src/gromacs/nbnxm/nbnxm_gpu_data_mgmt.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2016 by the GROMACS development team.
   5  * Copyright (c) 2017,2018,2019,2020,2021, by the GROMACS development team, led by
   6  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   7  * and including many others, as listed in the AUTHORS file in the
   8  * top-level source directory and at http://www.gromacs.org.
   9  *
  10  * GROMACS is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public License
  12  * as published by the Free Software Foundation; either version 2.1
  13  * of the License, or (at your option) any later version.
  14  *
  15  * GROMACS is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with GROMACS; if not, see
  22  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  23  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  24  *
  25  * If you want to redistribute modifications to GROMACS, please
  26  * consider that scientific software is very special. Version
  27  * control is crucial - bugs must be traceable. We will be happy to
  28  * consider code for inclusion in the official distribution, but
  29  * derived work must not be called official GROMACS. Details are found
  30  * in the README & COPYING files - if they are missing, get the
  31  * official version at http://www.gromacs.org.
  32  *
  33  * To help us fund GROMACS development, we humbly ask that you cite
  34  * the research papers on the package. Check out http://www.gromacs.org.
  35  */
  36 /*! \internal \file
  37  *  \brief Define common implementation of nbnxm_gpu_data_mgmt.h
  38  *
  39  *  \author Anca Hamuraru <anca@streamcomputing.eu>
  40  *  \author Dimitrios Karkoulis <dimitris.karkoulis@gmail.com>
  41  *  \author Teemu Virolainen <teemu@streamcomputing.eu>
  42  *  \author Szilárd Páll <pall.szilard@gmail.com>
  43  *  \author Artem Zhmurov <zhmurov@gmail.com>
  44  *
  45  *  \ingroup module_nbnxm
  46  */
  47 #include "gmxpre.h"
  48
  49 #include "config.h"
  50
  51 #if GMX_GPU_CUDA
  52 #    include "cuda/nbnxm_cuda_types.h"
  53 #endif
  54
  55 #if GMX_GPU_OPENCL
  56 #    include "opencl/nbnxm_ocl_types.h"
  57 #endif
  58
  59 #include "nbnxm_gpu_data_mgmt.h"
  60
  61 #include "gromacs/hardware/device_information.h"
  62 #include "gromacs/mdtypes/interaction_const.h"
  63 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  64 #include "gromacs/timing/gpu_timing.h"
  65 #include "gromacs/utility/cstringutil.h"
  66
  67 #include "nbnxm_gpu.h"
  68 #include "pairlistsets.h"
  69
  70 namespace Nbnxm
  71 {
  72
  73 void init_ewald_coulomb_force_table(const EwaldCorrectionTables& tables,
  74                                     NBParamGpu*                  nbp,
  75                                     const DeviceContext&         deviceContext)
  76 {
  77     if (nbp->coulomb_tab)
  78     {
  79         destroyParamLookupTable(&nbp->coulomb_tab, nbp->coulomb_tab_texobj);
  80     }
  81
  82     nbp->coulomb_tab_scale = tables.scale;
  83     initParamLookupTable(
  84             &nbp->coulomb_tab, &nbp->coulomb_tab_texobj, tables.tableF.data(), tables.tableF.size(), deviceContext);
  85 }
  86
  87 enum ElecType nbnxn_gpu_pick_ewald_kernel_type(const interaction_const_t& ic,
  88                                                const DeviceInformation gmx_unused& deviceInfo)
  89 {
  90     bool bTwinCut = (ic.rcoulomb != ic.rvdw);
  91
  92     /* Benchmarking/development environment variables to force the use of
  93        analytical or tabulated Ewald kernel. */
  94     const bool forceAnalyticalEwald = (getenv("GMX_GPU_NB_ANA_EWALD") != nullptr);
  95     const bool forceTabulatedEwald  = (getenv("GMX_GPU_NB_TAB_EWALD") != nullptr);
  96     const bool forceTwinCutoffEwald = (getenv("GMX_GPU_NB_EWALD_TWINCUT") != nullptr);
  97
  98     if (forceAnalyticalEwald && forceTabulatedEwald)
  99     {
 100         gmx_incons(
 101                 "Both analytical and tabulated Ewald GPU non-bonded kernels "
 102                 "requested through environment variables.");
 103     }
 104
 105     /* By default, use analytical Ewald except with CUDA on NVIDIA CC 7.0 and 8.0.
