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 low-level functions and fields of GPU state propagator object.
39 * \author Artem Zhmurov <zhmurov@gmail.com>
41 * \ingroup module_mdtypes
43 #ifndef GMX_MDTYPES_STATE_PROPAGATOR_DATA_GPU_IMPL_H
44 #define GMX_MDTYPES_STATE_PROPAGATOR_DATA_GPU_IMPL_H
50 #include "gromacs/gpu_utils/devicebuffer.h"
51 #if GMX_GPU == GMX_GPU_CUDA
52 #include "gromacs/gpu_utils/gpueventsynchronizer.cuh"
53 #elif GMX_GPU == GMX_GPU_OPENCL
54 #include "gromacs/gpu_utils/gpueventsynchronizer_ocl.h"
56 #include "gromacs/math/vectypes.h"
57 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
58 #include "gromacs/utility/classhelpers.h"
59 #include "gromacs/utility/enumerationhelpers.h"
64 class StatePropagatorDataGpu::Impl
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.
113 Impl(const void *pmeStream,
114 const void *localStream,
115 const void *nonLocalStream,
116 const void *deviceContext,
117 GpuApiCallBehavior transferKind,
123 /*! \brief Set the ranges for local and non-local atoms and reallocates buffers.
126 * The coordinates buffer is (re)allocated, when required by PME, with a padding,
127 * the size of which is set by the constructor. The padding region clearing kernel
128 * is scheduled in the \p pmeStream_ (unlike the coordinates H2D) as only the PME
129 * task uses this padding area.
131 * \param[in] numAtomsLocal Number of atoms in local domain.
132 * \param[in] numAtomsAll Total number of atoms to handle.
134 void reinit(int numAtomsLocal, int numAtomsAll);
136 /*! \brief Returns the range of atoms to be copied based on the copy type (all, local or non-local).
138 * \todo There are at least three versions of the function with this functionality in the code:
139 * this one and two more in NBNXM. These should be unified in a shape of a general function
142 * \param[in] atomLocality If all, local or non-local ranges are needed.
144 * \returns Tuple, containing the index of the first atom in the range and the total number of atoms in the range.
146 std::tuple<int, int> getAtomRangesFromAtomLocality(AtomLocality atomLocality);
149 /*! \brief Get the positions buffer on the GPU.
151 * \returns GPU positions buffer.
153 DeviceBuffer<float> getCoordinates();
155 /*! \brief Copy positions to the GPU memory.
157 * \param[in] h_x Positions in the host memory.
158 * \param[in] atomLocality Locality of the particles to copy.
160 void copyCoordinatesToGpu(gmx::ArrayRef<const gmx::RVec> h_x,
161 AtomLocality atomLocality);
163 /*! \brief Get the event synchronizer on the H2D coordinates copy.
165 * \param[in] atomLocality Locality of the particles to wait for.
167 * \returns The event to synchronize the stream that consumes coordinates on device.
169 GpuEventSynchronizer* getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality);
171 /*! \brief Copy positions from the GPU memory.
173 * \param[in] h_x Positions buffer in the host memory.
174 * \param[in] atomLocality Locality of the particles to copy.
176 void copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x,
177 AtomLocality atomLocality);
179 /*! \brief Wait until coordinates are available on the host.
181 * \param[in] atomLocality Locality of the particles to wait for.
183 void waitCoordinatesReadyOnHost(AtomLocality atomLocality);
186 /*! \brief Get the velocities buffer on the GPU.
188 * \returns GPU velocities buffer.
190 DeviceBuffer<float> getVelocities();
192 /*! \brief Copy velocities to the GPU memory.
194 * \param[in] h_v Velocities in the host memory.
195 * \param[in] atomLocality Locality of the particles to copy.
197 void copyVelocitiesToGpu(gmx::ArrayRef<const gmx::RVec> h_v,
198 AtomLocality atomLocality);
200 /*! \brief Copy velocities from the GPU memory.
202 * \param[in] h_v Velocities buffer in the host memory.
203 * \param[in] atomLocality Locality of the particles to copy.
205 void copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v,
206 AtomLocality atomLocality);
209 /*! \brief Get the force buffer on the GPU.
211 * \returns GPU force buffer.
