src/gromacs/nbnxm/kerneldispatch.cpp

   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 #include "gmxpre.h"
  37
  38 #include "gromacs/gmxlib/nrnb.h"
  39 #include "gromacs/math/vectypes.h"
  40 #include "gromacs/mdlib/force_flags.h"
  41 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  42 #include "gromacs/mdtypes/enerdata.h"
  43 #include "gromacs/mdtypes/interaction_const.h"
  44 #include "gromacs/mdtypes/md_enums.h"
  45 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  46 #include "gromacs/nbnxm/nbnxm.h"
  47 #include "gromacs/nbnxm/nbnxm_simd.h"
  48 #include "gromacs/nbnxm/kernels_reference/kernel_gpu_ref.h"
  49 #include "gromacs/simd/simd.h"
  50 #include "gromacs/utility/gmxassert.h"
  51 #include "gromacs/utility/real.h"
  52
  53 #include "kernel_common.h"
  54 #define INCLUDE_KERNELFUNCTION_TABLES
  55 #include "gromacs/nbnxm/kernels_reference/kernel_ref.h"
  56 #ifdef GMX_NBNXN_SIMD_2XNN
  57 #include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
  58 #endif
  59 #ifdef GMX_NBNXN_SIMD_4XN
  60 #include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
  61 #endif
  62 #undef INCLUDE_FUNCTION_TABLES
  63
  64 /*! \brief Clears the energy group output buffers
  65  *
  66  * \param[in,out] out  nbnxn kernel output struct
  67  */
  68 static void clearGroupEnergies(nbnxn_atomdata_output_t *out)
  69 {
  70     std::fill(out->Vvdw.begin(), out->Vvdw.end(), 0.0_real);
  71     std::fill(out->Vc.begin(), out->Vc.end(), 0.0_real);
  72     std::fill(out->VSvdw.begin(), out->VSvdw.end(), 0.0_real);
  73     std::fill(out->VSc.begin(), out->VSc.end(), 0.0_real);
  74 }
  75
  76 /*! \brief Reduce the group-pair energy buffers produced by a SIMD kernel
  77  * to single terms in the output buffers.
  78  *
  79  * The SIMD kernels produce a large number of energy buffer in SIMD registers
  80  * to avoid scattered reads and writes.
  81  *
  82  * \tparam        unrollj         The unroll size for j-particles in the SIMD kernel
  83  * \param[in]     numGroups       The number of energy groups
  84  * \param[in]     numGroups_2log  Log2 of numGroups, rounded up
  85  * \param[in,out] out             Struct with energy buffers
  86  */
  87 template <int unrollj> static void
  88 reduceGroupEnergySimdBuffers(int                       numGroups,
  89                              int                       numGroups_2log,
  90                              nbnxn_atomdata_output_t  *out)
  91 {
  92     const int                 unrollj_half     = unrollj/2;
  93     /* Energies are stored in SIMD registers with size 2^numGroups_2log */
  94     const int                 numGroupsStorage = (1 << numGroups_2log);
  95
  96     const real * gmx_restrict vVdwSimd     = out->VSvdw.data();
  97     const real * gmx_restrict vCoulombSimd = out->VSc.data();
  98     real * gmx_restrict       vVdw         = out->Vvdw.data();
  99     real * gmx_restrict       vCoulomb     = out->Vc.data();
 100
 101     /* The size of the SIMD energy group buffer array is:
 102      * numGroups*numGroups*numGroupsStorage*unrollj_half*simd_width
 103      */
 104     for (int i = 0; i < numGroups; i++)
 105     {
 106         for (int j1 = 0; j1 < numGroups; j1++)
 107         {
 108             for (int j0 = 0; j0 < numGroups; j0++)
 109             {
 110                 int c = ((i*numGroups + j1)*numGroupsStorage + j0)*unrollj_half*unrollj;
 111                 for (int s = 0; s < unrollj_half; s++)
 112                 {
 113                     vVdw    [i*numGroups + j0] += vVdwSimd    [c + 0];
 114                     vVdw    [i*numGroups + j1] += vVdwSimd    [c + 1];
 115                     vCoulomb[i*numGroups + j0] += vCoulombSimd[c + 0];
 116                     vCoulomb[i*numGroups + j1] += vCoulombSimd[c + 1];
 117                     c                          += unrollj + 2;
 118                 }
 119             }
 120         }
 121     }
 122 }
 123
 124 /*! \brief Dispatches the non-bonded N versus M atom cluster CPU kernels.
