(exclusion_filter + 1 * UNROLLJ));
 233     filter_S2 = load<SimdBitMask>(reinterpret_cast<const real*>(exclusion_filter + 2 * UNROLLJ));
 234     filter_S3 = load<SimdBitMask>(reinterpret_cast<const real*>(exclusion_filter + 3 * UNROLLJ));
 235 #endif
 236
 237 #ifdef CALC_COUL_RF
 238     /* Reaction-field constants */
 239     mrc_3_S = SimdReal(-2 * ic->k_rf);
 240 #    ifdef CALC_ENERGIES
 241     hrc_3_S  = SimdReal(ic->k_rf);
 242     moh_rc_S = SimdReal(-ic->c_rf);
 243 #    endif
 244 #endif
 245
 246 #ifdef CALC_COUL_TAB
 247
 248     invtsp_S = SimdReal(ic->coulombEwaldTables->scale);
 249 #    ifdef CALC_ENERGIES
 250     mhalfsp_S = SimdReal(-0.5_real / ic->coulombEwaldTables->scale);
 251 #    endif
 252
 253 #    ifdef TAB_FDV0
 254     tab_coul_F = ic->coulombEwaldTables->tableFDV0.data();
 255 #    else
 256     tab_coul_F = ic->coulombEwaldTables->tableF.data();
 257 #        ifdef CALC_ENERGIES
 258     tab_coul_V = ic->coulombEwaldTables->tableV.data();
 259 #        endif
 260 #    endif
 261 #endif /* CALC_COUL_TAB */
 262
 263 #ifdef CALC_COUL_EWALD
 264     beta2_S = SimdReal(ic->ewaldcoeff_q * ic->ewaldcoeff_q);
 265     beta_S  = SimdReal(ic->ewaldcoeff_q);
 266 #endif
 267
 268 #if (defined CALC_COUL_TAB || defined CALC_COUL_EWALD) && defined CALC_ENERGIES
 269     sh_ewald_S = SimdReal(ic->sh_ewald);
 270 #endif
 271
 272     /* LJ function constants */
 273 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
 274     SimdReal sixth_S(1.0 / 6.0);
 275     SimdReal twelveth_S(1.0 / 12.0);
 276 #endif
 277
 278 #if defined LJ_CUT && defined CALC_ENERGIES
 279     /* We shift the potential by cpot, which can be zero */
 280     p6_cpot_S  = SimdReal(ic->dispersion_shift.cpot);
 281     p12_cpot_S = SimdReal(ic->repulsion_shift.cpot);
 282 #endif
 283 #ifdef LJ_POT_SWITCH
 284     rswitch_S = SimdReal(ic->rvdw_switch);
 285     swV3_S    = SimdReal(ic->vdw_switch.c3);
 286     swV4_S    = SimdReal(ic->vdw_switch.c4);
 287     swV5_S    = SimdReal(ic->vdw_switch.c5);
 288     swF2_S    = SimdReal(3 * ic->vdw_switch.c3);
 289     swF3_S    = SimdReal(4 * ic->vdw_switch.c4);
 290     swF4_S    = SimdReal(5 * ic->vdw_switch.c5);
 291 #endif
 292 #ifdef LJ_FORCE_SWITCH
 293     rswitch_S = SimdReal(ic->rvdw_switch);
 294     p6_fc2_S  = SimdReal(ic->dispersion_shift.c2);
 295     p6_fc3_S  = SimdReal(ic->dispersion_shift.c3);
 296     p12_fc2_S = SimdReal(ic->repulsion_shift.c2);
 297     p12_fc3_S = SimdReal(ic->repulsion_shift.c3);
 298 #    ifdef CALC_ENERGIES
 299     {
 300         SimdReal mthird_S(-1.0 / 3.0);
 301         SimdReal mfourth_S(-1.0 / 4.0);
 302
 303         p6_vc3_S     = mthird_S * p6_fc2_S;
 304         p6_vc4_S     = mfourth_S * p6_fc3_S;
 305         p6_6cpot_S   = SimdReal(ic->dispersion_shift.cpot / 6);
 306         p12_vc3_S    = mthird_S * p12_fc2_S;
 307         p12_vc4_S    = mfourth_S * p12_fc3_S;
 308         p12_12cpot_S = SimdReal(ic->repulsion_shift.cpot / 12);
 309     }
 310 #    endif
 311 #endif
 312 #ifdef LJ_EWALD_GEOM
 313     mone_S           = SimdReal(-1.0);
 314     half_S           = SimdReal(0.5);
 315     lj_ewaldcoeff2   = ic->ewaldcoeff_lj * ic->ewaldcoeff_lj;
 316     lj_ewaldcoeff6_6 = lj_ewaldcoeff2 * lj_ewaldcoeff2 * lj_ewaldcoeff2 / 6;
 317     lje_c2_S         = SimdReal(lj_ewaldcoeff2);
 318     lje_c6_6_S       = SimdReal(lj_ewaldcoeff6_6);
 319 #    ifdef CALC_ENERGIES
 320     /* Determine the grid potential at the cut-off */
