src/gromacs/mdlib/nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn.c

   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  * Note: this file was generated by the Verlet kernel generator for
  37  * kernel type 4xn.
  38  */
  39
  40 #ifdef HAVE_CONFIG_H
  41 #include <config.h>
  42 #endif
  43
  44 #include "typedefs.h"
  45
  46 #include "gromacs/mdlib/nbnxn_simd.h"
  47
  48 #ifdef GMX_NBNXN_SIMD_4XN
  49
  50 #include "gromacs/simd/vector_operations.h"
  51
  52 #if !(GMX_SIMD_REAL_WIDTH == 2 || GMX_SIMD_REAL_WIDTH == 4 || GMX_SIMD_REAL_WIDTH == 8)
  53 #error "unsupported SIMD width"
  54 #endif
  55
  56 #define GMX_SIMD_J_UNROLL_SIZE 1
  57 #include "nbnxn_kernel_simd_4xn.h"
  58 #include "../nbnxn_kernel_common.h"
  59 #include "gmx_omp_nthreads.h"
  60 #include "types/force_flags.h"
  61 #include "gmx_fatal.h"
  62
  63 /*! \brief Kinds of electrostatic treatments in SIMD Verlet kernels
  64  */
  65 enum {
  66     coulktRF, coulktTAB, coulktTAB_TWIN, coulktEWALD, coulktEWALD_TWIN, coulktNR
  67 };
  68
  69 /*! \brief Kinds of Van der Waals treatments in SIMD Verlet kernels
  70  */
  71 enum {
  72     vdwktLJCUT_COMBGEOM, vdwktLJCUT_COMBLB, vdwktLJCUT_COMBNONE, vdwktLJFORCESWITCH, vdwktLJPOTSWITCH, vdwktLJEWALDCOMBGEOM, vdwktNR
  73 };
  74
  75 /* Declare and define the kernel function pointer lookup tables.
  76  * The minor index of the array goes over both the LJ combination rules,
  77  * which is only supported by plain cut-off, and the LJ switch/PME functions.
  78  */
  79 static p_nbk_func_noener p_nbk_noener[coulktNR][vdwktNR] =
  80 {
  81     {
  82         nbnxn_kernel_ElecRF_VdwLJCombGeom_F_4xn,
  83         nbnxn_kernel_ElecRF_VdwLJCombLB_F_4xn,
  84         nbnxn_kernel_ElecRF_VdwLJ_F_4xn,
  85         nbnxn_kernel_ElecRF_VdwLJFSw_F_4xn,
  86         nbnxn_kernel_ElecRF_VdwLJPSw_F_4xn,
  87         nbnxn_kernel_ElecRF_VdwLJEwCombGeom_F_4xn,
  88     },
  89     {
  90         nbnxn_kernel_ElecQSTab_VdwLJCombGeom_F_4xn,
  91         nbnxn_kernel_ElecQSTab_VdwLJCombLB_F_4xn,
  92         nbnxn_kernel_ElecQSTab_VdwLJ_F_4xn,
  93         nbnxn_kernel_ElecQSTab_VdwLJFSw_F_4xn,
  94         nbnxn_kernel_ElecQSTab_VdwLJPSw_F_4xn,
  95         nbnxn_kernel_ElecQSTab_VdwLJEwCombGeom_F_4xn,
  96     },
  97     {
  98         nbnxn_kernel_ElecQSTabTwinCut_VdwLJCombGeom_F_4xn,
  99         nbnxn_kernel_ElecQSTabTwinCut_VdwLJCombLB_F_4xn,
 100         nbnxn_kernel_ElecQSTabTwinCut_VdwLJ_F_4xn,
 101         nbnxn_kernel_ElecQSTabTwinCut_VdwLJFSw_F_4xn,
 102         nbnxn_kernel_ElecQSTabTwinCut_VdwLJPSw_F_4xn,
 103         nbnxn_kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_F_4xn,
 104     },
 105     {
 106         nbnxn_kernel_ElecEw_VdwLJCombGeom_F_4xn,
 107         nbnxn_kernel_ElecEw_VdwLJCombLB_F_4xn,
 108         nbnxn_kernel_ElecEw_VdwLJ_F_4xn,
 109         nbnxn_kernel_ElecEw_VdwLJFSw_F_4xn,
 110         nbnxn_kernel_ElecEw_VdwLJPSw_F_4xn,
 111         nbnxn_kernel_ElecEw_VdwLJEwCombGeom_F_4xn,
 112     },
 113     {
 114         nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_F_4xn,
 115         nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_F_4xn,
 116         nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_4xn,
 117         nbnxn_kernel_ElecEwTwinCut_VdwLJFSw_F_4xn,
 118         nbnxn_kernel_ElecEwTwinCut_VdwLJPSw_F_4xn,
 119         nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_4xn,
 120     },
 121 };
 122
 123 static p_nbk_func_ener p_nbk_ener[coulktNR][vdwktNR] =
 124 {
 125     {
 126         nbnxn_kernel_ElecRF_VdwLJCombGeom_VF_4xn,
 127         nbnxn_kernel_ElecRF_VdwLJCombLB_VF_4xn,
 128         nbnxn_kernel_ElecRF_VdwLJ_VF_4xn,
 129         nbnxn_kernel_ElecRF_VdwLJFSw_VF_4xn,
 130         nbnxn_kernel_ElecRF_VdwLJPSw_VF_4xn,
