src/gromacs/gpu_utils/device_stream_manager.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2019,2020, 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 /*! \internal \file
  36  *
  37  * \brief Implements GPU stream manager.
  38  *
  39  * \author Mark Abraham <mark.j.abraham@gmail.com>
  40  * \author Artem Zhmurov <zhmurov@gmail.com>
  41  *
  42  * \ingroup module_gpu_utils
  43  */
  44 #include "gmxpre.h"
  45
  46 #include "device_stream_manager.h"
  47
  48 #include "gromacs/gpu_utils/device_context.h"
  49 #include "gromacs/gpu_utils/device_stream.h"
  50 #include "gromacs/mdtypes/simulation_workload.h"
  51 #include "gromacs/utility/enumerationhelpers.h"
  52 #include "gromacs/utility/exceptions.h"
  53 #include "gromacs/utility/gmxassert.h"
  54 #include "gromacs/utility/stringutil.h"
  55
  56 namespace gmx
  57 {
  58
  59 /*! \libinternal
  60  * \brief Impl class to manages the lifetime of the GPU streams.
  61  *
  62  * If supported by the GPU API, the available runtime and the
  63  * indicated device, some streams will be configured at high
  64  * priority. Otherwise, all streams will share the default priority
  65  * appropriate to the situation.
  66  */
  67 class DeviceStreamManager::Impl
  68 {
  69 public:
  70     /*! \brief Constructor.
  71      *
  72      * \throws InternalError  If any of the required resources could not be initialized.
  73      */
  74     Impl(const DeviceInformation& deviceInfo,
  75          bool                     havePpDomainDecomposition,
  76          SimulationWorkload       simulationWork,
  77          bool                     useTiming);
  78     ~Impl();
  79
  80     //! Device context.
  81     DeviceContext context_;
  82     //! GPU command streams.
  83     EnumerationArray<DeviceStreamType, std::unique_ptr<DeviceStream>> streams_;
  84 };
  85
  86 // DeviceStreamManager::Impl
  87 DeviceStreamManager::Impl::Impl(const DeviceInformation& deviceInfo,
  88                                 const bool               havePpDomainDecomposition,
  89                                 const SimulationWorkload simulationWork,
  90                                 const bool               useTiming) :
  91     context_(deviceInfo)
  92 {
  93     try
  94     {
  95         streams_[DeviceStreamType::NonBondedLocal] =
  96                 std::make_unique<DeviceStream>(context_, DeviceStreamPriority::Normal, useTiming);
  97
  98         if (simulationWork.useGpuPme)
  99         {
 100             /* Creating a PME GPU stream:
 101              * - default high priority with CUDA
 102              * - no priorities implemented yet with OpenCL; see #2532
 103              */
 104             streams_[DeviceStreamType::Pme] =
 105                     std::make_unique<DeviceStream>(context_, DeviceStreamPriority::High, useTiming);
 106         }
 107
 108         if (havePpDomainDecomposition)
 109         {
 110             streams_[DeviceStreamType::NonBondedNonLocal] =
 111                     std::make_unique<DeviceStream>(context_, DeviceStreamPriority::High, useTiming);
 112         }
 113         // Update stream is used both for coordinates transfers and for GPU update/constraints
 114         if (simulationWork.useGpuPme || simulationWork.useGpuUpdate || simulationWork.useGpuBufferOps)
 115         {
 116             streams_[DeviceStreamType::UpdateAndConstraints] =
 117                     std::make_unique<DeviceStream>(context_, DeviceStreamPriority::Normal, useTiming);
 118         }
 119         if (simulationWork.useGpuPmePpCommunication)
 120         {
 121             streams_[DeviceStreamType::PmePpTransfer] =
 122                     std::make_unique<DeviceStream>(context_, DeviceStreamPriority::Normal, useTiming);
 123         }
 124     }
 125     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 126 }
 127
 128 DeviceStreamManager::Impl::~Impl() = default;
 129
 130 // DeviceStreamManager
 131 DeviceStreamManager::DeviceStreamManager(const DeviceInformation& deviceInfo,
 132                                          const bool               havePpDomainDecomposition,
 133                                          const SimulationWorkload simulationWork,
 134                                          const bool               useTiming) :
 135     impl_(new Impl(deviceInfo, havePpDomainDecomposition, simulationWork, useTiming))
 136 {
 137 }
 138
 139 DeviceStreamManager::~DeviceStreamManager() = default;
 140
 141 const DeviceInformation& DeviceStreamManager::deviceInfo() const
 142 {
 143     return impl_->context_.deviceInfo();
 144 }
 145
 146 const DeviceContext& DeviceStreamManager::context() const
 147 {
 148     return impl_->context_;
 149 }
 150
 151 const DeviceStream& DeviceStreamManager::stream(DeviceStreamType streamToGet) const
 152 {
 153     return *impl_->streams_[streamToGet];
 154 }
 155
 156 const DeviceStream& DeviceStreamManager::bondedStream(bool hasPPDomainDecomposition) const
 157 {
 158     if (hasPPDomainDecomposition)
 159     {
 160         GMX_RELEASE_ASSERT(stream(DeviceStreamType::NonBondedNonLocal).isValid(),
 161                            "GPU non-bonded non-local stream should be valid in order to use GPU "
 162                            "version of bonded forces with domain decomposition.");
 163         return stream(DeviceStreamType::NonBondedNonLocal);
 164     }
 165     else
 166     {
 167         GMX_RELEASE_ASSERT(stream(DeviceStreamType::NonBondedLocal).isValid(),
 168                            "GPU non-bonded local stream should be valid in order to use GPU "
 169                            "version of bonded forces without domain decomposition.");
 170         return stream(DeviceStreamType::NonBondedLocal);
 171     }
 172 }
 173
 174 bool DeviceStreamManager::streamIsValid(DeviceStreamType streamToCheck) const
 175 {
 176     return impl_->streams_[streamToCheck] != nullptr && impl_->streams_[streamToCheck]->isValid();
 177 }
 178
 179 } // namespace gmx