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/devicebuffer_datatype.h"
51 #include "gromacs/gpu_utils/gpu_utils.h" //only for GpuApiCallBehavior
52 #include "gromacs/gpu_utils/gputraits.cuh"
53 #include "gromacs/utility/gmxassert.h"
54 #include "gromacs/utility/stringutil.h"
57 * Allocates a device-side buffer.
58 * It is currently a caller's responsibility to call it only on not-yet allocated buffers.
60 * \tparam ValueType Raw value type of the \p buffer.
61 * \param[in,out] buffer Pointer to the device-side buffer.
62 * \param[in] numValues Number of values to accomodate.
63 * \param[in] deviceContext The buffer's dummy device context - not managed explicitly in CUDA RT.
65 template<typename ValueType>
66 void allocateDeviceBuffer(DeviceBuffer<ValueType>* buffer, size_t numValues, const DeviceContext& /* deviceContext */)
68 GMX_ASSERT(buffer, "needs a buffer pointer");
69 cudaError_t stat = cudaMalloc((void**)buffer, numValues * sizeof(ValueType));
70 GMX_RELEASE_ASSERT(stat == cudaSuccess, "cudaMalloc failure");
74 * Frees a device-side buffer.
75 * This does not reset separately stored size/capacity integers,
76 * as this is planned to be a destructor of DeviceBuffer as a proper class,
77 * and no calls on \p buffer should be made afterwards.
79 * \param[in] buffer Pointer to the buffer to free.
81 template<typename DeviceBuffer>
82 void freeDeviceBuffer(DeviceBuffer* buffer)
84 GMX_ASSERT(buffer, "needs a buffer pointer");
87 GMX_RELEASE_ASSERT(cudaFree(*buffer) == cudaSuccess, "cudaFree failed");
92 * Performs the host-to-device data copy, synchronous or asynchronously on request.
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*/)
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");
143 * Performs the device-to-host data copy, synchronous or asynchronously on request.
145 * \tparam ValueType Raw value type of the \p buffer.
146 * \param[in,out] hostBuffer Pointer to the raw host-side memory, also typed \p ValueType
147 * \param[in] buffer Pointer to the device-side buffer
148 * \param[in] startingOffset Offset (in values) at the device-side buffer to copy from.
149 * \param[in] numValues Number of values to copy.
150 * \param[in] deviceStream GPU stream to perform asynchronous copy in.
151 * \param[in] transferKind Copy type: synchronous or asynchronous.
152 * \param[out] timingEvent A dummy pointer to the H2D copy timing event to be filled in.
153 * Not used in CUDA implementation.
155 template<typename ValueType>
156 void copyFromDeviceBuffer(ValueType* hostBuffer,
157 DeviceBuffer<ValueType>* buffer,
158 size_t startingOffset,
160 const DeviceStream& deviceStream,
161 GpuApiCallBehavior transferKind,
162 CommandEvent* /*timingEvent*/)
168 GMX_ASSERT(buffer, "needs a buffer pointer");
169 GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
172 const size_t bytes = numValues * sizeof(ValueType);
173 switch (transferKind)
175 case GpuApiCallBehavior::Async:
176 GMX_ASSERT(isHostMemoryPinned(hostBuffer),
177 "Destination host buffer was not pinned for CUDA");
178 stat = cudaMemcpyAsync(hostBuffer, *((ValueType**)buffer) + startingOffset, bytes,
179 cudaMemcpyDeviceToHost, deviceStream.stream());
180 GMX_RELEASE_ASSERT(stat == cudaSuccess, "Asynchronous D2H copy failed");
183 case GpuApiCallBehavior::Sync:
184 stat = cudaMemcpy(hostBuffer, *((ValueType**)buffer) + startingOffset, bytes,
185 cudaMemcpyDeviceToHost);
186 GMX_RELEASE_ASSERT(stat == cudaSuccess, "Synchronous D2H copy failed");
194 * Clears the device buffer asynchronously.
196 * \tparam ValueType Raw value type of the \p buffer.
197 * \param[in,out] buffer Pointer to the device-side buffer
198 * \param[in] startingOffset Offset (in values) at the device-side buffer to start clearing at.
