2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2019,2020,2021, 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
56 #include "gromacs/gpu_utils/devicebuffer_datatype.h"
57 #include "gromacs/gpu_utils/gpu_utils.h"
58 #include "gromacs/math/vectypes.h"
59 #include "gromacs/mdtypes/simulation_workload.h"
60 #include "gromacs/utility/arrayref.h"
61 #include "gromacs/utility/classhelpers.h"
67 class GpuEventSynchronizer;
72 class DeviceStreamManager;
74 class StatePropagatorDataGpu
77 /*! \brief Constructor
79 * The buffers are reallocated only at the reinit call, the padding is
80 * used there for the coordinates buffer. It is needed for PME and added at
81 * the end of the buffer. It is assumed that if the rank has PME duties on the
82 * GPU, all coordinates are copied to the GPU and hence, for this rank, the
83 * coordinates buffer is not split into local and non-local ranges. For other
84 * ranks, the padding size is zero. This works because only one rank ever does
85 * PME work on the GPU, and if that rank also does PP work that is the only
86 * rank. So all coordinates are always transferred.
88 * In OpenCL, only pmeStream is used since it is the only stream created in
89 * PME context. The local and non-local streams are only needed when buffer
90 * ops are offloaded. This feature is currently not available in OpenCL and
91 * hence these streams are not set in these builds.
93 * \param[in] deviceStreamManager Object that owns the DeviceContext and DeviceStreams.
94 * \param[in] transferKind H2D/D2H transfer call behavior (synchronous or not).
95 * \param[in] allocationBlockSizeDivisor Deterines padding size for coordinates buffer.
96 * \param[in] wcycle Wall cycle counter data.
98 StatePropagatorDataGpu(const DeviceStreamManager& deviceStreamManager,
99 GpuApiCallBehavior transferKind,
100 int allocationBlockSizeDivisor,
101 gmx_wallcycle* wcycle);
103 /*! \brief Constructor to use in PME-only rank and in tests.
105 * This constructor should be used if only a coordinate device buffer should be managed
106 * using a single stream. Any operation on force or velocity buffer as well as copy of
107 * non-local coordinates will exit with assertion failure. Note, that the pmeStream can
108 * not be a nullptr and the constructor will exit with an assertion failure.
110 * \todo Currently, unsupported copy operations are blocked by assertion that the stream
111 * not nullptr. This should be improved.
113 * \param[in] pmeStream Device PME stream, nullptr is not allowed.
114 * \param[in] deviceContext Device context, nullptr allowed for non-OpenCL builds.
115 * \param[in] transferKind H2D/D2H transfer call behavior (synchronous or not).
116 * \param[in] allocationBlockSizeDivisor Determines padding size for coordinates buffer.
117 * \param[in] wcycle Wall cycle counter data.
119 StatePropagatorDataGpu(const DeviceStream* pmeStream,
120 const DeviceContext& deviceContext,
121 GpuApiCallBehavior transferKind,
122 int allocationBlockSizeDivisor,
123 gmx_wallcycle* wcycle);
126 StatePropagatorDataGpu(StatePropagatorDataGpu&& other) noexcept;
128 StatePropagatorDataGpu& operator=(StatePropagatorDataGpu&& other) noexcept;
130 ~StatePropagatorDataGpu();
132 /*! \brief Set the ranges for local and non-local atoms and reallocates buffers.
134 * Reallocates coordinate, velocities and force buffers on the device.
137 * The coordinates buffer is (re)allocated, when required by PME, with a padding,
138 * the size of which is set by the constructor. The padding region clearing kernel
139 * is scheduled in the \p pmeStream_ (unlike the coordinates H2D) as only the PME
140 * task uses this padding area.
143 * The force buffer is cleared if its size increases, so that previously unused
144 * memory is cleared before forces are accumulated.
146 * \param[in] numAtomsLocal Number of atoms in local domain.
147 * \param[in] numAtomsAll Total number of atoms to handle.
149 void reinit(int numAtomsLocal, int numAtomsAll);
151 /*! \brief Returns the range of atoms to be copied based on the copy type (all, local or non-local).
153 * \todo There are at least three versions of the function with this functionality in the code:
154 * this one and two more in NBNXM. These should be unified in a shape of a general function
157 * \param[in] atomLocality If all, local or non-local ranges are needed.
159 * \returns Tuple, containing the index of the first atom in the range and the total number of atoms in the range.
161 std::tuple<int, int> getAtomRangesFromAtomLocality(AtomLocality atomLocality) const;
164 /*! \brief Get the positions buffer on the GPU.
166 * \returns GPU positions buffer.
168 DeviceBuffer<RVec> getCoordinates();
170 /*! \brief Copy positions to the GPU memory.
172 * Use \ref getCoordinatesReadyOnDeviceEvent to get the associated event synchronizer or
173 * \ref waitCoordinatesCopiedToDevice to wait for the copy completion.
174 * Note: the event is not marked in OpenCL, because it is not used.
176 * \param[in] h_x Positions in the host memory.
177 * \param[in] atomLocality Locality of the particles to copy.
179 void copyCoordinatesToGpu(gmx::ArrayRef<const gmx::RVec> h_x, AtomLocality atomLocality);
181 /*! \brief Get the event synchronizer of the coordinates ready for the consumption on the device.
183 * Returns the event synchronizer which indicates that the coordinates are ready for the
184 * consumption on the device. Takes into account that the producer may be different.
186 * If the update is offloaded, and the current step is not a DD/search step, the returned
187 * synchronizer indicates the completion of GPU update-constraint kernels. Otherwise, on search
188 * steps and if update is not offloaded, the coordinates are provided by the H2D copy and the
189 * returned synchronizer indicates that the copy is complete.
191 * \param[in] atomLocality Locality of the particles to wait for.