 321     SimdReal lje_vc_S(ic->sh_lj_ewald);
 322 #    endif
 323 #endif
 324
 325     /* The kernel either supports rcoulomb = rvdw or rcoulomb >= rvdw */
 326     rc2_S = SimdReal(ic->rcoulomb * ic->rcoulomb);
 327 #ifdef VDW_CUTOFF_CHECK
 328     rcvdw2_S = SimdReal(ic->rvdw * ic->rvdw);
 329 #endif
 330
 331     minRsq_S = SimdReal(NBNXN_MIN_RSQ);
 332
 333     const real* gmx_restrict q        = nbatParams.q.data();
 334     const real               facel    = ic->epsfac;
 335     const real* gmx_restrict shiftvec = shift_vec[0];
 336     const real* gmx_restrict x        = nbat->x().data();
 337
 338 #ifdef FIX_LJ_C
 339     alignas(GMX_SIMD_ALIGNMENT) real pvdw_c6[2 * UNROLLI * UNROLLJ];
 340     real*                            pvdw_c12 = pvdw_c6 + UNROLLI * UNROLLJ;
 341
 342     for (int jp = 0; jp < UNROLLJ; jp++)
 343     {
 344         pvdw_c6[0 * UNROLLJ + jp] = nbat->nbfp[0 * 2];
 345         pvdw_c6[1 * UNROLLJ + jp] = nbat->nbfp[0 * 2];
 346         pvdw_c6[2 * UNROLLJ + jp] = nbat->nbfp[0 * 2];
 347         pvdw_c6[3 * UNROLLJ + jp] = nbat->nbfp[0 * 2];
 348
 349         pvdw_c12[0 * UNROLLJ + jp] = nbat->nbfp[0 * 2 + 1];
 350         pvdw_c12[1 * UNROLLJ + jp] = nbat->nbfp[0 * 2 + 1];
 351         pvdw_c12[2 * UNROLLJ + jp] = nbat->nbfp[0 * 2 + 1];
 352         pvdw_c12[3 * UNROLLJ + jp] = nbat->nbfp[0 * 2 + 1];
 353     }
 354     SimdReal c6_S0 = load<SimdReal>(pvdw_c6 + 0 * UNROLLJ);
 355     SimdReal c6_S1 = load<SimdReal>(pvdw_c6 + 1 * UNROLLJ);
 356     SimdReal c6_S2 = load<SimdReal>(pvdw_c6 + 2 * UNROLLJ);
 357     SimdReal c6_S3 = load<SimdReal>(pvdw_c6 + 3 * UNROLLJ);
 358
 359     SimdReal c12_S0 = load<SimdReal>(pvdw_c12 + 0 * UNROLLJ);
 360     SimdReal c12_S1 = load<SimdReal>(pvdw_c12 + 1 * UNROLLJ);
 361     SimdReal c12_S2 = load<SimdReal>(pvdw_c12 + 2 * UNROLLJ);
 362     SimdReal c12_S3 = load<SimdReal>(pvdw_c12 + 3 * UNROLLJ);
 363 #endif /* FIX_LJ_C */
 364
 365 #ifdef ENERGY_GROUPS
 366     egps_ishift  = nbatParams.neg_2log;
 367     egps_imask   = (1 << egps_ishift) - 1;
 368     egps_jshift  = 2 * nbatParams.neg_2log;
 369     egps_jmask   = (1 << egps_jshift) - 1;
 370     egps_jstride = (UNROLLJ >> 1) * UNROLLJ;
 371     /* Major division is over i-particle energy groups, determine the stride */
 372     Vstride_i = nbatParams.nenergrp * (1 << nbatParams.neg_2log) * egps_jstride;
 373 #endif
 374
 375     l_cj = nbl->cj.data();
 376
 377     ninner = 0;
 378
 379     for (const nbnxn_ci_t& ciEntry : nbl->ci)
 380     {
 381         ish    = (ciEntry.shift & NBNXN_CI_SHIFT);
 382         ish3   = ish * 3;
 383         cjind0 = ciEntry.cj_ind_start;
 384         cjind1 = ciEntry.cj_ind_end;
 385         ci     = ciEntry.ci;
 386         ci_sh  = (ish == CENTRAL ? ci : -1);
 387
 388         shX_S = SimdReal(shiftvec[ish3]);
 389         shY_S = SimdReal(shiftvec[ish3 + 1]);
 390         shZ_S = SimdReal(shiftvec[ish3 + 2]);
 391
 392 #if UNROLLJ <= 4
 393         int sci  = ci * STRIDE;
 394         int scix = sci * DIM;
 395 #    if defined LJ_COMB_LB || defined LJ_COMB_GEOM || defined LJ_EWALD_GEOM
 396         int sci2 = sci * 2;
 397 #    endif
 398 #else
 399         int sci  = (ci >> 1) * STRIDE;
 400         int scix = sci * DIM + (ci & 1) * (STRIDE >> 1);
 401 #    if defined LJ_COMB_LB || defined LJ_COMB_GEOM || defined LJ_EWALD_GEOM
 402         int sci2 = sci * 2 + (ci & 1) * (STRIDE >> 1);
 403 #    endif
