src/gromacs/mdlib/nbnxn_kernels/simd_2xnn/nbnxn_kernel_simd_2xnn_outer.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, 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 /* Half-width SIMD operations are required here.
  38  * As the 4xn kernels are the "standard" kernels and some special operations
  39  * are required only here, we define those in nbnxn_kernel_simd_utils_...
  40  *
  41  * Half-width SIMD real type:
  42  * gmx_mm_hpr
  43  *
  44  * Half-width SIMD operations
  45  * Load reals at half-width aligned pointer b into half-width SIMD register a:
  46  * gmx_load_hpr(a, b)
  47  * Set all entries in half-width SIMD register *a to b:
  48  * gmx_set1_hpr(a, b)
  49  * Load one real at b and one real at b+1 into halves of a, respectively:
  50  * gmx_load1p1_pr(a, b)
  51  * Load reals at half-width aligned pointer b into two halves of a:
  52  * gmx_loaddh_pr(a, b)
  53  * Store half-width SIMD register b into half width aligned memory a:
  54  * gmx_store_hpr(a, b)
  55  * gmx_add_hpr(a, b)
  56  * gmx_sub_hpr(a, b)
  57  * Sum over 4 half SIMD registers:
  58  * gmx_sum4_hpr(a, b)
  59  * Sum the elements of halfs of each input register and store sums in out:
  60  * gmx_mm_transpose_sum4h_pr(a, b)
  61  * Extract two half-width registers *b, *c from a full width register a:
  62  * gmx_pr_to_2hpr(a, b, c)
  63  */
  64
  65
  66 {
  67     const nbnxn_ci_t   *nbln;
  68     const nbnxn_cj_t   *l_cj;
  69     const int          *type;
  70     const real         *q;
  71     const real         *shiftvec;
  72     const real         *x;
  73     const real         *nbfp0, *nbfp1, *nbfp2 = NULL, *nbfp3 = NULL;
  74     real                facel;
  75     real               *nbfp_ptr;
  76     int                 n, ci, ci_sh;
  77     int                 ish, ish3;
  78     gmx_bool            do_LJ, half_LJ, do_coul, do_self;
  79     int                 sci, scix, sciy, sciz, sci2;
  80     int                 cjind0, cjind1, cjind;
  81     int                 ip, jp;
  82
  83 #ifdef ENERGY_GROUPS
  84     int         Vstride_i;
  85     int         egps_ishift, egps_imask;
  86     int         egps_jshift, egps_jmask, egps_jstride;
  87     int         egps_i;
  88     real       *vvdwtp[UNROLLI];
  89     real       *vctp[UNROLLI];
  90 #endif
  91
  92     gmx_simd_real_t  shX_S;
  93     gmx_simd_real_t  shY_S;
  94     gmx_simd_real_t  shZ_S;
  95     gmx_simd_real_t  ix_S0, iy_S0, iz_S0;
  96     gmx_simd_real_t  ix_S2, iy_S2, iz_S2;
  97     gmx_simd_real_t  fix_S0, fiy_S0, fiz_S0;
  98     gmx_simd_real_t  fix_S2, fiy_S2, fiz_S2;
  99     /* We use an i-force SIMD register width of 4 */
 100     /* The pr4 stuff is defined in nbnxn_kernel_simd_utils.h */
 101     gmx_simd4_real_t fix_S, fiy_S, fiz_S;
 102
 103     gmx_simd_real_t  diagonal_jmi_S;
 104 #if UNROLLI == UNROLLJ
 105     gmx_simd_bool_t  diagonal_mask_S0, diagonal_mask_S2;