 125  *
 126  * OpenMP parallelization is performed within this function.
 127  * Energy reduction, but not force and shift force reduction, is performed
 128  * within this function.
 129  *
 130  * \param[in]     nbvg          The group (local/non-local) to compute interaction for
 131  * \param[in,out] nbat          The atomdata for the interactions
 132  * \param[in]     ic            Non-bonded interaction constants
 133  * \param[in]     shiftVectors  The PBC shift vectors
 134  * \param[in]     forceFlags    Flags that tell what to compute
 135  * \param[in]     clearF        Enum that tells if to clear the force output buffer
 136  * \param[out]    fshift        Shift force output buffer
 137  * \param[out]    vCoulomb      Output buffer for Coulomb energies
 138  * \param[out]    vVdw          Output buffer for Van der Waals energies
 139  */
 140 static void
 141 nbnxn_kernel_cpu(const nonbonded_verlet_group_t *nbvg,
 142                  nbnxn_atomdata_t               *nbat,
 143                  const interaction_const_t      &ic,
 144                  rvec                           *shiftVectors,
 145                  int                             forceFlags,
 146                  int                             clearF,
 147                  real                           *fshift,
 148                  real                           *vCoulomb,
 149                  real                           *vVdw)
 150 {
 151
 152     int                      coulkt;
 153     if (EEL_RF(ic.eeltype) || ic.eeltype == eelCUT)
 154     {
 155         coulkt = coulktRF;
 156     }
 157     else
 158     {
 159         if (nbvg->ewald_excl == ewaldexclTable)
 160         {
 161             if (ic.rcoulomb == ic.rvdw)
 162             {
 163                 coulkt = coulktTAB;
 164             }
 165             else
 166             {
 167                 coulkt = coulktTAB_TWIN;
 168             }
 169         }
 170         else
 171         {
 172             if (ic.rcoulomb == ic.rvdw)
 173             {
 174                 coulkt = coulktEWALD;
 175             }
 176             else
 177             {
 178                 coulkt = coulktEWALD_TWIN;
 179             }
 180         }
 181     }
 182
 183     const nbnxn_atomdata_t::Params &nbatParams = nbat->params();
 184
 185     int vdwkt = 0;
 186     if (ic.vdwtype == evdwCUT)
 187     {
 188         switch (ic.vdw_modifier)
 189         {
 190             case eintmodNONE:
 191             case eintmodPOTSHIFT:
 192                 switch (nbatParams.comb_rule)
 193                 {
 194                     case ljcrGEOM: vdwkt = vdwktLJCUT_COMBGEOM; break;
 195                     case ljcrLB:   vdwkt = vdwktLJCUT_COMBLB;   break;
 196                     case ljcrNONE: vdwkt = vdwktLJCUT_COMBNONE; break;
 197                     default:
 198                         GMX_RELEASE_ASSERT(false, "Unknown combination rule");
 199                 }
 200                 break;
 201             case eintmodFORCESWITCH:
 202                 vdwkt = vdwktLJFORCESWITCH;
 203                 break;
 204             case eintmodPOTSWITCH:
 205                 vdwkt = vdwktLJPOTSWITCH;
 206                 break;
 207             default:
 208                 GMX_RELEASE_ASSERT(false, "Unsupported VdW interaction modifier");
 209         }
 210     }
 211     else if (ic.vdwtype == evdwPME)
 212     {
 213         if (ic.ljpme_comb_rule == eljpmeGEOM)
 214         {
 215             vdwkt = vdwktLJEWALDCOMBGEOM;
 216         }
 217         else
 218         {
 219             vdwkt = vdwktLJEWALDCOMBLB;
 220             /* At setup we (should have) selected the C reference kernel */
 221             GMX_RELEASE_ASSERT(nbvg->kernel_type == nbnxnk4x4_PlainC, "Only the C reference nbnxn SIMD kernel supports LJ-PME with LB combination rules");
 222         }
 223     }
 224     else
 225     {
 226         GMX_RELEASE_ASSERT(false, "Unsupported VdW interaction type");
 227     }
 228
 229     int                        nnbl = nbvg->nbl_lists.nnbl;
 230     NbnxnPairlistCpu * const * nbl  = nbvg->nbl_lists.nbl;
 231
 232     int gmx_unused             nthreads = gmx_omp_nthreads_get(emntNonbonded);
 233 #pragma omp parallel for schedule(static) num_threads(nthreads)
 234     for (int nb = 0; nb < nnbl; nb++)
 235     {
 236         // Presently, the kernels do not call C++ code that can throw,
 237         // so no need for a try/catch pair in this OpenMP region.