 404         sci += (ci & 1) * (STRIDE >> 1);
 405 #endif
 406
 407         /* We have 5 LJ/C combinations, but use only three inner loops,
 408          * as the other combinations are unlikely and/or not much faster:
 409          * inner half-LJ + C for half-LJ + C / no-LJ + C
 410          * inner LJ + C      for full-LJ + C
 411          * inner LJ          for full-LJ + no-C / half-LJ + no-C
 412          */
 413         do_LJ   = ((ciEntry.shift & NBNXN_CI_DO_LJ(0)) != 0);
 414         do_coul = ((ciEntry.shift & NBNXN_CI_DO_COUL(0)) != 0);
 415         half_LJ = (((ciEntry.shift & NBNXN_CI_HALF_LJ(0)) != 0) || !do_LJ) && do_coul;
 416
 417 #ifdef ENERGY_GROUPS
 418         egps_i = nbatParams.energrp[ci];
 419         {
 420             int ia, egp_ia;
 421
 422             for (ia = 0; ia < UNROLLI; ia++)
 423             {
 424                 egp_ia     = (egps_i >> (ia * egps_ishift)) & egps_imask;
 425                 vvdwtp[ia] = Vvdw + egp_ia * Vstride_i;
 426                 vctp[ia]   = Vc + egp_ia * Vstride_i;
 427             }
 428         }
 429 #endif
 430
 431 #ifdef CALC_ENERGIES
 432 #    ifdef LJ_EWALD_GEOM
 433         gmx_bool do_self = TRUE;
 434 #    else
 435         gmx_bool do_self = do_coul;
 436 #    endif
 437 #    if UNROLLJ == 4
 438         if (do_self && l_cj[ciEntry.cj_ind_start].cj == ci_sh)
 439 #    endif
 440 #    if UNROLLJ == 2
 441             if (do_self && l_cj[ciEntry.cj_ind_start].cj == (ci_sh << 1))
 442 #    endif
 443 #    if UNROLLJ == 8
 444                 if (do_self && l_cj[ciEntry.cj_ind_start].cj == (ci_sh >> 1))
 445 #    endif
 446                 {
 447                     if (do_coul)
 448                     {
 449                         real Vc_sub_self;
 450                         int  ia;
 451
 452 #    ifdef CALC_COUL_RF
 453                         Vc_sub_self = 0.5 * ic->c_rf;
 454 #    endif
 455 #    ifdef CALC_COUL_TAB
 456 #        ifdef TAB_FDV0
 457                         Vc_sub_self = 0.5 * tab_coul_F[2];
 458 #        else
 459                         Vc_sub_self = 0.5 * tab_coul_V[0];
 460 #        endif
 461 #    endif
 462 #    ifdef CALC_COUL_EWALD
 463                         /* beta/sqrt(pi) */
 464                         Vc_sub_self = 0.5 * ic->ewaldcoeff_q * M_2_SQRTPI;
 465 #    endif
 466
 467                         for (ia = 0; ia < UNROLLI; ia++)
 468                         {
 469                             real qi;
 470
 471                             qi = q[sci + ia];
 472 #    ifdef ENERGY_GROUPS
 473                             vctp[ia][((egps_i >> (ia * egps_ishift)) & egps_imask) * egps_jstride]
 474 #    else
 475                     Vc[0]
 476 #    endif
 477                                     -= facel * qi * qi * Vc_sub_self;
 478                         }
 479                     }
 480
 481 #    ifdef LJ_EWALD_GEOM
 482                     {
 483                         int ia;
 484
 485                         for (ia = 0; ia < UNROLLI; ia++)
 486                         {
 487                             real c6_i;
 488
 489                             c6_i = nbatParams.nbfp[nbatParams.type[sci + ia] * (nbatParams.numTypes + 1) * 2]
 490                                    / 6;
 491 #        ifdef ENERGY_GROUPS
 492                             vvdwtp[ia][((egps_i >> (ia * egps_ishift)) & egps_imask) * egps_jstride]
 493 #        else
 494                             Vvdw[0]