 106      */
 107     const bool c_useTabulatedEwaldDefault =
 108 #if GMX_GPU_CUDA
 109             (deviceInfo.prop.major == 7 && deviceInfo.prop.minor == 0)
 110             || (deviceInfo.prop.major == 8 && deviceInfo.prop.minor == 0);
 111 #else
 112             false;
 113 #endif
 114     bool bUseAnalyticalEwald = !c_useTabulatedEwaldDefault;
 115     if (forceAnalyticalEwald)
 116     {
 117         bUseAnalyticalEwald = true;
 118         if (debug)
 119         {
 120             fprintf(debug, "Using analytical Ewald GPU kernels\n");
 121         }
 122     }
 123     else if (forceTabulatedEwald)
 124     {
 125         bUseAnalyticalEwald = false;
 126
 127         if (debug)
 128         {
 129             fprintf(debug, "Using tabulated Ewald GPU kernels\n");
 130         }
 131     }
 132
 133     /* Use twin cut-off kernels if requested by bTwinCut or the env. var.
 134        forces it (use it for debugging/benchmarking only). */
 135     if (!bTwinCut && !forceTwinCutoffEwald)
 136     {
 137         return bUseAnalyticalEwald ? ElecType::EwaldAna : ElecType::EwaldTab;
 138     }
 139     else
 140     {
 141         return bUseAnalyticalEwald ? ElecType::EwaldAnaTwin : ElecType::EwaldTabTwin;
 142     }
 143 }
 144
 145 void set_cutoff_parameters(NBParamGpu* nbp, const interaction_const_t* ic, const PairlistParams& listParams)
 146 {
 147     nbp->ewald_beta        = ic->ewaldcoeff_q;
 148     nbp->sh_ewald          = ic->sh_ewald;
 149     nbp->epsfac            = ic->epsfac;
 150     nbp->two_k_rf          = 2.0 * ic->k_rf;
 151     nbp->c_rf              = ic->c_rf;
 152     nbp->rvdw_sq           = ic->rvdw * ic->rvdw;
 153     nbp->rcoulomb_sq       = ic->rcoulomb * ic->rcoulomb;
 154     nbp->rlistOuter_sq     = listParams.rlistOuter * listParams.rlistOuter;
 155     nbp->rlistInner_sq     = listParams.rlistInner * listParams.rlistInner;
 156     nbp->useDynamicPruning = listParams.useDynamicPruning;
 157
 158     nbp->sh_lj_ewald   = ic->sh_lj_ewald;
 159     nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
 160
 161     nbp->rvdw_switch      = ic->rvdw_switch;
 162     nbp->dispersion_shift = ic->dispersion_shift;
 163     nbp->repulsion_shift  = ic->repulsion_shift;
 164     nbp->vdw_switch       = ic->vdw_switch;
 165 }
 166
 167 void gpu_pme_loadbal_update_param(const nonbonded_verlet_t* nbv, const interaction_const_t* ic)
 168 {
 169     if (!nbv || !nbv->useGpu())
 170     {
 171         return;
 172     }
 173     NbnxmGpu*   nb  = nbv->gpu_nbv;
 174     NBParamGpu* nbp = nb->nbparam;
 175
 176     set_cutoff_parameters(nbp, ic, nbv->pairlistSets().params());
 177
 178     nbp->elecType = nbnxn_gpu_pick_ewald_kernel_type(*ic, nb->deviceContext_->deviceInfo());
 179
 180     GMX_RELEASE_ASSERT(ic->coulombEwaldTables, "Need valid Coulomb Ewald correction tables");
 181     init_ewald_coulomb_force_table(*ic->coulombEwaldTables, nbp, *nb->deviceContext_);
 182 }
 183
 184 void init_plist(gpu_plist* pl)
 185 {
 186     /* initialize to nullptr pointers to data that is not allocated here and will
 187        need reallocation in nbnxn_gpu_init_pairlist */
 188     pl->sci   = nullptr;
 189     pl->cj4   = nullptr;
 190     pl->imask = nullptr;
 191     pl->excl  = nullptr;
 192
 193     /* size -1 indicates that the respective array hasn't been initialized yet */
 194     pl->na_c                   = -1;
 195     pl->nsci                   = -1;
 196     pl->sci_nalloc             = -1;
 197     pl->ncj4                   = -1;
 198     pl->cj4_nalloc             = -1;
 199     pl->nimask                 = -1;
 200     pl->imask_nalloc           = -1;
 201     pl->nexcl                  = -1;
 202     pl->excl_nalloc            = -1;
 203     pl->haveFreshList          = false;
 204     pl->rollingPruningNumParts = 0;
 205     pl->rollingPruningPart     = 0;
 206 }
 207
 208 void init_timings(gmx_wallclock_gpu_nbnxn_t* t)
 209 {
 210     t->nb_h2d_t = 0.0;
 211     t->nb_d2h_t = 0.0;