 106 #else
 107     gmx_simd_bool_t  diagonal_mask0_S0, diagonal_mask0_S2;
 108     gmx_simd_bool_t  diagonal_mask1_S0, diagonal_mask1_S2;
 109 #endif
 110
 111     unsigned            *exclusion_filter;
 112     gmx_exclfilter       filter_S0, filter_S2;
 113
 114     gmx_simd_real_t      zero_S = gmx_simd_set1_r(0.0);
 115
 116     gmx_simd_real_t      one_S = gmx_simd_set1_r(1.0);
 117     gmx_simd_real_t      iq_S0 = gmx_simd_setzero_r();
 118     gmx_simd_real_t      iq_S2 = gmx_simd_setzero_r();
 119
 120 #ifdef CALC_COUL_RF
 121     gmx_simd_real_t      mrc_3_S;
 122 #ifdef CALC_ENERGIES
 123     gmx_simd_real_t      hrc_3_S, moh_rc_S;
 124 #endif
 125 #endif
 126
 127 #ifdef CALC_COUL_TAB
 128     /* Coulomb table variables */
 129     gmx_simd_real_t   invtsp_S;
 130     const real       *tab_coul_F;
 131 #ifndef TAB_FDV0
 132     const real       *tab_coul_V;
 133 #endif
 134     /* Thread-local working buffers for force and potential lookups */
 135     int               ti0_array[2*GMX_SIMD_REAL_WIDTH], *ti0 = NULL;
 136     int               ti2_array[2*GMX_SIMD_REAL_WIDTH], *ti2 = NULL;
 137 #ifdef CALC_ENERGIES
 138     gmx_simd_real_t   mhalfsp_S;
 139 #endif
 140 #endif
 141
 142 #ifdef CALC_COUL_EWALD
 143     gmx_simd_real_t beta2_S, beta_S;
 144 #endif
 145
 146 #if defined CALC_ENERGIES && (defined CALC_COUL_EWALD || defined CALC_COUL_TAB)
 147     gmx_simd_real_t  sh_ewald_S;
 148 #endif
 149
 150 #if defined LJ_CUT && defined CALC_ENERGIES
 151     gmx_simd_real_t   p6_cpot_S, p12_cpot_S;
 152 #endif
 153 #ifdef LJ_POT_SWITCH
 154     gmx_simd_real_t   rswitch_S;
 155     gmx_simd_real_t   swV3_S, swV4_S, swV5_S;
 156     gmx_simd_real_t   swF2_S, swF3_S, swF4_S;
 157 #endif
 158 #ifdef LJ_FORCE_SWITCH
 159     gmx_simd_real_t   rswitch_S;
 160     gmx_simd_real_t   p6_fc2_S, p6_fc3_S;
 161     gmx_simd_real_t   p12_fc2_S, p12_fc3_S;
 162 #ifdef CALC_ENERGIES
 163     gmx_simd_real_t   p6_vc3_S, p6_vc4_S;
 164     gmx_simd_real_t   p12_vc3_S, p12_vc4_S;
 165     gmx_simd_real_t   p6_6cpot_S, p12_12cpot_S;
 166 #endif
 167 #endif
 168 #ifdef LJ_EWALD_GEOM
 169     real              lj_ewaldcoeff2, lj_ewaldcoeff6_6;
 170     gmx_simd_real_t   mone_S, half_S, lje_c2_S, lje_c6_6_S, lje_vc_S;
 171 #endif
 172
 173 #ifdef LJ_COMB_LB
 174     const real       *ljc;
 175
 176     gmx_simd_real_t   hsig_i_S0, seps_i_S0;
 177     gmx_simd_real_t   hsig_i_S2, seps_i_S2;
 178 #else
 179 #ifdef FIX_LJ_C
 180     real              pvdw_array[2*UNROLLI*UNROLLJ+GMX_SIMD_REAL_WIDTH];
 181     real             *pvdw_c6, *pvdw_c12;
 182     gmx_simd_real_t   c6_S0, c12_S0;
 183     gmx_simd_real_t   c6_S2, c12_S2;
 184 #endif
 185
 186 #if defined LJ_COMB_GEOM || defined LJ_EWALD_GEOM
 187     const real       *ljc;
 188
 189     gmx_simd_real_t   c6s_S0, c12s_S0;
 190     gmx_simd_real_t   c6s_S2  = gmx_simd_setzero_r();
 191     gmx_simd_real_t   c12s_S2 = gmx_simd_setzero_r();
