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/cuda_arch_utils.cuh"
49 #include "gromacs/gpu_utils/device_context.h"
50 #include "gromacs/gpu_utils/device_stream.h"
51 #include "gromacs/gpu_utils/devicebuffer_datatype.h"
52 #include "gromacs/gpu_utils/gpu_utils.h" //only for GpuApiCallBehavior
53 #include "gromacs/gpu_utils/gputraits.cuh"
54 #include "gromacs/utility/gmxassert.h"
55 #include "gromacs/utility/stringutil.h"
58 * Allocates a device-side buffer.
59 * It is currently a caller's responsibility to call it only on not-yet allocated buffers.
61 * \tparam ValueType Raw value type of the \p buffer.
62 * \param[in,out] buffer Pointer to the device-side buffer.
63 * \param[in] numValues Number of values to accommodate.
64 * \param[in] deviceContext The buffer's dummy device context - not managed explicitly in CUDA RT.
66 template<typename ValueType>
67 void allocateDeviceBuffer(DeviceBuffer<ValueType>* buffer, size_t numValues, const DeviceContext& /* deviceContext */)
69 GMX_ASSERT(buffer, "needs a buffer pointer");
70 cudaError_t stat = cudaMalloc((void**)buffer, numValues * sizeof(ValueType));
71 GMX_RELEASE_ASSERT(stat == cudaSuccess, "cudaMalloc failure");
75 * Frees a device-side buffer.
76 * This does not reset separately stored size/capacity integers,
77 * as this is planned to be a destructor of DeviceBuffer as a proper class,
78 * and no calls on \p buffer should be made afterwards.
80 * \param[in] buffer Pointer to the buffer to free.
82 template<typename DeviceBuffer>
83 void freeDeviceBuffer(DeviceBuffer* buffer)
85 GMX_ASSERT(buffer, "needs a buffer pointer");
88 GMX_RELEASE_ASSERT(cudaFree(*buffer) == cudaSuccess, "cudaFree failed");
93 * Performs the host-to-device data copy, synchronous or asynchronously on request.
95 * \tparam ValueType Raw value type of the \p buffer.
96 * \param[in,out] buffer Pointer to the device-side buffer
97 * \param[in] hostBuffer Pointer to the raw host-side memory, also typed \p ValueType
98 * \param[in] startingOffset Offset (in values) at the device-side buffer to copy into.
99 * \param[in] numValues Number of values to copy.
100 * \param[in] deviceStream GPU stream to perform asynchronous copy in.
101 * \param[in] transferKind Copy type: synchronous or asynchronous.
102 * \param[out] timingEvent A dummy pointer to the H2D copy timing event to be filled in.
103 * Not used in CUDA implementation.
105 template<typename ValueType>
106 void copyToDeviceBuffer(DeviceBuffer<ValueType>* buffer,
107 const ValueType* hostBuffer,
108 size_t startingOffset,
110 const DeviceStream& deviceStream,
111 GpuApiCallBehavior transferKind,
112 CommandEvent* /*timingEvent*/)
118 GMX_ASSERT(buffer, "needs a buffer pointer");
119 GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
121 const size_t bytes = numValues * sizeof(ValueType);
123 switch (transferKind)
125 case GpuApiCallBehavior::Async:
126 GMX_ASSERT(isHostMemoryPinned(hostBuffer),
127 "Source host buffer was not pinned for CUDA");
128 stat = cudaMemcpyAsync(*((ValueType**)buffer) + startingOffset, hostBuffer, bytes,
129 cudaMemcpyHostToDevice, deviceStream.stream());
130 GMX_RELEASE_ASSERT(stat == cudaSuccess, "Asynchronous H2D copy failed");
133 case GpuApiCallBehavior::Sync:
134 stat = cudaMemcpy(*((ValueType**)buffer) + startingOffset, hostBuffer, bytes,
135 cudaMemcpyHostToDevice);
136 GMX_RELEASE_ASSERT(stat == cudaSuccess, "Synchronous H2D copy failed");
144 * Performs the device-to-host data copy, synchronous or asynchronously on request.
146 * \tparam ValueType Raw value type of the \p buffer.
147 * \param[in,out] hostBuffer Pointer to the raw host-side memory, also typed \p ValueType
148 * \param[in] buffer Pointer to the device-side buffer
149 * \param[in] startingOffset Offset (in values) at the device-side buffer to copy from.
150 * \param[in] numValues Number of values to copy.
151 * \param[in] deviceStream GPU stream to perform asynchronous copy in.
152 * \param[in] transferKind Copy type: synchronous or asynchronous.
153 * \param[out] timingEvent A dummy pointer to the H2D copy timing event to be filled in.
154 * Not used in CUDA implementation.