213 DeviceBuffer<float> getForces();
215 /*! \brief Copy forces to the GPU memory.
217 * \param[in] h_f Forces in the host memory.
218 * \param[in] atomLocality Locality of the particles to copy.
220 void copyForcesToGpu(gmx::ArrayRef<const gmx::RVec> h_f,
221 AtomLocality atomLocality);
223 /*! \brief Copy forces from the GPU memory.
225 * \param[in] h_f Forces buffer in the host memory.
226 * \param[in] atomLocality Locality of the particles to copy.
228 void copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f,
229 AtomLocality atomLocality);
231 /*! \brief Getter for the update stream.
233 * \todo This is temporary here, until the management of this stream is taken over.
235 * \returns The device command stream to use in update-constraints.
237 void* getUpdateStream();
239 /*! \brief Getter for the number of local atoms.
241 * \returns The number of local atoms.
245 /*! \brief Getter for the total number of atoms.
247 * \returns The total number of atoms.
254 CommandStream pmeStream_ = nullptr;
255 //! GPU NBNXM local stream.
256 CommandStream localStream_ = nullptr;
257 //! GPU NBNXM non-local stream
258 CommandStream nonLocalStream_ = nullptr;
259 //! GPU Update-constreaints stream.
260 CommandStream updateStream_ = nullptr;
262 // Streams to use for coordinates H2S and D2H copies (one event for each atom locality)
263 EnumerationArray<AtomLocality, CommandStream> xCopyStreams_ = {{nullptr}};
265 //! An array of events that indicate H2D copy is complete (one event for each atom locality)
266 EnumerationArray<AtomLocality, GpuEventSynchronizer> xReadyOnDevice_;
267 //! An array of events that indicate D2H copy is complete (one event for each atom locality)
268 EnumerationArray<AtomLocality, GpuEventSynchronizer> xReadyOnHost_;
270 /*! \brief GPU context (for OpenCL builds)
271 * \todo Make a Context class usable in CPU code
273 DeviceContext deviceContext_ = nullptr;
274 //! Default GPU calls behavior
275 GpuApiCallBehavior transferKind_ = GpuApiCallBehavior::Async;
276 //! Padding size for the coordinates buffer
277 int paddingSize_ = 0;
279 //! Number of local atoms
280 int numAtomsLocal_ = -1;
281 //! Total number of atoms
282 int numAtomsAll_ = -1;
284 //! Device positions buffer
285 DeviceBuffer<float> d_x_;
286 //! Number of particles saved in the positions buffer
288 //! Allocation size for the positions buffer
289 int d_xCapacity_ = -1;
291 //! Device velocities buffer
292 DeviceBuffer<float> d_v_;
293 //! Number of particles saved in the velocities buffer
295 //! Allocation size for the velocities buffer
296 int d_vCapacity_ = -1;
298 //! Device force buffer
299 DeviceBuffer<float> d_f_;
300 //! Number of particles saved in the force buffer
302 //! Allocation size for the force buffer
303 int d_fCapacity_ = -1;
305 /*! \brief Performs the copy of data from host to device buffer.
307 * \todo Template on locality.
309 * \param[out] d_data Device-side buffer.
310 * \param[in] h_data Host-side buffer.
311 * \param[in] dataSize Device-side data allocation size.
312 * \param[in] atomLocality If all, local or non-local ranges should be copied.
313 * \param[in] commandStream GPU stream to execute copy in.
315 void copyToDevice(DeviceBuffer<float> d_data,
316 gmx::ArrayRef<const gmx::RVec> h_data,
318 AtomLocality atomLocality,
319 CommandStream commandStream);
321 /*! \brief Performs the copy of data from device to host buffer.
323 * \param[out] h_data Host-side buffer.
324 * \param[in] d_data Device-side buffer.
325 * \param[in] dataSize Device-side data allocation size.
326 * \param[in] atomLocality If all, local or non-local ranges should be copied.
327 * \param[in] commandStream GPU stream to execute copy in.
329 void copyFromDevice(gmx::ArrayRef<gmx::RVec> h_data,
330 DeviceBuffer<float> d_data,
332 AtomLocality atomLocality,
333 CommandStream commandStream);
338 #endif // GMX_MDTYPES_STATE_PROPAGATOR_DATA_GPU_IMPL_H