 192 #endif
 193 #endif /* LJ_COMB_LB */
 194
 195     gmx_simd_real_t  vctot_S, Vvdwtot_S;
 196     gmx_simd_real_t  sixth_S, twelveth_S;
 197
 198     gmx_simd_real_t  avoid_sing_S;
 199     gmx_simd_real_t  rc2_S;
 200 #ifdef VDW_CUTOFF_CHECK
 201     gmx_simd_real_t  rcvdw2_S;
 202 #endif
 203
 204 #ifdef CALC_ENERGIES
 205     /* cppcheck-suppress unassignedVariable */
 206     real       tmpsum_array[2*GMX_SIMD_REAL_WIDTH], *tmpsum;
 207 #endif
 208 #ifdef CALC_SHIFTFORCES
 209     /* cppcheck-suppress unassignedVariable */
 210     real       shf_array[2*GMX_SIMD_REAL_WIDTH], *shf;
 211 #endif
 212
 213     int ninner;
 214
 215 #ifdef COUNT_PAIRS
 216     int npair = 0;
 217 #endif
 218
 219 #if defined LJ_COMB_GEOM || defined LJ_COMB_LB || defined LJ_EWALD_GEOM
 220     ljc = nbat->lj_comb;
 221 #endif
 222 #if !(defined LJ_COMB_GEOM || defined LJ_COMB_LB)
 223     /* No combination rule used */
 224     nbfp_ptr    = (4 == nbfp_stride) ? nbat->nbfp_s4 : nbat->nbfp;
 225 #endif
 226
 227     /* Load j-i for the first i */
 228     diagonal_jmi_S    = gmx_simd_load_r(nbat->simd_2xnn_diagonal_j_minus_i);
 229     /* Generate all the diagonal masks as comparison results */
 230 #if UNROLLI == UNROLLJ
 231     diagonal_mask_S0  = gmx_simd_cmplt_r(zero_S, diagonal_jmi_S);
 232     diagonal_jmi_S    = gmx_simd_sub_r(diagonal_jmi_S, one_S);
 233     diagonal_jmi_S    = gmx_simd_sub_r(diagonal_jmi_S, one_S);
 234     diagonal_mask_S2  = gmx_simd_cmplt_r(zero_S, diagonal_jmi_S);
 235 #else
 236 #if 2*UNROLLI == UNROLLJ
 237     diagonal_mask0_S0 = gmx_simd_cmplt_r(zero_S, diagonal_jmi_S);
 238     diagonal_jmi_S    = gmx_simd_sub_r(diagonal_jmi_S, one_S);
 239     diagonal_jmi_S    = gmx_simd_sub_r(diagonal_jmi_S, one_S);
 240     diagonal_mask0_S2 = gmx_simd_cmplt_r(zero_S, diagonal_jmi_S);
 241     diagonal_jmi_S    = gmx_simd_sub_r(diagonal_jmi_S, one_S);
 242     diagonal_jmi_S    = gmx_simd_sub_r(diagonal_jmi_S, one_S);
 243     diagonal_mask1_S0 = gmx_simd_cmplt_r(zero_S, diagonal_jmi_S);
 244     diagonal_jmi_S    = gmx_simd_sub_r(diagonal_jmi_S, one_S);
 245     diagonal_jmi_S    = gmx_simd_sub_r(diagonal_jmi_S, one_S);
 246     diagonal_mask1_S2 = gmx_simd_cmplt_r(zero_S, diagonal_jmi_S);
 247 #endif
 248 #endif
 249
 250     /* Load masks for topology exclusion masking. filter_stride is
 251        static const, so the conditional will be optimized away. */
 252     if (1 == filter_stride)
 253     {
 254         exclusion_filter = nbat->simd_exclusion_filter1;
 255     }
 256     else /* (2 == filter_stride) */
 257     {
 258         exclusion_filter = nbat->simd_exclusion_filter2;
 259     }
 260
 261     /* Here we cast the exclusion filters from unsigned * to int * or real *.
 262      * Since we only check bits, the actual value they represent does not
 263      * matter, as long as both filter and mask data are treated the same way.
