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