 238         nbnxn_atomdata_output_t *out = &nbat->out[nb];
 239
 240         if (clearF == enbvClearFYes)
 241         {
 242             clear_f(nbat, nb, out->f.data());
 243         }
 244
 245         real *fshift_p;
 246         if ((forceFlags & GMX_FORCE_VIRIAL) && nnbl == 1)
 247         {
 248             fshift_p = fshift;
 249         }
 250         else
 251         {
 252             fshift_p = out->fshift.data();
 253
 254             if (clearF == enbvClearFYes)
 255             {
 256                 clear_fshift(fshift_p);
 257             }
 258         }
 259
 260         if (!(forceFlags & GMX_FORCE_ENERGY))
 261         {
 262             /* Don't calculate energies */
 263             switch (nbvg->kernel_type)
 264             {
 265                 case nbnxnk4x4_PlainC:
 266                     nbnxn_kernel_noener_ref[coulkt][vdwkt](nbl[nb], nbat,
 267                                                            &ic,
 268                                                            shiftVectors,
 269                                                            out->f.data(),
 270                                                            fshift_p);
 271                     break;
 272 #ifdef GMX_NBNXN_SIMD_2XNN
 273                 case nbnxnk4xN_SIMD_2xNN:
 274                     nbnxm_kernel_noener_simd_2xmm[coulkt][vdwkt](nbl[nb], nbat,
 275                                                                  &ic,
 276                                                                  shiftVectors,
 277                                                                  out->f.data(),
 278                                                                  fshift_p);
 279                     break;
 280 #endif
 281 #ifdef GMX_NBNXN_SIMD_4XN
 282                 case nbnxnk4xN_SIMD_4xN:
 283                     nbnxm_kernel_noener_simd_4xm[coulkt][vdwkt](nbl[nb], nbat,
 284                                                                 &ic,
 285                                                                 shiftVectors,
 286                                                                 out->f.data(),
 287                                                                 fshift_p);
 288                     break;
 289 #endif
 290                 default:
 291                     GMX_RELEASE_ASSERT(false, "Unsupported kernel architecture");
 292             }
 293         }
 294         else if (out->Vvdw.size() == 1)
 295         {
 296             /* A single energy group (pair) */
 297             out->Vvdw[0] = 0;
 298             out->Vc[0]   = 0;
 299
 300             switch (nbvg->kernel_type)
 301             {
 302                 case nbnxnk4x4_PlainC:
 303                     nbnxn_kernel_ener_ref[coulkt][vdwkt](nbl[nb], nbat,
 304                                                          &ic,
 305                                                          shiftVectors,
 306                                                          out->f.data(),
 307                                                          fshift_p,
 308                                                          out->Vvdw.data(),
 309                                                          out->Vc.data());
 310                     break;
 311 #ifdef GMX_NBNXN_SIMD_2XNN
 312                 case nbnxnk4xN_SIMD_2xNN:
 313                     nbnxm_kernel_ener_simd_2xmm[coulkt][vdwkt](nbl[nb], nbat,
 314                                                                &ic,
 315                                                                shiftVectors,
 316                                                                out->f.data(),
 317                                                                fshift_p,
 318                                                                out->Vvdw.data(),
 319                                                                out->Vc.data());
 320                     break;
 321 #endif
 322 #ifdef GMX_NBNXN_SIMD_4XN
 323                 case nbnxnk4xN_SIMD_4xN:
 324                     nbnxm_kernel_ener_simd_4xm[coulkt][vdwkt](nbl[nb], nbat,
 325                                                               &ic,
 326                                                               shiftVectors,
 327                                                               out->f.data(),
 328                                                               fshift_p,
 329                                                               out->Vvdw.data(),
 330                                                               out->Vc.data());
 331                     break;