 131         nbnxn_kernel_ElecRF_VdwLJEwCombGeom_VF_4xn,
 132     },
 133     {
 134         nbnxn_kernel_ElecQSTab_VdwLJCombGeom_VF_4xn,
 135         nbnxn_kernel_ElecQSTab_VdwLJCombLB_VF_4xn,
 136         nbnxn_kernel_ElecQSTab_VdwLJ_VF_4xn,
 137         nbnxn_kernel_ElecQSTab_VdwLJFSw_VF_4xn,
 138         nbnxn_kernel_ElecQSTab_VdwLJPSw_VF_4xn,
 139         nbnxn_kernel_ElecQSTab_VdwLJEwCombGeom_VF_4xn,
 140     },
 141     {
 142         nbnxn_kernel_ElecQSTabTwinCut_VdwLJCombGeom_VF_4xn,
 143         nbnxn_kernel_ElecQSTabTwinCut_VdwLJCombLB_VF_4xn,
 144         nbnxn_kernel_ElecQSTabTwinCut_VdwLJ_VF_4xn,
 145         nbnxn_kernel_ElecQSTabTwinCut_VdwLJFSw_VF_4xn,
 146         nbnxn_kernel_ElecQSTabTwinCut_VdwLJPSw_VF_4xn,
 147         nbnxn_kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VF_4xn,
 148     },
 149     {
 150         nbnxn_kernel_ElecEw_VdwLJCombGeom_VF_4xn,
 151         nbnxn_kernel_ElecEw_VdwLJCombLB_VF_4xn,
 152         nbnxn_kernel_ElecEw_VdwLJ_VF_4xn,
 153         nbnxn_kernel_ElecEw_VdwLJFSw_VF_4xn,
 154         nbnxn_kernel_ElecEw_VdwLJPSw_VF_4xn,
 155         nbnxn_kernel_ElecEw_VdwLJEwCombGeom_VF_4xn,
 156     },
 157     {
 158         nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_VF_4xn,
 159         nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_VF_4xn,
 160         nbnxn_kernel_ElecEwTwinCut_VdwLJ_VF_4xn,
 161         nbnxn_kernel_ElecEwTwinCut_VdwLJFSw_VF_4xn,
 162         nbnxn_kernel_ElecEwTwinCut_VdwLJPSw_VF_4xn,
 163         nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_4xn,
 164     },
 165 };
 166
 167 static p_nbk_func_ener p_nbk_energrp[coulktNR][vdwktNR] =
 168 {
 169     {
 170         nbnxn_kernel_ElecRF_VdwLJCombGeom_VgrpF_4xn,
 171         nbnxn_kernel_ElecRF_VdwLJCombLB_VgrpF_4xn,
 172         nbnxn_kernel_ElecRF_VdwLJ_VgrpF_4xn,
 173         nbnxn_kernel_ElecRF_VdwLJFSw_VgrpF_4xn,
 174         nbnxn_kernel_ElecRF_VdwLJPSw_VgrpF_4xn,
 175         nbnxn_kernel_ElecRF_VdwLJEwCombGeom_VgrpF_4xn,
 176     },
 177     {
 178         nbnxn_kernel_ElecQSTab_VdwLJCombGeom_VgrpF_4xn,
 179         nbnxn_kernel_ElecQSTab_VdwLJCombLB_VgrpF_4xn,
 180         nbnxn_kernel_ElecQSTab_VdwLJ_VgrpF_4xn,
 181         nbnxn_kernel_ElecQSTab_VdwLJFSw_VgrpF_4xn,
 182         nbnxn_kernel_ElecQSTab_VdwLJPSw_VgrpF_4xn,
 183         nbnxn_kernel_ElecQSTab_VdwLJEwCombGeom_VgrpF_4xn,
 184     },
 185     {
 186         nbnxn_kernel_ElecQSTabTwinCut_VdwLJCombGeom_VgrpF_4xn,
 187         nbnxn_kernel_ElecQSTabTwinCut_VdwLJCombLB_VgrpF_4xn,
 188         nbnxn_kernel_ElecQSTabTwinCut_VdwLJ_VgrpF_4xn,
 189         nbnxn_kernel_ElecQSTabTwinCut_VdwLJFSw_VgrpF_4xn,
 190         nbnxn_kernel_ElecQSTabTwinCut_VdwLJPSw_VgrpF_4xn,
 191         nbnxn_kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VgrpF_4xn,
 192     },
 193     {
 194         nbnxn_kernel_ElecEw_VdwLJCombGeom_VgrpF_4xn,
 195         nbnxn_kernel_ElecEw_VdwLJCombLB_VgrpF_4xn,
 196         nbnxn_kernel_ElecEw_VdwLJ_VgrpF_4xn,
 197         nbnxn_kernel_ElecEw_VdwLJFSw_VgrpF_4xn,
 198         nbnxn_kernel_ElecEw_VdwLJPSw_VgrpF_4xn,
 199         nbnxn_kernel_ElecEw_VdwLJEwCombGeom_VgrpF_4xn,
 200     },
 201     {
 202         nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_VgrpF_4xn,
 203         nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_VgrpF_4xn,
 204         nbnxn_kernel_ElecEwTwinCut_VdwLJ_VgrpF_4xn,
 205         nbnxn_kernel_ElecEwTwinCut_VdwLJFSw_VgrpF_4xn,
 206         nbnxn_kernel_ElecEwTwinCut_VdwLJPSw_VgrpF_4xn,
 207         nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VgrpF_4xn,
 208     },
 209 };
 210
 211
 212 static void
 213 reduce_group_energies(int ng, int ng_2log,
 214                       const real *VSvdw, const real *VSc,
