src/gromacs/mdtypes/state_propagator_data_gpu.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 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 /*! \internal \file
  36  *
  37  * \brief Declaration of interfaces for GPU state data propagator object.
  38  *
  39  * This object stores and manages positions, velocities and forces for
  40  * all particles in the system on the GPU.
  41  *
  42  * \todo Add cycle counters.
  43  * \todo Add synchronization points.
  44  *
  45  * \author Artem Zhmurov <zhmurov@gmail.com>
  46  *
  47  * \inlibraryapi
  48  * \ingroup module_mdtypes
  49  */
  50 #ifndef GMX_MDTYPES_STATE_PROPAGATOR_DATA_GPU_H
  51 #define GMX_MDTYPES_STATE_PROPAGATOR_DATA_GPU_H
  52
  53 #include "gromacs/gpu_utils/devicebuffer_datatype.h"
  54 #include "gromacs/gpu_utils/gpu_utils.h"
  55 #include "gromacs/math/vectypes.h"
  56 #include "gromacs/utility/arrayref.h"
  57 #include "gromacs/utility/classhelpers.h"
  58
  59 namespace gmx
  60 {
  61
  62 class StatePropagatorDataGpu
  63 {
  64     public:
  65
  66         /*! \brief Atom locality indicator: local, non-local, all.
  67          *
  68          * \todo This should be managed by a separate object, since the localities
  69          *       are used here and in buffer ops.
  70          */
  71         enum class AtomLocality : int
  72         {
  73             Local    = 0, //!< Local atoms
  74             NonLocal = 1, //!< Non-local atoms
  75             All      = 2, //!< Both local and non-local atoms
  76             Count    = 3  //!< The number of atom locality types
  77         };
  78
  79         /*! \brief Constructor
  80          *
  81          * The buffers are reallocated only at the reinit call, the padding is
  82          * used there for the coordinates buffer. It is needed for PME and added at
  83          * the end of the buffer. It is assumed that if the rank has PME duties on the
  84          * GPU, all coordinates are copied to the GPU and hence, for this rank, the
  85          * coordinates buffer is not split into local and non-local ranges. For other
  86          * ranks, the padding size is zero. This works because only one rank ever does
  87          * PME work on the GPU, and if that rank also does PP work that is the only
  88          * rank. So all coordinates are always transferred.
  89          *
  90          * \note \p commandStream and \p gpuContext are allowed to be nullptr if
  91          *       StatePropagatorDataGpu is not used in the OpenCL run (e.g. if PME
  92          *       does not run on the GPU).
  93          *
  94          * \todo Make \p CommandStream visible in the CPU parts of the code so we
  95          *       will not have to pass a void*.
  96          * \todo Make \p Context visible in CPU parts of the code so we will not
  97          *       have to pass a void*.
  98          *
  99          *  \param[in] commandStream  GPU stream, nullptr allowed.
 100          *  \param[in] gpuContext     GPU context, nullptr allowed.
 101          *  \param[in] transferKind   H2D/D2H transfer call behavior (synchronous or not).
 102          *  \param[in] paddingSize    Padding size for coordinates buffer.
 103          */
 104         StatePropagatorDataGpu(const void        *commandStream,
 105                                const void        *gpuContext,
 106                                GpuApiCallBehavior transferKind,
 107                                int                paddingSize);
 108         //! Move constructor
 109         StatePropagatorDataGpu(StatePropagatorDataGpu &&other) noexcept;
 110         //! Move assignment
 111         StatePropagatorDataGpu &operator=(StatePropagatorDataGpu &&other) noexcept;
 112         //! Destructor
 113         ~StatePropagatorDataGpu();
 114
 115         /*! \brief Set the ranges for local and non-local atoms and reallocates buffers.
 116          *
 117          * The coordinates buffer is reallocated with the padding added at the end. The
 118          * size of padding is set by the constructor.
 119          *
 120          *  \param[in] numAtomsLocal  Number of atoms in local domain.
 121          *  \param[in] numAtomsAll    Total number of atoms to handle.
 122          */
 123         void reinit(int numAtomsLocal, int numAtomsAll);
 124
 125         /*! \brief Returns the range of atoms to be copied based on the copy type (all, local or non-local).