192 * \param[in] simulationWork The simulation lifetime flags.
193 * \param[in] stepWork The step lifetime flags.
194 * \param[in] gpuCoordinateHaloLaunched Event recorded when GPU coordinate halo has been launched.
196 * \returns The event to synchronize the stream that consumes coordinates on device.
198 GpuEventSynchronizer* getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
199 const SimulationWorkload& simulationWork,
200 const StepWorkload& stepWork,
201 GpuEventSynchronizer* gpuCoordinateHaloLaunched = nullptr);
203 /*! \brief Blocking wait until coordinates are copied to the device.
205 * Synchronizes the stream in which the copy was executed.
207 * \param[in] atomLocality Locality of the particles to wait for.
209 void waitCoordinatesCopiedToDevice(AtomLocality atomLocality);
211 /*! \brief Setter for the event synchronizer for the update is done on th GPU
213 * \param[in] xUpdatedOnDeviceEvent The event to synchronize the stream coordinates wre updated on device.
215 void setXUpdatedOnDeviceEvent(GpuEventSynchronizer* xUpdatedOnDeviceEvent);
217 /*! \brief Copy positions from the GPU memory, with an optional explicit dependency.
219 * \param[in] h_x Positions buffer in the host memory.
220 * \param[in] atomLocality Locality of the particles to copy.
221 * \param[in] dependency Dependency event for this operation.
223 void copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x,
224 AtomLocality atomLocality,
225 GpuEventSynchronizer* dependency = nullptr);
227 /*! \brief Wait until coordinates are available on the host.
229 * \param[in] atomLocality Locality of the particles to wait for.
231 void waitCoordinatesReadyOnHost(AtomLocality atomLocality);
234 /*! \brief Get the velocities buffer on the GPU.
236 * \returns GPU velocities buffer.
238 DeviceBuffer<RVec> getVelocities();
240 /*! \brief Copy velocities to the GPU memory.
242 * Does not mark any event, because we don't use it anywhere at the moment.
244 * \param[in] h_v Velocities in the host memory.
245 * \param[in] atomLocality Locality of the particles to copy.
247 void copyVelocitiesToGpu(gmx::ArrayRef<const gmx::RVec> h_v, AtomLocality atomLocality);
249 /*! \brief Copy velocities from the GPU memory.
251 * \param[in] h_v Velocities buffer in the host memory.
252 * \param[in] atomLocality Locality of the particles to copy.
254 void copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v, AtomLocality atomLocality);
256 /*! \brief Wait until velocities are available on the host.
258 * \param[in] atomLocality Locality of the particles to wait for.
260 void waitVelocitiesReadyOnHost(AtomLocality atomLocality);
263 /*! \brief Get the force buffer on the GPU.
265 * \returns GPU force buffer.
267 DeviceBuffer<RVec> getForces();
269 /*! \brief Copy forces to the GPU memory.
271 * \param[in] h_f Forces in the host memory.
272 * \param[in] atomLocality Locality of the particles to copy.
274 void copyForcesToGpu(gmx::ArrayRef<const gmx::RVec> h_f, AtomLocality atomLocality);
276 /*! \brief Clear forces in the GPU memory.
278 * \param[in] atomLocality Locality of the particles to clear.
279 * \param[in] dependency Dependency event for this operation.
281 void clearForcesOnGpu(AtomLocality atomLocality, GpuEventSynchronizer* dependency);
283 /*! \brief Get the event synchronizer for the forces ready on device.
285 * Returns either of the event synchronizers, depending on the offload scenario
286 * for the current simulation timestep:
287 * 1. The forces are copied to the device (when GPU buffer ops are off)
288 * 2. The forces are reduced on the device (GPU buffer ops are on)
290 * \param[in] stepWork Step workload flags
291 * \param[in] simulationWork Simulation workload flags
293 * \returns The event to synchronize the stream that consumes forces on device.
295 GpuEventSynchronizer* getLocalForcesReadyOnDeviceEvent(StepWorkload stepWork,
296 SimulationWorkload simulationWork);
298 /*! \brief Getter for the event synchronizer for the forces are reduced on the GPU.
300 * \param[in] atomLocality Locality of the particles to wait for.
301 * \returns The event to mark when forces are reduced on the GPU.
303 GpuEventSynchronizer* fReducedOnDevice(AtomLocality atomLocality);
305 /*! \brief Getter for the event synchronizer for the forces are ready on the GPU.
307 * \param[in] atomLocality Locality of the particles to wait for.
308 * \returns The event to mark when forces are ready on the GPU.
310 GpuEventSynchronizer* fReadyOnDevice(AtomLocality atomLocality);
312 /*! \brief Copy forces from the GPU memory.
314 * \param[in] h_f Forces buffer in the host memory.
315 * \param[in] atomLocality Locality of the particles to copy.
317 void copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f, AtomLocality atomLocality);
319 /*! \brief Wait until forces are available on the host.
321 * \param[in] atomLocality Locality of the particles to wait for.
323 void waitForcesReadyOnHost(AtomLocality atomLocality);
325 /*! \brief Getter for the update stream.
327 * \todo This is temporary here, until the management of this stream is taken over.
329 * \returns The device command stream to use in update-constraints.
331 const DeviceStream* getUpdateStream();
333 /*! \brief Getter for the number of local atoms.
335 * \returns The number of local atoms.
337 int numAtomsLocal() const;
339 /*! \brief Getter for the total number of atoms.
341 * \returns The total number of atoms.
343 int numAtomsAll() const;
347 std::unique_ptr<Impl> impl_;
348 GMX_DISALLOW_COPY_AND_ASSIGN(StatePropagatorDataGpu);
353 #endif // GMX_MDTYPES_STATE_PROPAGATOR_DATA_GPU_H