 264      */
 265     filter_S0 = gmx_load_exclusion_filter(exclusion_filter + 0*2*UNROLLJ*filter_stride);
 266     filter_S2 = gmx_load_exclusion_filter(exclusion_filter + 1*2*UNROLLJ*filter_stride);
 267
 268 #ifdef CALC_COUL_RF
 269     /* Reaction-field constants */
 270     mrc_3_S  = gmx_simd_set1_r(-2*ic->k_rf);
 271 #ifdef CALC_ENERGIES
 272     hrc_3_S  = gmx_simd_set1_r(ic->k_rf);
 273     moh_rc_S = gmx_simd_set1_r(-ic->c_rf);
 274 #endif
 275 #endif
 276
 277 #ifdef CALC_COUL_TAB
 278     /* Generate aligned table index pointers */
 279     ti0 = prepare_table_load_buffer(ti0_array);
 280     ti2 = prepare_table_load_buffer(ti2_array);
 281
 282     invtsp_S  = gmx_simd_set1_r(ic->tabq_scale);
 283 #ifdef CALC_ENERGIES
 284     mhalfsp_S = gmx_simd_set1_r(-0.5/ic->tabq_scale);
 285 #endif
 286
 287 #ifdef TAB_FDV0
 288     tab_coul_F = ic->tabq_coul_FDV0;
 289 #else
 290     tab_coul_F = ic->tabq_coul_F;
 291     tab_coul_V = ic->tabq_coul_V;
 292 #endif
 293 #endif /* CALC_COUL_TAB */
 294
 295 #ifdef CALC_COUL_EWALD
 296     beta2_S = gmx_simd_set1_r(ic->ewaldcoeff_q*ic->ewaldcoeff_q);
 297     beta_S  = gmx_simd_set1_r(ic->ewaldcoeff_q);
 298 #endif
 299
 300 #if (defined CALC_COUL_TAB || defined CALC_COUL_EWALD) && defined CALC_ENERGIES
 301     sh_ewald_S = gmx_simd_set1_r(ic->sh_ewald);
 302 #endif
 303
 304     /* LJ function constants */
 305 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
 306     sixth_S      = gmx_simd_set1_r(1.0/6.0);
 307     twelveth_S   = gmx_simd_set1_r(1.0/12.0);
 308 #endif
 309
 310 #if defined LJ_CUT && defined CALC_ENERGIES
 311     /* We shift the potential by cpot, which can be zero */
 312     p6_cpot_S    = gmx_simd_set1_r(ic->dispersion_shift.cpot);
 313     p12_cpot_S   = gmx_simd_set1_r(ic->repulsion_shift.cpot);
 314 #endif
 315 #ifdef LJ_POT_SWITCH
 316     rswitch_S = gmx_simd_set1_r(ic->rvdw_switch);
 317     swV3_S    = gmx_simd_set1_r(ic->vdw_switch.c3);
 318     swV4_S    = gmx_simd_set1_r(ic->vdw_switch.c4);
 319     swV5_S    = gmx_simd_set1_r(ic->vdw_switch.c5);
 320     swF2_S    = gmx_simd_set1_r(3*ic->vdw_switch.c3);
 321     swF3_S    = gmx_simd_set1_r(4*ic->vdw_switch.c4);
 322     swF4_S    = gmx_simd_set1_r(5*ic->vdw_switch.c5);
 323 #endif
 324 #ifdef LJ_FORCE_SWITCH
 325     rswitch_S = gmx_simd_set1_r(ic->rvdw_switch);
 326     p6_fc2_S  = gmx_simd_set1_r(ic->dispersion_shift.c2);
 327     p6_fc3_S  = gmx_simd_set1_r(ic->dispersion_shift.c3);
 328     p12_fc2_S = gmx_simd_set1_r(ic->repulsion_shift.c2);
 329     p12_fc3_S = gmx_simd_set1_r(ic->repulsion_shift.c3);
 330 #ifdef CALC_ENERGIES
 331     {
 332         gmx_simd_real_t mthird_S  = gmx_simd_set1_r(-1.0/3.0);
 333         gmx_simd_real_t mfourth_S = gmx_simd_set1_r(-1.0/4.0);
 334
 335         p6_vc3_S     = gmx_simd_mul_r(mthird_S,  p6_fc2_S);
 336         p6_vc4_S     = gmx_simd_mul_r(mfourth_S, p6_fc3_S);
 337         p6_6cpot_S   = gmx_simd_set1_r(ic->dispersion_shift.cpot/6);
 338         p12_vc3_S    = gmx_simd_mul_r(mthird_S,  p12_fc2_S);
 339         p12_vc4_S    = gmx_simd_mul_r(mfourth_S, p12_fc3_S);
 340         p12_12cpot_S = gmx_simd_set1_r(ic->repulsion_shift.cpot/12);
 341     }