 495 #        endif
 496                                     += 0.5 * c6_i * lj_ewaldcoeff6_6;
 497                         }
 498                     }
 499 #    endif /* LJ_EWALD_GEOM */
 500                 }
 501 #endif
 502
 503         /* Load i atom data */
 504         int sciy = scix + STRIDE;
 505         int sciz = sciy + STRIDE;
 506         ix_S0    = SimdReal(x[scix]) + shX_S;
 507         ix_S1    = SimdReal(x[scix + 1]) + shX_S;
 508         ix_S2    = SimdReal(x[scix + 2]) + shX_S;
 509         ix_S3    = SimdReal(x[scix + 3]) + shX_S;
 510         iy_S0    = SimdReal(x[sciy]) + shY_S;
 511         iy_S1    = SimdReal(x[sciy + 1]) + shY_S;
 512         iy_S2    = SimdReal(x[sciy + 2]) + shY_S;
 513         iy_S3    = SimdReal(x[sciy + 3]) + shY_S;
 514         iz_S0    = SimdReal(x[sciz]) + shZ_S;
 515         iz_S1    = SimdReal(x[sciz + 1]) + shZ_S;
 516         iz_S2    = SimdReal(x[sciz + 2]) + shZ_S;
 517         iz_S3    = SimdReal(x[sciz + 3]) + shZ_S;
 518
 519         if (do_coul)
 520         {
 521             iq_S0 = SimdReal(facel * q[sci]);
 522             iq_S1 = SimdReal(facel * q[sci + 1]);
 523             iq_S2 = SimdReal(facel * q[sci + 2]);
 524             iq_S3 = SimdReal(facel * q[sci + 3]);
 525         }
 526
 527 #ifdef LJ_COMB_LB
 528         hsig_i_S0 = SimdReal(ljc[sci2 + 0]);
 529         hsig_i_S1 = SimdReal(ljc[sci2 + 1]);
 530         hsig_i_S2 = SimdReal(ljc[sci2 + 2]);
 531         hsig_i_S3 = SimdReal(ljc[sci2 + 3]);
 532         seps_i_S0 = SimdReal(ljc[sci2 + STRIDE + 0]);
 533         seps_i_S1 = SimdReal(ljc[sci2 + STRIDE + 1]);
 534         seps_i_S2 = SimdReal(ljc[sci2 + STRIDE + 2]);
 535         seps_i_S3 = SimdReal(ljc[sci2 + STRIDE + 3]);
 536 #else
 537 #    ifdef LJ_COMB_GEOM
 538         SimdReal c6s_S0 = SimdReal(ljc[sci2 + 0]);
 539         SimdReal c6s_S1 = SimdReal(ljc[sci2 + 1]);
 540         SimdReal c6s_S2, c6s_S3;
 541         if (!half_LJ)
 542         {
 543             c6s_S2 = SimdReal(ljc[sci2 + 2]);
 544             c6s_S3 = SimdReal(ljc[sci2 + 3]);
 545         }
 546         else
 547         {
 548             c6s_S2 = setZero();
 549             c6s_S3 = setZero();
 550         }
 551         SimdReal c12s_S0 = SimdReal(ljc[sci2 + STRIDE + 0]);
 552         SimdReal c12s_S1 = SimdReal(ljc[sci2 + STRIDE + 1]);
 553         SimdReal c12s_S2, c12s_S3;
 554         if (!half_LJ)
 555         {
 556             c12s_S2 = SimdReal(ljc[sci2 + STRIDE + 2]);
 557             c12s_S3 = SimdReal(ljc[sci2 + STRIDE + 3]);
 558         }
 559         else
 560         {
 561             c12s_S2 = setZero();
 562             c12s_S3 = setZero();
 563         }
 564 #    else
 565         const int   numTypes = nbatParams.numTypes;
 566         const real* nbfp0    = nbfp_ptr + type[sci] * numTypes * c_simdBestPairAlignment;
 567         const real* nbfp1    = nbfp_ptr + type[sci + 1] * numTypes * c_simdBestPairAlignment;
 568         const real *nbfp2 = nullptr, *nbfp3 = nullptr;
 569         if (!half_LJ)
 570         {
 571             nbfp2 = nbfp_ptr + type[sci + 2] * numTypes * c_simdBestPairAlignment;
 572             nbfp3 = nbfp_ptr + type[sci + 3] * numTypes * c_simdBestPairAlignment;
 573         }
 574 #    endif
 575 #endif
 576 #ifdef LJ_EWALD_GEOM
 577         /* We need the geometrically combined C6 for the PME grid correction */
 578         SimdReal c6s_S0 = SimdReal(ljc[sci2 + 0]);
