2 * This file is part of the GROMACS molecular simulation package.
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.
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/mdtypes/simulation_workload.h"
57 #include "gromacs/utility/arrayref.h"
58 #include "gromacs/utility/classhelpers.h"
62 class GpuEventSynchronizer;
68 class StatePropagatorDataGpu
71 /*! \brief Constructor
73 * The buffers are reallocated only at the reinit call, the padding is
74 * used there for the coordinates buffer. It is needed for PME and added at
75 * the end of the buffer. It is assumed that if the rank has PME duties on the
76 * GPU, all coordinates are copied to the GPU and hence, for this rank, the
77 * coordinates buffer is not split into local and non-local ranges. For other
78 * ranks, the padding size is zero. This works because only one rank ever does
79 * PME work on the GPU, and if that rank also does PP work that is the only
80 * rank. So all coordinates are always transferred.
82 * In OpenCL, only pmeStream is used since it is the only stream created in
83 * PME context. The local and non-local streams are only needed when buffer
84 * ops are offloaded. This feature is currently not available in OpenCL and
85 * hence these streams are not set in these builds.
87 * \note In CUDA, the update stream is created in the constructor as a temporary
88 * solution, in place until the stream manager is introduced.
89 * Note that this makes it impossible to construct this object in CUDA
90 * builds executing on a host without any CUDA-capable device available.
92 * \note In CUDA, \p deviceContext is unused, hence always nullptr;
93 * all stream arguments can also be nullptr in runs where the
94 * respective streams are not required.
95 * In OpenCL, \p deviceContext needs to be a valid device context.
96 * In OpenCL runs StatePropagatorDataGpu is currently only used
97 * with PME offload, and only on ranks with PME duty. Hence, the
98 * \p pmeStream argument needs to be a valid OpenCL queue object
99 * which must have been created in \p deviceContext.
101 * \todo Make a \p CommandStream visible in the CPU parts of the code so we
102 * will not have to pass a void*.
103 * \todo Make a \p DeviceContext object visible in CPU parts of the code so we
104 * will not have to pass a void*.
106 * \param[in] pmeStream Device PME stream, nullptr allowed.
107 * \param[in] localStream Device NBNXM local stream, nullptr allowed.
108 * \param[in] nonLocalStream Device NBNXM non-local stream, nullptr allowed.
109 * \param[in] deviceContext Device context, nullptr allowed.
110 * \param[in] transferKind H2D/D2H transfer call behavior (synchronous or not).
111 * \param[in] paddingSize Padding size for coordinates buffer.
112 * \param[in] wcycle Wall cycle counter data.
114 StatePropagatorDataGpu(const void* pmeStream,
115 const void* localStream,
116 const void* nonLocalStream,
117 const void* deviceContext,
118 GpuApiCallBehavior transferKind,
120 gmx_wallcycle* wcycle);
122 /*! \brief Constructor to use in PME-only rank and in tests.
124 * This constructor should be used if only a coordinate device buffer should be managed
125 * using a single stream. Any operation on force or velocity buffer as well as copy of
126 * non-local coordinates will exit with assertion failure. Note, that the pmeStream can
127 * not be a nullptr and the constructor will exit with an assertion failure.
129 * \todo Currently, unsupported copy operations are blocked by assertion that the stream
130 * not nullptr. This should be improved.
132 * \param[in] pmeStream Device PME stream, nullptr is not allowed.
133 * \param[in] deviceContext Device context, nullptr allowed for non-OpenCL builds.
134 * \param[in] transferKind H2D/D2H transfer call behavior (synchronous or not).
135 * \param[in] paddingSize Padding size for coordinates buffer.
136 * \param[in] wcycle Wall cycle counter data.
138 StatePropagatorDataGpu(const void* pmeStream,
139 const void* deviceContext,
140 GpuApiCallBehavior transferKind,
142 gmx_wallcycle* wcycle);
145 StatePropagatorDataGpu(StatePropagatorDataGpu&& other) noexcept;
147 StatePropagatorDataGpu& operator=(StatePropagatorDataGpu&& other) noexcept;
149 ~StatePropagatorDataGpu();
151 /*! \brief Set the ranges for local and non-local atoms and reallocates buffers.
153 * Reallocates coordinate, velocities and force buffers on the device.
156 * The coordinates buffer is (re)allocated, when required by PME, with a padding,
157 * the size of which is set by the constructor. The padding region clearing kernel
158 * is scheduled in the \p pmeStream_ (unlike the coordinates H2D) as only the PME
159 * task uses this padding area.
162 * The force buffer is cleared if its size increases, so that previously unused
163 * memory is cleared before forces are accumulated.
165 * \param[in] numAtomsLocal Number of atoms in local domain.
166 * \param[in] numAtomsAll Total number of atoms to handle.
168 void reinit(int numAtomsLocal, int numAtomsAll);
170 /*! \brief Returns the range of atoms to be copied based on the copy type (all, local or non-local).
172 * \todo There are at least three versions of the function with this functionality in the code:
173 * this one and two more in NBNXM. These should be unified in a shape of a general function
176 * \param[in] atomLocality If all, local or non-local ranges are needed.
178 * \returns Tuple, containing the index of the first atom in the range and the total number of atoms in the range.
180 std::tuple<int, int> getAtomRangesFromAtomLocality(AtomLocality atomLocality);
183 /*! \brief Get the positions buffer on the GPU.
185 * \returns GPU positions buffer.
187 DeviceBuffer<RVec> getCoordinates();
189 /*! \brief Copy positions to the GPU memory.
191 * \param[in] h_x Positions in the host memory.