 342 #endif
 343 #endif
 344 #ifdef LJ_EWALD_GEOM
 345     mone_S           = gmx_simd_set1_r(-1.0);
 346     half_S           = gmx_simd_set1_r(0.5);
 347     lj_ewaldcoeff2   = ic->ewaldcoeff_lj*ic->ewaldcoeff_lj;
 348     lj_ewaldcoeff6_6 = lj_ewaldcoeff2*lj_ewaldcoeff2*lj_ewaldcoeff2/6;
 349     lje_c2_S         = gmx_simd_set1_r(lj_ewaldcoeff2);
 350     lje_c6_6_S       = gmx_simd_set1_r(lj_ewaldcoeff6_6);
 351     /* Determine the grid potential at the cut-off */
 352     lje_vc_S         = gmx_simd_set1_r(ic->sh_lj_ewald);
 353 #endif
 354
 355     /* The kernel either supports rcoulomb = rvdw or rcoulomb >= rvdw */
 356     rc2_S    = gmx_simd_set1_r(ic->rcoulomb*ic->rcoulomb);
 357 #ifdef VDW_CUTOFF_CHECK
 358     rcvdw2_S = gmx_simd_set1_r(ic->rvdw*ic->rvdw);
 359 #endif
 360
 361     avoid_sing_S = gmx_simd_set1_r(NBNXN_AVOID_SING_R2_INC);
 362
 363     q                   = nbat->q;
 364     type                = nbat->type;
 365     facel               = ic->epsfac;
 366     shiftvec            = shift_vec[0];
 367     x                   = nbat->x;
 368
 369 #ifdef CALC_ENERGIES
 370     tmpsum   = gmx_simd_align_r(tmpsum_array);
 371 #endif
 372 #ifdef CALC_SHIFTFORCES
 373     shf      = gmx_simd_align_r(shf_array);
 374 #endif
 375
 376 #ifdef FIX_LJ_C
 377     pvdw_c6  = gmx_simd_align_r(pvdw_array);
 378     pvdw_c12 = pvdw_c6 + UNROLLI*UNROLLJ;
 379
 380     for (jp = 0; jp < UNROLLJ; jp++)
 381     {
 382         pvdw_c6 [0*UNROLLJ+jp] = nbat->nbfp[0*2];
 383         pvdw_c6 [1*UNROLLJ+jp] = nbat->nbfp[0*2];
 384         pvdw_c6 [2*UNROLLJ+jp] = nbat->nbfp[0*2];
 385         pvdw_c6 [3*UNROLLJ+jp] = nbat->nbfp[0*2];
 386
 387         pvdw_c12[0*UNROLLJ+jp] = nbat->nbfp[0*2+1];
 388         pvdw_c12[1*UNROLLJ+jp] = nbat->nbfp[0*2+1];
 389         pvdw_c12[2*UNROLLJ+jp] = nbat->nbfp[0*2+1];
 390         pvdw_c12[3*UNROLLJ+jp] = nbat->nbfp[0*2+1];
 391     }
 392     c6_S0            = gmx_simd_load_r(pvdw_c6 +0*UNROLLJ);
 393     c6_S1            = gmx_simd_load_r(pvdw_c6 +1*UNROLLJ);
 394     c6_S2            = gmx_simd_load_r(pvdw_c6 +2*UNROLLJ);
 395     c6_S3            = gmx_simd_load_r(pvdw_c6 +3*UNROLLJ);
 396
 397     c12_S0           = gmx_simd_load_r(pvdw_c12+0*UNROLLJ);
 398     c12_S1           = gmx_simd_load_r(pvdw_c12+1*UNROLLJ);
 399     c12_S2           = gmx_simd_load_r(pvdw_c12+2*UNROLLJ);
 400     c12_S3           = gmx_simd_load_r(pvdw_c12+3*UNROLLJ);
 401 #endif /* FIX_LJ_C */
 402
 403 #ifdef ENERGY_GROUPS
 404     egps_ishift  = nbat->neg_2log;
 405     egps_imask   = (1<<egps_ishift) - 1;
 406     egps_jshift  = 2*nbat->neg_2log;
 407     egps_jmask   = (1<<egps_jshift) - 1;
 408     egps_jstride = (UNROLLJ>>1)*UNROLLJ;
 409     /* Major division is over i-particle energy groups, determine the stride */
 410     Vstride_i    = nbat->nenergrp*(1<<nbat->neg_2log)*egps_jstride;
 411 #endif
 412
 413     l_cj = nbl->cj;
 414
 415     ninner = 0;
 416     for (n = 0; n < nbl->nci; n++)
 417     {
 418         nbln = &nbl->ci[n];
 419
 420         ish              = (nbln->shift & NBNXN_CI_SHIFT);
 421         ish3             = ish*3;
 422         cjind0           = nbln->cj_ind_start;