199 * \param[in] numValues Number of values to clear.
200 * \param[in] deviceStream GPU stream.
202 template<typename ValueType>
203 void clearDeviceBufferAsync(DeviceBuffer<ValueType>* buffer,
204 size_t startingOffset,
206 const DeviceStream& deviceStream)
208 GMX_ASSERT(buffer, "needs a buffer pointer");
209 const size_t bytes = numValues * sizeof(ValueType);
210 const char pattern = 0;
212 cudaError_t stat = cudaMemsetAsync(*((ValueType**)buffer) + startingOffset, pattern, bytes,
213 deviceStream.stream());
214 GMX_RELEASE_ASSERT(stat == cudaSuccess, "Couldn't clear the device buffer");
217 /*! \brief Check the validity of the device buffer.
219 * Checks if the buffer is not nullptr.
221 * \todo Add checks on the buffer size when it will be possible.
223 * \param[in] buffer Device buffer to be checked.
224 * \param[in] requiredSize Number of elements that the buffer will have to accommodate.
226 * \returns Whether the device buffer can be set.
229 static bool checkDeviceBuffer(DeviceBuffer<T> buffer, gmx_unused int requiredSize)
231 GMX_ASSERT(buffer != nullptr, "The device pointer is nullptr");
232 return buffer != nullptr;
235 //! Device texture wrapper.
236 using DeviceTexture = cudaTextureObject_t;
238 /*! \brief Create a texture object for an array of type ValueType.
240 * Creates the device buffer, copies data and binds texture object for an array of type ValueType.
242 * \todo Test if using textures is still relevant on modern hardware.
244 * \tparam ValueType Raw data type.
246 * \param[out] deviceBuffer Device buffer to store data in.
247 * \param[out] deviceTexture Device texture object to initialize.
248 * \param[in] hostBuffer Host buffer to get date from
249 * \param[in] numValues Number of elements in the buffer.
250 * \param[in] deviceContext GPU device context.
252 template<typename ValueType>
253 void initParamLookupTable(DeviceBuffer<ValueType>* deviceBuffer,
254 DeviceTexture* deviceTexture,
255 const ValueType* hostBuffer,
257 const DeviceContext& deviceContext)
263 GMX_ASSERT(hostBuffer, "Host buffer should be specified.");
265 allocateDeviceBuffer(deviceBuffer, numValues, deviceContext);
267 const size_t sizeInBytes = numValues * sizeof(ValueType);
270 cudaMemcpy(*((ValueType**)deviceBuffer), hostBuffer, sizeInBytes, cudaMemcpyHostToDevice);
274 gmx::formatString("Synchronous H2D copy failed (CUDA error: %s).", cudaGetErrorName(stat))
277 if (!c_disableCudaTextures)
282 memset(&rd, 0, sizeof(rd));
283 rd.resType = cudaResourceTypeLinear;
284 rd.res.linear.devPtr = *deviceBuffer;
285 rd.res.linear.desc = cudaCreateChannelDesc<ValueType>();
286 rd.res.linear.sizeInBytes = sizeInBytes;
288 memset(&td, 0, sizeof(td));
289 td.readMode = cudaReadModeElementType;
290 stat = cudaCreateTextureObject(deviceTexture, &rd, &td, nullptr);
291 GMX_RELEASE_ASSERT(stat == cudaSuccess,
292 gmx::formatString("cudaCreateTextureObject failed (CUDA error: %s).",
293 cudaGetErrorName(stat))
298 /*! \brief Unbind the texture and release the CUDA texture object.
300 * \tparam ValueType Raw data type
302 * \param[in,out] deviceBuffer Device buffer to store data in.
303 * \param[in,out] deviceTexture Device texture object to unbind.
305 template<typename ValueType>
306 void destroyParamLookupTable(DeviceBuffer<ValueType>* deviceBuffer, DeviceTexture& deviceTexture)
308 if (!c_disableCudaTextures && deviceTexture && deviceBuffer)
310 cudaError_t stat = cudaDestroyTextureObject(deviceTexture);
314 "cudaDestroyTextureObject on texture object failed (CUDA error: %s).",
315 cudaGetErrorName(stat))
318 freeDeviceBuffer(deviceBuffer);