 126          *
 127          * \todo There are at least three versions of the function with this functionality in the code:
 128          *       this one and two more in NBNXM. These should be unified in a shape of a general function
 129          *       in DD.
 130          *
 131          * \param[in]  atomLocality    If all, local or non-local ranges are needed.
 132          *
 133          * \returns Tuple, containing the index of the first atom in the range and the total number of atoms in the range.
 134          */
 135         std::tuple<int, int> getAtomRangesFromAtomLocality(AtomLocality  atomLocality);
 136
 137
 138         /*! \brief Get the positions buffer on the GPU.
 139          *
 140          *  \returns GPU positions buffer.
 141          */
 142         DeviceBuffer<float> getCoordinates();
 143
 144         /*! \brief Copy positions to the GPU memory.
 145          *
 146          *  \param[in] h_x           Positions in the host memory.
 147          *  \param[in] atomLocality  Locality of the particles to copy.
 148          */
 149         void copyCoordinatesToGpu(gmx::ArrayRef<const gmx::RVec>  h_x,
 150                                   AtomLocality                    atomLocality);
 151
 152         /*! \brief Copy positions from the GPU memory.
 153          *
 154          *  \param[in] h_x           Positions buffer in the host memory.
 155          *  \param[in] atomLocality  Locality of the particles to copy.
 156          */
 157         void copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec>  h_x,
 158                                     AtomLocality              atomLocality);
 159
 160
 161         /*! \brief Get the velocities buffer on the GPU.
 162          *
 163          *  \returns GPU velocities buffer.
 164          */
 165         DeviceBuffer<float> getVelocities();
 166
 167         /*! \brief Copy velocities to the GPU memory.
 168          *
 169          *  \param[in] h_v           Velocities in the host memory.
 170          *  \param[in] atomLocality  Locality of the particles to copy.
 171          */
 172         void copyVelocitiesToGpu(gmx::ArrayRef<const gmx::RVec>  h_v,
 173                                  AtomLocality                    atomLocality);
 174
 175         /*! \brief Copy velocities from the GPU memory.
 176          *
 177          *  \param[in] h_v           Velocities buffer in the host memory.
 178          *  \param[in] atomLocality  Locality of the particles to copy.
 179          */
 180         void copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec>  h_v,
 181                                    AtomLocality              atomLocality);
 182
 183
 184         /*! \brief Get the force buffer on the GPU.
 185          *
 186          *  \returns GPU force buffer.
 187          */
 188         DeviceBuffer<float> getForces();
 189
 190         /*! \brief Copy forces to the GPU memory.
 191          *
 192          *  \param[in] h_f           Forces in the host memory.
 193          *  \param[in] atomLocality  Locality of the particles to copy.
 194          */
 195         void copyForcesToGpu(gmx::ArrayRef<const gmx::RVec>  h_f,
 196                              AtomLocality                    atomLocality);
 197
 198         /*! \brief Copy forces from the GPU memory.
 199          *
 200          *  \param[in] h_f           Forces buffer in the host memory.
 201          *  \param[in] atomLocality  Locality of the particles to copy.
 202          */
 203         void copyForcesFromGpu(gmx::ArrayRef<gmx::RVec>  h_f,
 204                                AtomLocality              atomLocality);
 205         /*! \brief Synchronize the underlying GPU stream
 206          */
 207         void synchronizeStream();
 208
 209         /*! \brief Getter for the number of local atoms.
 210          *
 211          *  \returns The number of local atoms.
 212          */
 213         int numAtomsLocal();
 214
 215         /*! \brief Getter for the total number of atoms.
 216          *
 217          *  \returns The total number of atoms.
 218          */
 219         int numAtomsAll();
 220
 221     private:
 222         class Impl;
 223         gmx::PrivateImplPointer<Impl> impl_;
 224         GMX_DISALLOW_COPY_AND_ASSIGN(StatePropagatorDataGpu);
 225 };
 226
 227 }      // namespace gmx
 228
 229 #endif // GMX_MDTYPES_STATE_PROPAGATOR_DATA_GPU_H