 423         cjind1           = nbln->cj_ind_end;
 424         ci               = nbln->ci;
 425         ci_sh            = (ish == CENTRAL ? ci : -1);
 426
 427         shX_S = gmx_simd_load1_r(shiftvec+ish3);
 428         shY_S = gmx_simd_load1_r(shiftvec+ish3+1);
 429         shZ_S = gmx_simd_load1_r(shiftvec+ish3+2);
 430
 431 #if UNROLLJ <= 4
 432         sci              = ci*STRIDE;
 433         scix             = sci*DIM;
 434         sci2             = sci*2;
 435 #else
 436         sci              = (ci>>1)*STRIDE;
 437         scix             = sci*DIM + (ci & 1)*(STRIDE>>1);
 438         sci2             = sci*2 + (ci & 1)*(STRIDE>>1);
 439         sci             += (ci & 1)*(STRIDE>>1);
 440 #endif
 441
 442         /* We have 5 LJ/C combinations, but use only three inner loops,
 443          * as the other combinations are unlikely and/or not much faster:
 444          * inner half-LJ + C for half-LJ + C / no-LJ + C
 445          * inner LJ + C      for full-LJ + C
 446          * inner LJ          for full-LJ + no-C / half-LJ + no-C
 447          */
 448         do_LJ   = (nbln->shift & NBNXN_CI_DO_LJ(0));
 449         do_coul = (nbln->shift & NBNXN_CI_DO_COUL(0));
 450         half_LJ = ((nbln->shift & NBNXN_CI_HALF_LJ(0)) || !do_LJ) && do_coul;
 451 #ifdef LJ_EWALD_GEOM
 452         do_self = TRUE;
 453 #else
 454         do_self = do_coul;
 455 #endif
 456
 457 #ifdef ENERGY_GROUPS
 458         egps_i = nbat->energrp[ci];
 459         {
 460             int ia, egp_ia;
 461
 462             for (ia = 0; ia < UNROLLI; ia++)
 463             {
 464                 egp_ia     = (egps_i >> (ia*egps_ishift)) & egps_imask;
 465                 vvdwtp[ia] = Vvdw + egp_ia*Vstride_i;
 466                 vctp[ia]   = Vc   + egp_ia*Vstride_i;
 467             }
 468         }
 469 #endif
 470
 471 #ifdef CALC_ENERGIES
 472 #if UNROLLJ == 4
 473         if (do_self && l_cj[nbln->cj_ind_start].cj == ci_sh)
 474 #endif
 475 #if UNROLLJ == 8
 476         if (do_self && l_cj[nbln->cj_ind_start].cj == (ci_sh>>1))
 477 #endif
 478         {
 479             if (do_coul)
 480             {
 481                 real Vc_sub_self;
 482                 int  ia;
 483
 484 #ifdef CALC_COUL_RF
 485                 Vc_sub_self = 0.5*ic->c_rf;
 486 #endif
 487 #ifdef CALC_COUL_TAB
 488 #ifdef TAB_FDV0
 489                 Vc_sub_self = 0.5*tab_coul_F[2];
 490 #else
 491                 Vc_sub_self = 0.5*tab_coul_V[0];
 492 #endif
 493 #endif
 494 #ifdef CALC_COUL_EWALD
 495                 /* beta/sqrt(pi) */
 496                 Vc_sub_self = 0.5*ic->ewaldcoeff_q*M_2_SQRTPI;
 497 #endif
 498
 499                 for (ia = 0; ia < UNROLLI; ia++)
 500                 {
 501                     real qi;
 502
 503                     qi = q[sci+ia];
 504 #ifdef ENERGY_GROUPS
 505                     vctp[ia][((egps_i>>(ia*egps_ishift)) & egps_imask)*egps_jstride]
 506 #else
 507                     Vc[0]
 508 #endif
 509                         -= facel*qi*qi*Vc_sub_self;
 510                 }
 511             }
 512
 513 #ifdef LJ_EWALD_GEOM
 514             {
 515                 int  ia;
 516
 517                 for (ia = 0; ia < UNROLLI; ia++)
 518                 {