192 * \param[in] atomLocality Locality of the particles to copy.
194 void copyCoordinatesToGpu(gmx::ArrayRef<const gmx::RVec> h_x, AtomLocality atomLocality);
196 /*! \brief Get the event synchronizer of the coordinates ready for the consumption on the device.
198 * Returns the event synchronizer which indicates that the coordinates are ready for the
199 * consumption on the device. Takes into account that the producer may be different.
201 * If the update is offloaded, and the current step is not a DD/search step, the returned
202 * synchronizer indicates the completion of GPU update-constraint kernels. Otherwise, on search
203 * steps and if update is not offloaded, the coordinates are provided by the H2D copy and the
204 * returned synchronizer indicates that the copy is complete.
206 * \param[in] atomLocality Locality of the particles to wait for.
207 * \param[in] simulationWork The simulation lifetime flags.
208 * \param[in] stepWork The step lifetime flags.
210 * \returns The event to synchronize the stream that consumes coordinates on device.
212 GpuEventSynchronizer* getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
213 const SimulationWorkload& simulationWork,
214 const StepWorkload& stepWork);
216 /*! \brief Blocking wait until coordinates are copied to the device.
218 * Synchronizes the stream in which the copy was executed.
220 * \param[in] atomLocality Locality of the particles to wait for.
222 void waitCoordinatesCopiedToDevice(AtomLocality atomLocality);
224 /*! \brief Getter for the event synchronizer for the update is done on th GPU
226 * \returns The event to synchronize the stream coordinates wre updated on device.
228 GpuEventSynchronizer* xUpdatedOnDevice();
230 /*! \brief Copy positions from the GPU memory.
232 * \param[in] h_x Positions buffer in the host memory.
233 * \param[in] atomLocality Locality of the particles to copy.
235 void copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x, AtomLocality atomLocality);
237 /*! \brief Wait until coordinates are available on the host.
239 * \param[in] atomLocality Locality of the particles to wait for.
241 void waitCoordinatesReadyOnHost(AtomLocality atomLocality);
244 /*! \brief Get the velocities buffer on the GPU.
246 * \returns GPU velocities buffer.
248 DeviceBuffer<RVec> getVelocities();
250 /*! \brief Copy velocities to the GPU memory.
252 * \param[in] h_v Velocities in the host memory.
253 * \param[in] atomLocality Locality of the particles to copy.
255 void copyVelocitiesToGpu(gmx::ArrayRef<const gmx::RVec> h_v, AtomLocality atomLocality);
257 /*! \brief Get the event synchronizer for the H2D velocities copy.
259 * \param[in] atomLocality Locality of the particles to wait for.
261 * \returns The event to synchronize the stream that consumes velocities on device.
263 GpuEventSynchronizer* getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality);
265 /*! \brief Copy velocities from the GPU memory.
267 * \param[in] h_v Velocities buffer in the host memory.
268 * \param[in] atomLocality Locality of the particles to copy.
270 void copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v, AtomLocality atomLocality);
272 /*! \brief Wait until velocities are available on the host.
274 * \param[in] atomLocality Locality of the particles to wait for.
276 void waitVelocitiesReadyOnHost(AtomLocality atomLocality);
279 /*! \brief Get the force buffer on the GPU.
281 * \returns GPU force buffer.
283 DeviceBuffer<RVec> getForces();
285 /*! \brief Copy forces to the GPU memory.
287 * \param[in] h_f Forces in the host memory.
288 * \param[in] atomLocality Locality of the particles to copy.
290 void copyForcesToGpu(gmx::ArrayRef<const gmx::RVec> h_f, AtomLocality atomLocality);
292 /*! \brief Get the event synchronizer for the forces ready on device.
294 * Returns either of the event synchronizers, depending on the offload scenario
295 * for the current simulation timestep:
296 * 1. The forces are copied to the device (when GPU buffer ops are off)
297 * 2. The forces are reduced on the device (GPU buffer ops are on)
299 * \todo Pass step workload instead of the useGpuFBufferOps boolean.
301 * \param[in] atomLocality Locality of the particles to wait for.
302 * \param[in] useGpuFBufferOps If the force buffer ops are offloaded to the GPU.
304 * \returns The event to synchronize the stream that consumes forces on device.
306 GpuEventSynchronizer* getForcesReadyOnDeviceEvent(AtomLocality atomLocality, bool useGpuFBufferOps);
308 /*! \brief Getter for the event synchronizer for the forces are reduced on the GPU.
310 * \returns The event to mark when forces are reduced on the GPU.
312 GpuEventSynchronizer* fReducedOnDevice();
314 /*! \brief Copy forces from the GPU memory.
316 * \param[in] h_f Forces buffer in the host memory.
317 * \param[in] atomLocality Locality of the particles to copy.
319 void copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f, AtomLocality atomLocality);
321 /*! \brief Wait until forces are available on the host.
323 * \param[in] atomLocality Locality of the particles to wait for.
325 void waitForcesReadyOnHost(AtomLocality atomLocality);
327 /*! \brief Getter for the update stream.
329 * \todo This is temporary here, until the management of this stream is taken over.
331 * \returns The device command stream to use in update-constraints.
333 void* getUpdateStream();
335 /*! \brief Getter for the number of local atoms.
337 * \returns The number of local atoms.
341 /*! \brief Getter for the total number of atoms.
343 * \returns The total number of atoms.
349 gmx::PrivateImplPointer<Impl> impl_;
350 GMX_DISALLOW_COPY_AND_ASSIGN(StatePropagatorDataGpu);
355 #endif // GMX_MDTYPES_STATE_PROPAGATOR_DATA_GPU_H