2 * This file is part of the GROMACS molecular simulation package.
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.
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.
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.
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.
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.
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.
37 * \brief Declaration of interfaces for GPU state data propagator object.
39 * This object stores and manages positions, velocities and forces for
40 * all particles in the system on the GPU.
42 * \todo Add cycle counters.
43 * \todo Add synchronization points.
45 * \author Artem Zhmurov <zhmurov@gmail.com>
48 * \ingroup module_mdtypes
50 #ifndef GMX_MDTYPES_STATE_PROPAGATOR_DATA_GPU_H
51 #define GMX_MDTYPES_STATE_PROPAGATOR_DATA_GPU_H
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"
59 class GpuEventSynchronizer;
64 class StatePropagatorDataGpu
68 /*! \brief Atom locality indicator: local, non-local, all.
70 * \todo This should be managed by a separate object, since the localities
71 * are used here and in buffer ops.
73 enum class AtomLocality : int
75 Local = 0, //!< Local atoms
76 NonLocal = 1, //!< Non-local atoms
77 All = 2, //!< Both local and non-local atoms
78 Count = 3 //!< The number of atom locality types
81 /*! \brief Constructor
83 * The buffers are reallocated only at the reinit call, the padding is
84 * used there for the coordinates buffer. It is needed for PME and added at
85 * the end of the buffer. It is assumed that if the rank has PME duties on the
86 * GPU, all coordinates are copied to the GPU and hence, for this rank, the
87 * coordinates buffer is not split into local and non-local ranges. For other
88 * ranks, the padding size is zero. This works because only one rank ever does
89 * PME work on the GPU, and if that rank also does PP work that is the only
90 * rank. So all coordinates are always transferred.
92 * In OpenCL, only pmeStream is used since it is the only stream created in
93 * PME context. The local and non-local streams are only needed when buffer
94 * ops are offloaded. This feature is currently not available in OpenCL and
95 * hence these streams are not set in these builds.
97 * \note In CUDA, the update stream is created in the constructor as a temporary
98 * solution, in place until the stream manager is introduced.
99 * Note that this makes it impossible to construct this object in CUDA
100 * builds executing on a host without any CUDA-capable device available.
102 * \note In CUDA, \p deviceContext is unused, hence always nullptr;
103 * all stream arguments can also be nullptr in runs where the
104 * respective streams are not required.
105 * In OpenCL, \p deviceContext needs to be a valid device context.
106 * In OpenCL runs StatePropagatorDataGpu is currently only used
107 * with PME offload, and only on ranks with PME duty. Hence, the
108 * \p pmeStream argument needs to be a valid OpenCL queue object
109 * which must have been created in \p deviceContext.
111 * \todo Make a \p CommandStream visible in the CPU parts of the code so we
112 * will not have to pass a void*.
113 * \todo Make a \p DeviceContext object visible in CPU parts of the code so we
114 * will not have to pass a void*.
116 * \param[in] pmeStream Device PME stream, nullptr allowed.
117 * \param[in] localStream Device NBNXM local stream, nullptr allowed.
118 * \param[in] nonLocalStream Device NBNXM non-local stream, nullptr allowed.
119 * \param[in] deviceContext Device context, nullptr allowed.
120 * \param[in] transferKind H2D/D2H transfer call behavior (synchronous or not).
121 * \param[in] paddingSize Padding size for coordinates buffer.
123 StatePropagatorDataGpu(const void *pmeStream,
124 const void *localStream,
125 const void *nonLocalStream,
126 const void *deviceContext,
127 GpuApiCallBehavior transferKind,
130 StatePropagatorDataGpu(StatePropagatorDataGpu &&other) noexcept;
132 StatePropagatorDataGpu &operator=(StatePropagatorDataGpu &&other) noexcept;
134 ~StatePropagatorDataGpu();
136 /*! \brief Set the ranges for local and non-local atoms and reallocates buffers.
139 * The coordinates buffer is (re)allocated, when required by PME, with a padding,
140 * the size of which is set by the constructor. The padding region clearing kernel
141 * is scheduled in the \p pmeStream_ (unlike the coordinates H2D) as only the PME
142 * task uses this padding area.