 212     t->nb_c     = 0;
 213     t->pl_h2d_t = 0.0;
 214     t->pl_h2d_c = 0;
 215     for (int i = 0; i < 2; i++)
 216     {
 217         for (int j = 0; j < 2; j++)
 218         {
 219             t->ktime[i][j].t = 0.0;
 220             t->ktime[i][j].c = 0;
 221         }
 222     }
 223     t->pruneTime.c        = 0;
 224     t->pruneTime.t        = 0.0;
 225     t->dynamicPruneTime.c = 0;
 226     t->dynamicPruneTime.t = 0.0;
 227 }
 228
 229 //! This function is documented in the header file
 230 void gpu_init_pairlist(NbnxmGpu* nb, const NbnxnPairlistGpu* h_plist, const InteractionLocality iloc)
 231 {
 232     char sbuf[STRLEN];
 233     // Timing accumulation should happen only if there was work to do
 234     // because getLastRangeTime() gets skipped with empty lists later
 235     // which leads to the counter not being reset.
 236     bool                bDoTime      = (nb->bDoTime && !h_plist->sci.empty());
 237     const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
 238     gpu_plist*          d_plist      = nb->plist[iloc];
 239
 240     if (d_plist->na_c < 0)
 241     {
 242         d_plist->na_c = h_plist->na_ci;
 243     }
 244     else
 245     {
 246         if (d_plist->na_c != h_plist->na_ci)
 247         {
 248             sprintf(sbuf,
 249                     "In init_plist: the #atoms per cell has changed (from %d to %d)",
 250                     d_plist->na_c,
 251                     h_plist->na_ci);
 252             gmx_incons(sbuf);
 253         }
 254     }
 255
 256     gpu_timers_t::Interaction& iTimers = nb->timers->interaction[iloc];
 257
 258     if (bDoTime)
 259     {
 260         iTimers.pl_h2d.openTimingRegion(deviceStream);
 261         iTimers.didPairlistH2D = true;
 262     }
 263
 264     // TODO most of this function is same in CUDA and OpenCL, move into the header
 265     const DeviceContext& deviceContext = *nb->deviceContext_;
 266
 267     reallocateDeviceBuffer(
 268             &d_plist->sci, h_plist->sci.size(), &d_plist->nsci, &d_plist->sci_nalloc, deviceContext);
 269     copyToDeviceBuffer(&d_plist->sci,
 270                        h_plist->sci.data(),
 271                        0,
 272                        h_plist->sci.size(),
 273                        deviceStream,
 274                        GpuApiCallBehavior::Async,
 275                        bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
 276
 277     reallocateDeviceBuffer(
 278             &d_plist->cj4, h_plist->cj4.size(), &d_plist->ncj4, &d_plist->cj4_nalloc, deviceContext);
 279     copyToDeviceBuffer(&d_plist->cj4,
 280                        h_plist->cj4.data(),
 281                        0,
 282                        h_plist->cj4.size(),
 283                        deviceStream,
 284                        GpuApiCallBehavior::Async,
 285                        bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
 286
 287     reallocateDeviceBuffer(&d_plist->imask,
 288                            h_plist->cj4.size() * c_nbnxnGpuClusterpairSplit,
 289                            &d_plist->nimask,
 290                            &d_plist->imask_nalloc,
 291                            deviceContext);
 292
 293     reallocateDeviceBuffer(
 294             &d_plist->excl, h_plist->excl.size(), &d_plist->nexcl, &d_plist->excl_nalloc, deviceContext);
 295     copyToDeviceBuffer(&d_plist->excl,
 296                        h_plist->excl.data(),
 297                        0,
 298                        h_plist->excl.size(),
 299                        deviceStream,
 300                        GpuApiCallBehavior::Async,
 301                        bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
 302
 303     if (bDoTime)
 304     {
 305         iTimers.pl_h2d.closeTimingRegion(deviceStream);
 306     }
 307
 308     /* need to prune the pair list during the next step */
 309     d_plist->haveFreshList = true;
 310 }
 311
 312 //! This function is documented in the header file