 332 #endif
 333                 default:
 334                     GMX_RELEASE_ASSERT(false, "Unsupported kernel architecture");
 335             }
 336         }
 337         else
 338         {
 339             /* Calculate energy group contributions */
 340             clearGroupEnergies(out);
 341
 342             int unrollj = 0;
 343
 344             switch (nbvg->kernel_type)
 345             {
 346                 case nbnxnk4x4_PlainC:
 347                     unrollj = c_nbnxnCpuIClusterSize;
 348                     nbnxn_kernel_energrp_ref[coulkt][vdwkt](nbl[nb], nbat,
 349                                                             &ic,
 350                                                             shiftVectors,
 351                                                             out->f.data(),
 352                                                             fshift_p,
 353                                                             out->Vvdw.data(),
 354                                                             out->Vc.data());
 355                     break;
 356 #ifdef GMX_NBNXN_SIMD_2XNN
 357                 case nbnxnk4xN_SIMD_2xNN:
 358                     unrollj = GMX_SIMD_REAL_WIDTH/2;
 359                     nbnxm_kernel_energrp_simd_2xmm[coulkt][vdwkt](nbl[nb], nbat,
 360                                                                   &ic,
 361                                                                   shiftVectors,
 362                                                                   out->f.data(),
 363                                                                   fshift_p,
 364                                                                   out->VSvdw.data(),
 365                                                                   out->VSc.data());
 366                     break;
 367 #endif
 368 #ifdef GMX_NBNXN_SIMD_4XN
 369                 case nbnxnk4xN_SIMD_4xN:
 370                     unrollj = GMX_SIMD_REAL_WIDTH;
 371                     nbnxm_kernel_energrp_simd_4xm[coulkt][vdwkt](nbl[nb], nbat,
 372                                                                  &ic,
 373                                                                  shiftVectors,
 374                                                                  out->f.data(),
 375                                                                  fshift_p,
 376                                                                  out->VSvdw.data(),
 377                                                                  out->VSc.data());
 378                     break;
 379 #endif
 380                 default:
 381                     GMX_RELEASE_ASSERT(false, "Unsupported kernel architecture");
 382             }
 383
 384             if (nbvg->kernel_type != nbnxnk4x4_PlainC)
 385             {
 386                 switch (unrollj)
 387                 {
 388                     case 2:
 389                         reduceGroupEnergySimdBuffers<2>(nbatParams.nenergrp,
 390                                                         nbatParams.neg_2log,
 391                                                         out);
 392                         break;
 393                     case 4:
 394                         reduceGroupEnergySimdBuffers<4>(nbatParams.nenergrp,
 395                                                         nbatParams.neg_2log,
 396                                                         out);
 397                         break;
 398                     case 8:
 399                         reduceGroupEnergySimdBuffers<8>(nbatParams.nenergrp,
 400                                                         nbatParams.neg_2log,
 401                                                         out);
 402                         break;
 403                     default:
 404                         GMX_RELEASE_ASSERT(false, "Unsupported j-unroll size");
 405                 }
 406             }
 407         }
 408     }
 409
 410     if (forceFlags & GMX_FORCE_ENERGY)
 411     {
 412         reduce_energies_over_lists(nbat, nnbl, vVdw, vCoulomb);
 413     }
 414 }
 415
 416 static void accountFlops(t_nrnb                           *nrnb,
 417                          const nonbonded_verlet_t         &nbv,
 418                          const Nbnxm::InteractionLocality  iLocality,
 419                          const interaction_const_t        &ic,
 420                          const int                         forceFlags)
 421 {
 422     const nonbonded_verlet_group_t &nbvg            = nbv.grp[iLocality];