 579         SimdReal c6s_S1 = SimdReal(ljc[sci2 + 1]);
 580         SimdReal c6s_S2, c6s_S3;
 581         if (!half_LJ)
 582         {
 583             c6s_S2 = SimdReal(ljc[sci2 + 2]);
 584             c6s_S3 = SimdReal(ljc[sci2 + 3]);
 585         }
 586 #endif
 587
 588         /* Zero the potential energy for this list */
 589 #ifdef CALC_ENERGIES
 590         SimdReal Vvdwtot_S = setZero();
 591         SimdReal vctot_S   = setZero();
 592 #endif
 593
 594         /* Clear i atom forces */
 595         fix_S0 = setZero();
 596         fix_S1 = setZero();
 597         fix_S2 = setZero();
 598         fix_S3 = setZero();
 599         fiy_S0 = setZero();
 600         fiy_S1 = setZero();
 601         fiy_S2 = setZero();
 602         fiy_S3 = setZero();
 603         fiz_S0 = setZero();
 604         fiz_S1 = setZero();
 605         fiz_S2 = setZero();
 606         fiz_S3 = setZero();
 607
 608         cjind = cjind0;
 609
 610         /* Currently all kernels use (at least half) LJ */
 611 #define CALC_LJ
 612         if (half_LJ)
 613         {
 614             /* Coulomb: all i-atoms, LJ: first half i-atoms */
 615 #define CALC_COULOMB
 616 #define HALF_LJ
 617 #define CHECK_EXCLS
 618             while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
 619             {
 620 #include "kernel_inner.h"
 621                 cjind++;
 622             }
 623 #undef CHECK_EXCLS
 624             for (; (cjind < cjind1); cjind++)
 625             {
 626 #include "kernel_inner.h"
 627             }
 628 #undef HALF_LJ
 629 #undef CALC_COULOMB
 630         }
 631         else if (do_coul)
 632         {
 633             /* Coulomb: all i-atoms, LJ: all i-atoms */
 634 #define CALC_COULOMB
 635 #define CHECK_EXCLS
 636             while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
 637             {
 638 #include "kernel_inner.h"
 639                 cjind++;
 640             }
 641 #undef CHECK_EXCLS
 642             for (; (cjind < cjind1); cjind++)
 643             {
 644 #include "kernel_inner.h"
 645             }
 646 #undef CALC_COULOMB
 647         }
 648         else
 649         {
 650             /* Coulomb: none, LJ: all i-atoms */
 651 #define CHECK_EXCLS
 652             while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
 653             {
 654 #include "kernel_inner.h"
 655                 cjind++;
 656             }
 657 #undef CHECK_EXCLS
 658             for (; (cjind < cjind1); cjind++)
 659             {
 660 #include "kernel_inner.h"
 661             }
 662         }
 663 #undef CALC_LJ
 664         ninner += cjind1 - cjind0;
 665
 666         /* Add accumulated i-forces to the force array */
 667         real fShiftX = reduceIncr4ReturnSum(f + scix, fix_S0, fix_S1, fix_S2, fix_S3);
 668         real fShiftY = reduceIncr4ReturnSum(f + sciy, fiy_S0, fiy_S1, fiy_S2, fiy_S3);
 669         real fShiftZ = reduceIncr4ReturnSum(f + sciz, fiz_S0, fiz_S1, fiz_S2, fiz_S3);
 670
 671
 672 #ifdef CALC_SHIFTFORCES
 673         fshift[ish3 + 0] += fShiftX;
 674         fshift[ish3 + 1] += fShiftY;
 675         fshift[ish3 + 2] += fShiftZ;
 676 #endif
 677
 678 #ifdef CALC_ENERGIES
 679         if (do_coul)
 680         {
 681             *Vc += reduce(vctot_S);
 682         }
 683
 684         *Vvdw += reduce(Vvdwtot_S);
 685 #endif
 686
 687         /* Outer loop uses 6 flops/iteration */
 688     }
 689
 690 #ifdef COUNT_PAIRS
 691     printf("atom pairs %d\n", npair);
 692 #endif
 693 }