144 * \param[in] numAtomsLocal Number of atoms in local domain.
145 * \param[in] numAtomsAll Total number of atoms to handle.
147 void reinit(int numAtomsLocal, int numAtomsAll);
149 /*! \brief Returns the range of atoms to be copied based on the copy type (all, local or non-local).
151 * \todo There are at least three versions of the function with this functionality in the code:
152 * this one and two more in NBNXM. These should be unified in a shape of a general function
155 * \param[in] atomLocality If all, local or non-local ranges are needed.
157 * \returns Tuple, containing the index of the first atom in the range and the total number of atoms in the range.
159 std::tuple<int, int> getAtomRangesFromAtomLocality(AtomLocality atomLocality);
162 /*! \brief Get the positions buffer on the GPU.
164 * \returns GPU positions buffer.
166 DeviceBuffer<float> getCoordinates();
168 /*! \brief Copy positions to the GPU memory.
170 * \param[in] h_x Positions in the host memory.
171 * \param[in] atomLocality Locality of the particles to copy.
173 void copyCoordinatesToGpu(gmx::ArrayRef<const gmx::RVec> h_x,
174 AtomLocality atomLocality);
176 /*! \brief Get the event synchronizer on the H2D coordinates copy.
178 * \param[in] atomLocality Locality of the particles to wait for.
180 * \returns The event to synchronize the stream that consumes coordinates on device.
182 GpuEventSynchronizer* getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality);
184 /*! \brief Copy positions from the GPU memory.
186 * \param[in] h_x Positions buffer in the host memory.
187 * \param[in] atomLocality Locality of the particles to copy.
189 void copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x,
190 AtomLocality atomLocality);
192 /*! \brief Wait until coordinates are available on the host.
194 * \param[in] atomLocality Locality of the particles to wait for.
196 void waitCoordinatesReadyOnHost(AtomLocality atomLocality);
199 /*! \brief Get the velocities buffer on the GPU.
201 * \returns GPU velocities buffer.
203 DeviceBuffer<float> getVelocities();
205 /*! \brief Copy velocities to the GPU memory.
207 * \param[in] h_v Velocities in the host memory.
208 * \param[in] atomLocality Locality of the particles to copy.
210 void copyVelocitiesToGpu(gmx::ArrayRef<const gmx::RVec> h_v,
211 AtomLocality atomLocality);
213 /*! \brief Copy velocities from the GPU memory.
215 * \param[in] h_v Velocities buffer in the host memory.
216 * \param[in] atomLocality Locality of the particles to copy.
218 void copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v,
219 AtomLocality atomLocality);
222 /*! \brief Get the force buffer on the GPU.
224 * \returns GPU force buffer.
226 DeviceBuffer<float> getForces();
228 /*! \brief Copy forces to the GPU memory.
230 * \param[in] h_f Forces in the host memory.
231 * \param[in] atomLocality Locality of the particles to copy.
233 void copyForcesToGpu(gmx::ArrayRef<const gmx::RVec> h_f,
234 AtomLocality atomLocality);
236 /*! \brief Copy forces from the GPU memory.
238 * \param[in] h_f Forces buffer in the host memory.
239 * \param[in] atomLocality Locality of the particles to copy.
241 void copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f,
242 AtomLocality atomLocality);
244 /*! \brief Getter for the update stream.
246 * \todo This is temporary here, until the management of this stream is taken over.
248 * \returns The device command stream to use in update-constraints.
250 void* getUpdateStream();
252 /*! \brief Getter for the number of local atoms.
254 * \returns The number of local atoms.
258 /*! \brief Getter for the total number of atoms.
260 * \returns The total number of atoms.
266 gmx::PrivateImplPointer<Impl> impl_;
267 GMX_DISALLOW_COPY_AND_ASSIGN(StatePropagatorDataGpu);
272 #endif // GMX_MDTYPES_STATE_PROPAGATOR_DATA_GPU_H