 519                     real c6_i;
 520
 521                     c6_i = nbat->nbfp[nbat->type[sci+ia]*(nbat->ntype + 1)*2]/6;
 522 #ifdef ENERGY_GROUPS
 523                     vvdwtp[ia][((egps_i>>(ia*egps_ishift)) & egps_imask)*egps_jstride]
 524 #else
 525                     Vvdw[0]
 526 #endif
 527                         += 0.5*c6_i*lj_ewaldcoeff6_6;
 528                 }
 529             }
 530 #endif      /* LJ_EWALD */
 531         }
 532 #endif
 533
 534         /* Load i atom data */
 535         sciy             = scix + STRIDE;
 536         sciz             = sciy + STRIDE;
 537         gmx_load1p1_pr(&ix_S0, x+scix);
 538         gmx_load1p1_pr(&ix_S2, x+scix+2);
 539         gmx_load1p1_pr(&iy_S0, x+sciy);
 540         gmx_load1p1_pr(&iy_S2, x+sciy+2);
 541         gmx_load1p1_pr(&iz_S0, x+sciz);
 542         gmx_load1p1_pr(&iz_S2, x+sciz+2);
 543         ix_S0          = gmx_simd_add_r(ix_S0, shX_S);
 544         ix_S2          = gmx_simd_add_r(ix_S2, shX_S);
 545         iy_S0          = gmx_simd_add_r(iy_S0, shY_S);
 546         iy_S2          = gmx_simd_add_r(iy_S2, shY_S);
 547         iz_S0          = gmx_simd_add_r(iz_S0, shZ_S);
 548         iz_S2          = gmx_simd_add_r(iz_S2, shZ_S);
 549
 550         if (do_coul)
 551         {
 552             gmx_simd_real_t facel_S;
 553
 554             facel_S    = gmx_simd_set1_r(facel);
 555
 556             gmx_load1p1_pr(&iq_S0, q+sci);
 557             gmx_load1p1_pr(&iq_S2, q+sci+2);
 558             iq_S0      = gmx_simd_mul_r(facel_S, iq_S0);
 559             iq_S2      = gmx_simd_mul_r(facel_S, iq_S2);
 560         }
 561
 562 #ifdef LJ_COMB_LB
 563         gmx_load1p1_pr(&hsig_i_S0, ljc+sci2+0);
 564         gmx_load1p1_pr(&hsig_i_S2, ljc+sci2+2);
 565         gmx_load1p1_pr(&seps_i_S0, ljc+sci2+STRIDE+0);
 566         gmx_load1p1_pr(&seps_i_S2, ljc+sci2+STRIDE+2);
 567 #else
 568 #ifdef LJ_COMB_GEOM
 569         gmx_load1p1_pr(&c6s_S0, ljc+sci2+0);
 570         if (!half_LJ)
 571         {
 572             gmx_load1p1_pr(&c6s_S2, ljc+sci2+2);
 573         }
 574         gmx_load1p1_pr(&c12s_S0, ljc+sci2+STRIDE+0);
 575         if (!half_LJ)
 576         {
 577             gmx_load1p1_pr(&c12s_S2, ljc+sci2+STRIDE+2);
 578         }
 579 #else
 580         nbfp0     = nbfp_ptr + type[sci  ]*nbat->ntype*nbfp_stride;
 581         nbfp1     = nbfp_ptr + type[sci+1]*nbat->ntype*nbfp_stride;
 582         if (!half_LJ)
 583         {
 584             nbfp2 = nbfp_ptr + type[sci+2]*nbat->ntype*nbfp_stride;
 585             nbfp3 = nbfp_ptr + type[sci+3]*nbat->ntype*nbfp_stride;
 586         }
 587 #endif
 588 #endif
 589 #ifdef LJ_EWALD_GEOM
 590         /* We need the geometrically combined C6 for the PME grid correction */
 591         gmx_load1p1_pr(&c6s_S0, ljc+sci2+0);
 592         if (!half_LJ)
 593         {
 594             gmx_load1p1_pr(&c6s_S2, ljc+sci2+2);
 595         }
 596 #endif
 597
 598         /* Zero the potential energy for this list */
 599         Vvdwtot_S        = gmx_simd_setzero_r();
 600         vctot_S          = gmx_simd_setzero_r();
 601
 602         /* Clear i atom forces */
 603         fix_S0           = gmx_simd_setzero_r();
