2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2018,2019,2020, 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.
35 #ifndef GMX_GPU_UTILS_DEVICEBUFFER_CUH
36 #define GMX_GPU_UTILS_DEVICEBUFFER_CUH
38 /*! \libinternal \file
39 * \brief Implements the DeviceBuffer type and routines for CUDA.
40 * Should only be included directly by the main DeviceBuffer file devicebuffer.h.
41 * TODO: the intent is for DeviceBuffer to become a class.
43 * \author Aleksei Iupinov <a.yupinov@gmail.com>
48 #include "gromacs/gpu_utils/device_context.h"
49 #include "gromacs/gpu_utils/devicebuffer_datatype.h"
50 #include "gromacs/gpu_utils/gpu_utils.h" //only for GpuApiCallBehavior
51 #include "gromacs/gpu_utils/gputraits.cuh"
52 #include "gromacs/utility/gmxassert.h"
55 * Allocates a device-side buffer.
56 * It is currently a caller's responsibility to call it only on not-yet allocated buffers.
58 * \tparam ValueType Raw value type of the \p buffer.
59 * \param[in,out] buffer Pointer to the device-side buffer.
60 * \param[in] numValues Number of values to accomodate.
61 * \param[in] deviceContext The buffer's dummy device context - not managed explicitly in CUDA RT.
63 template<typename ValueType>
64 void allocateDeviceBuffer(DeviceBuffer<ValueType>* buffer, size_t numValues, const DeviceContext& /* deviceContext */)
66 GMX_ASSERT(buffer, "needs a buffer pointer");
67 cudaError_t stat = cudaMalloc((void**)buffer, numValues * sizeof(ValueType));
68 GMX_RELEASE_ASSERT(stat == cudaSuccess, "cudaMalloc failure");
72 * Frees a device-side buffer.
73 * This does not reset separately stored size/capacity integers,
74 * as this is planned to be a destructor of DeviceBuffer as a proper class,
75 * and no calls on \p buffer should be made afterwards.
77 * \param[in] buffer Pointer to the buffer to free.
79 template<typename DeviceBuffer>
80 void freeDeviceBuffer(DeviceBuffer* buffer)
82 GMX_ASSERT(buffer, "needs a buffer pointer");
85 GMX_RELEASE_ASSERT(cudaFree(*buffer) == cudaSuccess, "cudaFree failed");
90 * Performs the host-to-device data copy, synchronous or asynchronously on request.
92 * TODO: This is meant to gradually replace cu/ocl_copy_h2d.
94 * \tparam ValueType Raw value type of the \p buffer.
95 * \param[in,out] buffer Pointer to the device-side buffer
96 * \param[in] hostBuffer Pointer to the raw host-side memory, also typed \p ValueType
97 * \param[in] startingOffset Offset (in values) at the device-side buffer to copy into.
98 * \param[in] numValues Number of values to copy.
99 * \param[in] deviceStream GPU stream to perform asynchronous copy in.
100 * \param[in] transferKind Copy type: synchronous or asynchronous.
101 * \param[out] timingEvent A dummy pointer to the H2D copy timing event to be filled in.
102 * Not used in CUDA implementation.
104 template<typename ValueType>
105 void copyToDeviceBuffer(DeviceBuffer<ValueType>* buffer,
106 const ValueType* hostBuffer,
107 size_t startingOffset,
109 const DeviceStream& deviceStream,
110 GpuApiCallBehavior transferKind,
111 CommandEvent* /*timingEvent*/)
115 return; // such calls are actually made with empty domains
117 GMX_ASSERT(buffer, "needs a buffer pointer");
118 GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
120 const size_t bytes = numValues * sizeof(ValueType);
122 switch (transferKind)
124 case GpuApiCallBehavior::Async:
125 GMX_ASSERT(isHostMemoryPinned(hostBuffer),
126 "Source host buffer was not pinned for CUDA");
127 stat = cudaMemcpyAsync(*((ValueType**)buffer) + startingOffset, hostBuffer, bytes,
128 cudaMemcpyHostToDevice, deviceStream.stream());
129 GMX_RELEASE_ASSERT(stat == cudaSuccess, "Asynchronous H2D copy failed");
132 case GpuApiCallBehavior::Sync:
133 stat = cudaMemcpy(*((ValueType**)buffer) + startingOffset, hostBuffer, bytes,
134 cudaMemcpyHostToDevice);
135 GMX_RELEASE_ASSERT(stat == cudaSuccess, "Synchronous H2D copy failed");
144 * Performs the device-to-host data copy, synchronous or asynchronously on request.
