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