 604         fix_S2           = gmx_simd_setzero_r();
 605         fiy_S0           = gmx_simd_setzero_r();
 606         fiy_S2           = gmx_simd_setzero_r();
 607         fiz_S0           = gmx_simd_setzero_r();
 608         fiz_S2           = gmx_simd_setzero_r();
 609
 610         cjind = cjind0;
 611
 612         /* Currently all kernels use (at least half) LJ */
 613 #define CALC_LJ
 614         if (half_LJ)
 615         {
 616             /* Coulomb: all i-atoms, LJ: first half i-atoms */
 617 #define CALC_COULOMB
 618 #define HALF_LJ
 619 #define CHECK_EXCLS
 620             while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
 621             {
 622 #include "nbnxn_kernel_simd_2xnn_inner.h"
 623                 cjind++;
 624             }
 625 #undef CHECK_EXCLS
 626             for (; (cjind < cjind1); cjind++)
 627             {
 628 #include "nbnxn_kernel_simd_2xnn_inner.h"
 629             }
 630 #undef HALF_LJ
 631 #undef CALC_COULOMB
 632         }
 633         else if (do_coul)
 634         {
 635             /* Coulomb: all i-atoms, LJ: all i-atoms */
 636 #define CALC_COULOMB
 637 #define CHECK_EXCLS
 638             while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
 639             {
 640 #include "nbnxn_kernel_simd_2xnn_inner.h"
 641                 cjind++;
 642             }
 643 #undef CHECK_EXCLS
 644             for (; (cjind < cjind1); cjind++)
 645             {
 646 #include "nbnxn_kernel_simd_2xnn_inner.h"
 647             }
 648 #undef CALC_COULOMB
 649         }
 650         else
 651         {
 652             /* Coulomb: none, LJ: all i-atoms */
 653 #define CHECK_EXCLS
 654             while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
 655             {
 656 #include "nbnxn_kernel_simd_2xnn_inner.h"
 657                 cjind++;
 658             }
 659 #undef CHECK_EXCLS
 660             for (; (cjind < cjind1); cjind++)
 661             {
 662 #include "nbnxn_kernel_simd_2xnn_inner.h"
 663             }
 664         }
 665 #undef CALC_LJ
 666         ninner += cjind1 - cjind0;
 667
 668         /* Add accumulated i-forces to the force array */
 669         fix_S = gmx_mm_transpose_sum4h_pr(fix_S0, fix_S2);
 670         gmx_simd4_store_r(f+scix, gmx_add_pr4(fix_S, gmx_simd4_load_r(f+scix)));
 671
 672         fiy_S = gmx_mm_transpose_sum4h_pr(fiy_S0, fiy_S2);
 673         gmx_simd4_store_r(f+sciy, gmx_add_pr4(fiy_S, gmx_simd4_load_r(f+sciy)));
 674
 675         fiz_S = gmx_mm_transpose_sum4h_pr(fiz_S0, fiz_S2);
 676         gmx_simd4_store_r(f+sciz, gmx_add_pr4(fiz_S, gmx_simd4_load_r(f+sciz)));
 677
 678 #ifdef CALC_SHIFTFORCES
 679         fshift[ish3+0] += gmx_sum_simd4(fix_S, shf);
 680         fshift[ish3+1] += gmx_sum_simd4(fiy_S, shf);
 681         fshift[ish3+2] += gmx_sum_simd4(fiz_S, shf);
 682 #endif
 683
 684 #ifdef CALC_ENERGIES
 685         if (do_coul)
 686         {
 687             *Vc += gmx_sum_simd(vctot_S, tmpsum);
 688         }
 689
 690         *Vvdw += gmx_sum_simd(Vvdwtot_S, tmpsum);
 691 #endif
 692
 693         /* Outer loop uses 6 flops/iteration */
 694     }
 695
 696 #ifdef COUNT_PAIRS
 697     printf("atom pairs %d\n", npair);
 698 #endif
 699 }