146 * TODO: This is meant to gradually replace cu/ocl_copy_d2h.
148 * \tparam ValueType Raw value type of the \p buffer.
149 * \param[in,out] hostBuffer Pointer to the raw host-side memory, also typed \p ValueType
150 * \param[in] buffer Pointer to the device-side buffer
151 * \param[in] startingOffset Offset (in values) at the device-side buffer to copy from.
152 * \param[in] numValues Number of values to copy.
153 * \param[in] deviceStream GPU stream to perform asynchronous copy in.
154 * \param[in] transferKind Copy type: synchronous or asynchronous.
155 * \param[out] timingEvent A dummy pointer to the H2D copy timing event to be filled in.
156 * Not used in CUDA implementation.
158 template<typename ValueType>
159 void copyFromDeviceBuffer(ValueType* hostBuffer,
160 DeviceBuffer<ValueType>* buffer,
161 size_t startingOffset,
163 const DeviceStream& deviceStream,
164 GpuApiCallBehavior transferKind,
165 CommandEvent* /*timingEvent*/)
167 GMX_ASSERT(buffer, "needs a buffer pointer");
168 GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
171 const size_t bytes = numValues * sizeof(ValueType);
172 switch (transferKind)
174 case GpuApiCallBehavior::Async:
175 GMX_ASSERT(isHostMemoryPinned(hostBuffer),
176 "Destination host buffer was not pinned for CUDA");
177 stat = cudaMemcpyAsync(hostBuffer, *((ValueType**)buffer) + startingOffset, bytes,
178 cudaMemcpyDeviceToHost, deviceStream.stream());
179 GMX_RELEASE_ASSERT(stat == cudaSuccess, "Asynchronous D2H copy failed");
182 case GpuApiCallBehavior::Sync:
183 stat = cudaMemcpy(hostBuffer, *((ValueType**)buffer) + startingOffset, bytes,
184 cudaMemcpyDeviceToHost);
185 GMX_RELEASE_ASSERT(stat == cudaSuccess, "Synchronous D2H copy failed");
193 * Clears the device buffer asynchronously.
195 * \tparam ValueType Raw value type of the \p buffer.
196 * \param[in,out] buffer Pointer to the device-side buffer
197 * \param[in] startingOffset Offset (in values) at the device-side buffer to start clearing at.
198 * \param[in] numValues Number of values to clear.
199 * \param[in] deviceStream GPU stream.
201 template<typename ValueType>
202 void clearDeviceBufferAsync(DeviceBuffer<ValueType>* buffer,
203 size_t startingOffset,
205 const DeviceStream& deviceStream)
207 GMX_ASSERT(buffer, "needs a buffer pointer");
208 const size_t bytes = numValues * sizeof(ValueType);
209 const char pattern = 0;
211 cudaError_t stat = cudaMemsetAsync(*((ValueType**)buffer) + startingOffset, pattern, bytes,
212 deviceStream.stream());
213 GMX_RELEASE_ASSERT(stat == cudaSuccess, "Couldn't clear the device buffer");
216 /*! \brief Check the validity of the device buffer.
218 * Checks if the buffer is not nullptr.
220 * \todo Add checks on the buffer size when it will be possible.
222 * \param[in] buffer Device buffer to be checked.
223 * \param[in] requiredSize Number of elements that the buffer will have to accommodate.
225 * \returns Whether the device buffer can be set.
228 static bool checkDeviceBuffer(DeviceBuffer<T> buffer, gmx_unused int requiredSize)
230 GMX_ASSERT(buffer != nullptr, "The device pointer is nullptr");
231 return buffer != nullptr;