 423     const bool                      usingGpuKernels = (nbvg.kernel_type == nbnxnk8x8x8_GPU);
 424
 425     int enr_nbnxn_kernel_ljc;
 426     if (EEL_RF(ic.eeltype) || ic.eeltype == eelCUT)
 427     {
 428         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_RF;
 429     }
 430     else if ((!usingGpuKernels && nbvg.ewald_excl == ewaldexclAnalytical) ||
 431              (usingGpuKernels && Nbnxm::gpu_is_kernel_ewald_analytical(nbv.gpu_nbv)))
 432     {
 433         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_EWALD;
 434     }
 435     else
 436     {
 437         enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_TAB;
 438     }
 439     int enr_nbnxn_kernel_lj = eNR_NBNXN_LJ;
 440     if (forceFlags & GMX_FORCE_ENERGY)
 441     {
 442         /* In eNR_??? the nbnxn F+E kernels are always the F kernel + 1 */
 443         enr_nbnxn_kernel_ljc += 1;
 444         enr_nbnxn_kernel_lj  += 1;
 445     }
 446
 447     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc,
 448              nbvg.nbl_lists.natpair_ljq);
 449     inc_nrnb(nrnb, enr_nbnxn_kernel_lj,
 450              nbvg.nbl_lists.natpair_lj);
 451     /* The Coulomb-only kernels are offset -eNR_NBNXN_LJ_RF+eNR_NBNXN_RF */
 452     inc_nrnb(nrnb, enr_nbnxn_kernel_ljc-eNR_NBNXN_LJ_RF+eNR_NBNXN_RF,
 453              nbvg.nbl_lists.natpair_q);
 454
 455     const bool calcEnergy = ((forceFlags & GMX_FORCE_ENERGY) != 0);
 456     if (ic.vdw_modifier == eintmodFORCESWITCH)
 457     {
 458         /* We add up the switch cost separately */
 459         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_FSW + (calcEnergy ? 1 : 0),
 460                  nbvg.nbl_lists.natpair_ljq + nbvg.nbl_lists.natpair_lj);
 461     }
 462     if (ic.vdw_modifier == eintmodPOTSWITCH)
 463     {
 464         /* We add up the switch cost separately */
 465         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_PSW + (calcEnergy ? 1 : 0),
 466                  nbvg.nbl_lists.natpair_ljq + nbvg.nbl_lists.natpair_lj);
 467     }
 468     if (ic.vdwtype == evdwPME)
 469     {
 470         /* We add up the LJ Ewald cost separately */
 471         inc_nrnb(nrnb, eNR_NBNXN_ADD_LJ_EWALD + (calcEnergy ? 1 : 0),
 472                  nbvg.nbl_lists.natpair_ljq + nbvg.nbl_lists.natpair_lj);
 473     }
 474 }
 475
 476 void NbnxnDispatchKernel(nonbonded_verlet_t        *nbv,
 477                          Nbnxm::InteractionLocality iLocality,
 478                          const interaction_const_t &ic,
 479                          int                        forceFlags,
 480                          int                        clearF,
 481                          t_forcerec                *fr,
 482                          gmx_enerdata_t            *enerd,
 483                          t_nrnb                    *nrnb)
 484 {
 485     const nonbonded_verlet_group_t &nbvg = nbv->grp[iLocality];
 486
 487     switch (nbvg.kernel_type)
 488     {
 489         case nbnxnk4x4_PlainC:
 490         case nbnxnk4xN_SIMD_4xN:
 491         case nbnxnk4xN_SIMD_2xNN:
 492             nbnxn_kernel_cpu(&nbvg,
 493                              nbv->nbat,
 494                              ic,
 495                              fr->shift_vec,
 496                              forceFlags,
 497                              clearF,
 498                              fr->fshift[0],
 499                              enerd->grpp.ener[egCOULSR],
 500                              fr->bBHAM ?
 501                              enerd->grpp.ener[egBHAMSR] :
 502                              enerd->grpp.ener[egLJSR]);
 503             break;
 504
 505         case nbnxnk8x8x8_GPU:
 506             Nbnxm::gpu_launch_kernel(nbv->gpu_nbv, forceFlags, iLocality);
 507             break;
 508
 509         case nbnxnk8x8x8_PlainC:
 510             nbnxn_kernel_gpu_ref(nbvg.nbl_lists.nblGpu[0],
 511                                  nbv->nbat, &ic,
 512                                  fr->shift_vec,
 513                                  forceFlags,
 514                                  clearF,
 515                                  nbv->nbat->out[0].f,
 516                                  fr->fshift[0],
 517                                  enerd->grpp.ener[egCOULSR],
 518                                  fr->bBHAM ?
 519                                  enerd->grpp.ener[egBHAMSR] :
 520                                  enerd->grpp.ener[egLJSR]);
 521             break;
 522
 523         default:
 524             GMX_RELEASE_ASSERT(false, "Invalid nonbonded kernel type passed!");
 525
 526     }
 527
 528     accountFlops(nrnb, *nbv, iLocality, ic, forceFlags);
 529 }