 313 gmx_wallclock_gpu_nbnxn_t* gpu_get_timings(NbnxmGpu* nb)
 314 {
 315     return (nb != nullptr && nb->bDoTime) ? nb->timings : nullptr;
 316 }
 317
 318 //! This function is documented in the header file
 319 void gpu_reset_timings(nonbonded_verlet_t* nbv)
 320 {
 321     if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
 322     {
 323         init_timings(nbv->gpu_nbv->timings);
 324     }
 325 }
 326
 327 bool gpu_is_kernel_ewald_analytical(const NbnxmGpu* nb)
 328 {
 329     return ((nb->nbparam->elecType == ElecType::EwaldAna)
 330             || (nb->nbparam->elecType == ElecType::EwaldAnaTwin));
 331 }
 332
 333 enum ElecType nbnxmGpuPickElectrostaticsKernelType(const interaction_const_t* ic,
 334                                                    const DeviceInformation&   deviceInfo)
 335 {
 336     if (ic->eeltype == eelCUT)
 337     {
 338         return ElecType::Cut;
 339     }
 340     else if (EEL_RF(ic->eeltype))
 341     {
 342         return ElecType::RF;
 343     }
 344     else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
 345     {
 346         return nbnxn_gpu_pick_ewald_kernel_type(*ic, deviceInfo);
 347     }
 348     else
 349     {
 350         /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
 351         GMX_THROW(gmx::InconsistentInputError(
 352                 gmx::formatString("The requested electrostatics type %s (%d) is not implemented in "
 353                                   "the GPU accelerated kernels!",
 354                                   EELTYPE(ic->eeltype),
 355                                   ic->eeltype)));
 356     }
 357 }
 358
 359
 360 enum VdwType nbnxmGpuPickVdwKernelType(const interaction_const_t* ic, LJCombinationRule ljCombinationRule)
 361 {
 362     if (ic->vdwtype == evdwCUT)
 363     {
 364         switch (ic->vdw_modifier)
 365         {
 366             case eintmodNONE:
 367             case eintmodPOTSHIFT:
 368                 switch (ljCombinationRule)
 369                 {
 370                     case LJCombinationRule::None: return VdwType::Cut;
 371                     case LJCombinationRule::Geometric: return VdwType::CutCombGeom;
 372                     case LJCombinationRule::LorentzBerthelot: return VdwType::CutCombLB;
 373                     default:
 374                         GMX_THROW(gmx::InconsistentInputError(gmx::formatString(
 375                                 "The requested LJ combination rule %s is not implemented in "
 376                                 "the GPU accelerated kernels!",
 377                                 enumValueToString(ljCombinationRule))));
 378                 }
 379             case eintmodFORCESWITCH: return VdwType::FSwitch;
 380             case eintmodPOTSWITCH: return VdwType::PSwitch;
 381             default:
 382                 GMX_THROW(gmx::InconsistentInputError(
 383                         gmx::formatString("The requested VdW interaction modifier %s (%d) is not "
 384                                           "implemented in the GPU accelerated kernels!",
 385                                           INTMODIFIER(ic->vdw_modifier),
 386                                           ic->vdw_modifier)));
 387         }
 388     }
 389     else if (ic->vdwtype == evdwPME)
 390     {
 391         if (ic->ljpme_comb_rule == eljpmeGEOM)
 392         {
 393             assert(ljCombinationRule == LJCombinationRule::Geometric);
 394             return VdwType::EwaldGeom;
 395         }
 396         else
 397         {
 398             assert(ljCombinationRule == LJCombinationRule::LorentzBerthelot);
 399             return VdwType::EwaldLB;
 400         }
 401     }
 402     else
 403     {
 404         GMX_THROW(gmx::InconsistentInputError(gmx::formatString(
 405                 "The requested VdW type %s (%d) is not implemented in the GPU accelerated kernels!",
 406                 EVDWTYPE(ic->vdwtype),
 407                 ic->vdwtype)));
 408     }
 409 }
 410
 411 } // namespace Nbnxm