156 template<typename ValueType>
157 void copyFromDeviceBuffer(ValueType* hostBuffer,
158 DeviceBuffer<ValueType>* buffer,
159 size_t startingOffset,
161 const DeviceStream& deviceStream,
162 GpuApiCallBehavior transferKind,
163 CommandEvent* /*timingEvent*/)
169 GMX_ASSERT(buffer, "needs a buffer pointer");
170 GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
173 const size_t bytes = numValues * sizeof(ValueType);
174 switch (transferKind)
176 case GpuApiCallBehavior::Async:
177 GMX_ASSERT(isHostMemoryPinned(hostBuffer),
178 "Destination host buffer was not pinned for CUDA");
179 stat = cudaMemcpyAsync(hostBuffer, *((ValueType**)buffer) + startingOffset, bytes,
180 cudaMemcpyDeviceToHost, deviceStream.stream());
181 GMX_RELEASE_ASSERT(stat == cudaSuccess, "Asynchronous D2H copy failed");
184 case GpuApiCallBehavior::Sync:
185 stat = cudaMemcpy(hostBuffer, *((ValueType**)buffer) + startingOffset, bytes,
186 cudaMemcpyDeviceToHost);
187 GMX_RELEASE_ASSERT(stat == cudaSuccess, "Synchronous D2H copy failed");
195 * Clears the device buffer asynchronously.
197 * \tparam ValueType Raw value type of the \p buffer.
198 * \param[in,out] buffer Pointer to the device-side buffer
199 * \param[in] startingOffset Offset (in values) at the device-side buffer to start clearing at.
200 * \param[in] numValues Number of values to clear.
201 * \param[in] deviceStream GPU stream.
203 template<typename ValueType>
204 void clearDeviceBufferAsync(DeviceBuffer<ValueType>* buffer,
205 size_t startingOffset,
207 const DeviceStream& deviceStream)
209 GMX_ASSERT(buffer, "needs a buffer pointer");
210 const size_t bytes = numValues * sizeof(ValueType);
211 const char pattern = 0;
213 cudaError_t stat = cudaMemsetAsync(*((ValueType**)buffer) + startingOffset, pattern, bytes,
214 deviceStream.stream());
215 GMX_RELEASE_ASSERT(stat == cudaSuccess, "Couldn't clear the device buffer");
218 /*! \brief Check the validity of the device buffer.
220 * Checks if the buffer is not nullptr.
222 * \todo Add checks on the buffer size when it will be possible.
224 * \param[in] buffer Device buffer to be checked.
225 * \param[in] requiredSize Number of elements that the buffer will have to accommodate.
227 * \returns Whether the device buffer can be set.
230 gmx_unused static bool checkDeviceBuffer(DeviceBuffer<T> buffer, gmx_unused int requiredSize)
232 GMX_ASSERT(buffer != nullptr, "The device pointer is nullptr");
233 return buffer != nullptr;
236 //! Device texture wrapper.
237 using DeviceTexture = cudaTextureObject_t;
239 /*! \brief Create a texture object for an array of type ValueType.
241 * Creates the device buffer, copies data and binds texture object for an array of type ValueType.
243 * \todo Test if using textures is still relevant on modern hardware.
245 * \tparam ValueType Raw data type.
247 * \param[out] deviceBuffer Device buffer to store data in.
248 * \param[out] deviceTexture Device texture object to initialize.
249 * \param[in] hostBuffer Host buffer to get date from
250 * \param[in] numValues Number of elements in the buffer.
251 * \param[in] deviceContext GPU device context.
253 template<typename ValueType>
254 void initParamLookupTable(DeviceBuffer<ValueType>* deviceBuffer,
255 DeviceTexture* deviceTexture,
256 const ValueType* hostBuffer,
258 const DeviceContext& deviceContext)
264 GMX_ASSERT(hostBuffer, "Host buffer should be specified.");
266 allocateDeviceBuffer(deviceBuffer, numValues, deviceContext);
268 const size_t sizeInBytes = numValues * sizeof(ValueType);
271 cudaMemcpy(*((ValueType**)deviceBuffer), hostBuffer, sizeInBytes, cudaMemcpyHostToDevice);
275 gmx::formatString("Synchronous H2D copy failed (CUDA error: %s).", cudaGetErrorName(stat))
278 if (!c_disableCudaTextures)
283 memset(&rd, 0, sizeof(rd));
284 rd.resType = cudaResourceTypeLinear;
285 rd.res.linear.devPtr = *deviceBuffer;
286 rd.res.linear.desc = cudaCreateChannelDesc<ValueType>();
287 rd.res.linear.sizeInBytes = sizeInBytes;
289 memset(&td, 0, sizeof(td));
290 td.readMode = cudaReadModeElementType;
291 stat = cudaCreateTextureObject(deviceTexture, &rd, &td, nullptr);
292 GMX_RELEASE_ASSERT(stat == cudaSuccess,
293 gmx::formatString("cudaCreateTextureObject failed (CUDA error: %s).",
294 cudaGetErrorName(stat))
299 /*! \brief Unbind the texture and release the CUDA texture object.
301 * \tparam ValueType Raw data type
303 * \param[in,out] deviceBuffer Device buffer to store data in.
304 * \param[in,out] deviceTexture Device texture object to unbind.
306 template<typename ValueType>
307 void destroyParamLookupTable(DeviceBuffer<ValueType>* deviceBuffer, DeviceTexture& deviceTexture)
309 if (!c_disableCudaTextures && deviceTexture && deviceBuffer)
311 cudaError_t stat = cudaDestroyTextureObject(deviceTexture);
315 "cudaDestroyTextureObject on texture object failed (CUDA error: %s).",
316 cudaGetErrorName(stat))
319 freeDeviceBuffer(deviceBuffer);