 215                       real *Vvdw, real *Vc)
 216 {
 217     const int unrollj      = GMX_SIMD_REAL_WIDTH/GMX_SIMD_J_UNROLL_SIZE;
 218     const int unrollj_half = unrollj/2;
 219     int       ng_p2, i, j, j0, j1, c, s;
 220
 221     ng_p2 = (1<<ng_2log);
 222
 223     /* The size of the x86 SIMD energy group buffer array is:
 224      * ng*ng*ng_p2*unrollj_half*simd_width
 225      */
 226     for (i = 0; i < ng; i++)
 227     {
 228         for (j = 0; j < ng; j++)
 229         {
 230             Vvdw[i*ng+j] = 0;
 231             Vc[i*ng+j]   = 0;
 232         }
 233
 234         for (j1 = 0; j1 < ng; j1++)
 235         {
 236             for (j0 = 0; j0 < ng; j0++)
 237             {
 238                 c = ((i*ng + j1)*ng_p2 + j0)*unrollj_half*unrollj;
 239                 for (s = 0; s < unrollj_half; s++)
 240                 {
 241                     Vvdw[i*ng+j0] += VSvdw[c+0];
 242                     Vvdw[i*ng+j1] += VSvdw[c+1];
 243                     Vc  [i*ng+j0] += VSc  [c+0];
 244                     Vc  [i*ng+j1] += VSc  [c+1];
 245                     c             += unrollj + 2;
 246                 }
 247             }
 248         }
 249     }
 250 }
 251
 252 #else /* GMX_NBNXN_SIMD_4XN */
 253
 254 #include "gmx_fatal.h"
 255
 256 #endif /* GMX_NBNXN_SIMD_4XN */
 257
 258 void
 259 nbnxn_kernel_simd_4xn(nbnxn_pairlist_set_t      gmx_unused *nbl_list,
 260                       const nbnxn_atomdata_t    gmx_unused *nbat,
 261                       const interaction_const_t gmx_unused *ic,
 262                       int                       gmx_unused  ewald_excl,
 263                       rvec                      gmx_unused *shift_vec,
 264                       int                       gmx_unused  force_flags,
 265                       int                       gmx_unused  clearF,
 266                       real                      gmx_unused *fshift,
 267                       real                      gmx_unused *Vc,
 268                       real                      gmx_unused *Vvdw)
 269 #ifdef GMX_NBNXN_SIMD_4XN
 270 {
 271     int                nnbl;
 272     nbnxn_pairlist_t **nbl;
 273     int                coulkt, vdwkt = 0;
 274     int                nb;
 275     int                nthreads gmx_unused;
 276
 277     nnbl = nbl_list->nnbl;
 278     nbl  = nbl_list->nbl;
 279
 280     if (EEL_RF(ic->eeltype) || ic->eeltype == eelCUT)
 281     {
 282         coulkt = coulktRF;
 283     }
 284     else
 285     {
 286         if (ewald_excl == ewaldexclTable)
 287         {
 288             if (ic->rcoulomb == ic->rvdw)
 289             {
 290                 coulkt = coulktTAB;
 291             }
 292             else
 293             {
 294                 coulkt = coulktTAB_TWIN;
 295             }
 296         }
 297         else
 298         {
 299             if (ic->rcoulomb == ic->rvdw)
 300             {
 301                 coulkt = coulktEWALD;
 302             }
 303             else
 304             {
 305                 coulkt = coulktEWALD_TWIN;
 306             }
 307         }
 308     }
 309
 310     if (ic->vdwtype == evdwCUT)
 311     {
 312         switch (ic->vdw_modifier)
 313         {
 314             case eintmodNONE:
 315             case eintmodPOTSHIFT:
 316                 switch (nbat->comb_rule)
 317                 {
 318                     case ljcrGEOM: vdwkt = vdwktLJCUT_COMBGEOM; break;
 319                     case ljcrLB:   vdwkt = vdwktLJCUT_COMBLB;   break;
 320                     case ljcrNONE: vdwkt = vdwktLJCUT_COMBNONE; break;
 321                     default:       gmx_incons("Unknown combination rule");
 322                 }
 323                 break;
 324             case eintmodFORCESWITCH:
 325                 vdwkt = vdwktLJFORCESWITCH;
 326                 break;
 327             case eintmodPOTSWITCH:
 328                 vdwkt = vdwktLJPOTSWITCH;
 329                 break;
 330             default:
 331                 gmx_incons("Unsupported VdW interaction modifier");
 332         }
 333     }
 334     else if (ic->vdwtype == evdwPME)
 335     {
 336         if (ic->ljpme_comb_rule == eljpmeLB)
 337         {
 338             gmx_incons("The nbnxn SIMD kernels don't suport LJ-PME with LB");
 339         }
 340         vdwkt = vdwktLJEWALDCOMBGEOM;
 341     }
 342     else
 343     {
 344         gmx_incons("Unsupported VdW interaction type");
 345     }
 346
 347     nthreads = gmx_omp_nthreads_get(emntNonbonded);
 348 #pragma omp parallel for schedule(static) num_threads(nthreads)
 349     for (nb = 0; nb < nnbl; nb++)
 350     {
 351         nbnxn_atomdata_output_t *out;
 352         real                    *fshift_p;
 353
 354         out = &nbat->out[nb];
 355
 356         if (clearF == enbvClearFYes)
 357         {
 358             clear_f(nbat, nb, out->f);
 359         }
 360
 361         if ((force_flags & GMX_FORCE_VIRIAL) && nnbl == 1)
 362         {
 363             fshift_p = fshift;
 364         }
 365         else
 366         {
 367             fshift_p = out->fshift;
 368
 369             if (clearF == enbvClearFYes)
 370             {
 371                 clear_fshift(fshift_p);
 372             }
 373         }
 374
 375         if (!(force_flags & GMX_FORCE_ENERGY))
 376         {
 377             /* Don't calculate energies */
 378             p_nbk_noener[coulkt][vdwkt](nbl[nb], nbat,
 379                                         ic,
 380                                         shift_vec,
 381                                         out->f,
 382                                         fshift_p);
 383         }
 384         else if (out->nV == 1)
 385         {
 386             /* No energy groups */
 387             out->Vvdw[0] = 0;
 388             out->Vc[0]   = 0;
 389
 390             p_nbk_ener[coulkt][vdwkt](nbl[nb], nbat,
 391                                       ic,
 392                                       shift_vec,
 393                                       out->f,
 394                                       fshift_p,
 395                                       out->Vvdw,
 396                                       out->Vc);
 397         }
 398         else
 399         {
 400             /* Calculate energy group contributions */
 401             int i;
 402
 403             for (i = 0; i < out->nVS; i++)
 404             {
 405                 out->VSvdw[i] = 0;
 406             }
 407             for (i = 0; i < out->nVS; i++)
 408             {
 409                 out->VSc[i] = 0;
 410             }
 411
 412             p_nbk_energrp[coulkt][vdwkt](nbl[nb], nbat,
 413                                          ic,
 414                                          shift_vec,
 415                                          out->f,
 416                                          fshift_p,
 417                                          out->VSvdw,
 418                                          out->VSc);
 419
 420             reduce_group_energies(nbat->nenergrp, nbat->neg_2log,
 421                                   out->VSvdw, out->VSc,
 422                                   out->Vvdw, out->Vc);
 423         }
 424     }
 425
 426     if (force_flags & GMX_FORCE_ENERGY)
 427     {
 428         reduce_energies_over_lists(nbat, nnbl, Vvdw, Vc);
 429     }
 430 }
 431 #else
 432 {
 433     gmx_incons("nbnxn_kernel_simd_4xn called when such kernels "
 434                " are not enabled.");
 435 }
 436 #endif
 437 #undef GMX_SIMD_J_UNROLL_SIZE