2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2018,2019,2020,2021, 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/cudautils.cuh"
50 #include "gromacs/gpu_utils/device_context.h"
51 #include "gromacs/gpu_utils/device_stream.h"
52 #include "gromacs/gpu_utils/devicebuffer_datatype.h"
53 #include "gromacs/gpu_utils/gpu_utils.h" //only for GpuApiCallBehavior
54 #include "gromacs/gpu_utils/gputraits.cuh"
55 #include "gromacs/utility/gmxassert.h"
56 #include "gromacs/utility/stringutil.h"
59 * Allocates a device-side buffer.
60 * It is currently a caller's responsibility to call it only on not-yet allocated buffers.
62 * \tparam ValueType Raw value type of the \p buffer.
63 * \param[in,out] buffer Pointer to the device-side buffer.
64 * \param[in] numValues Number of values to accommodate.
65 * \param[in] deviceContext The buffer's dummy device context - not managed explicitly in CUDA RT.
67 template<typename ValueType>
68 void allocateDeviceBuffer(DeviceBuffer<ValueType>* buffer, size_t numValues, const DeviceContext& /* deviceContext */)
70 GMX_ASSERT(buffer, "needs a buffer pointer");
71 // NOLINTNEXTLINE(google-readability-casting)
72 cudaError_t stat = cudaMalloc((void**)(buffer), numValues * sizeof(ValueType));
75 ("Allocation of the device buffer failed. " + gmx::getDeviceErrorString(stat)).c_str());
79 * Frees a device-side buffer.
80 * This does not reset separately stored size/capacity integers,
81 * as this is planned to be a destructor of DeviceBuffer as a proper class,
82 * and no calls on \p buffer should be made afterwards.
84 * \param[in] buffer Pointer to the buffer to free.
86 template<typename DeviceBuffer>
87 void freeDeviceBuffer(DeviceBuffer* buffer)
89 GMX_ASSERT(buffer, "needs a buffer pointer");
92 cudaError_t stat = cudaFree(*buffer);
95 ("Freeing of the device buffer failed. " + gmx::getDeviceErrorString(stat)).c_str());
100 * Performs the host-to-device data copy, synchronous or asynchronously on request.
102 * \tparam ValueType Raw value type of the \p buffer.
103 * \param[in,out] buffer Pointer to the device-side buffer
104 * \param[in] hostBuffer Pointer to the raw host-side memory, also typed \p ValueType
105 * \param[in] startingOffset Offset (in values) at the device-side buffer to copy into.
106 * \param[in] numValues Number of values to copy.
107 * \param[in] deviceStream GPU stream to perform asynchronous copy in.
108 * \param[in] transferKind Copy type: synchronous or asynchronous.
109 * \param[out] timingEvent A dummy pointer to the H2D copy timing event to be filled in.
110 * Not used in CUDA implementation.
112 template<typename ValueType>
113 void copyToDeviceBuffer(DeviceBuffer<ValueType>* buffer,
114 const ValueType* hostBuffer,
115 size_t startingOffset,
117 const DeviceStream& deviceStream,
118 GpuApiCallBehavior transferKind,
119 CommandEvent* /*timingEvent*/)
125 GMX_ASSERT(buffer, "needs a buffer pointer");
126 GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
128 const size_t bytes = numValues * sizeof(ValueType);
130 switch (transferKind)
132 case GpuApiCallBehavior::Async:
133 GMX_ASSERT(isHostMemoryPinned(hostBuffer), "Source host buffer was not pinned for CUDA");
134 // NOLINTNEXTLINE(google-readability-casting)
135 stat = cudaMemcpyAsync(*((ValueType**)buffer) + startingOffset,
138 cudaMemcpyHostToDevice,
139 deviceStream.stream());
142 ("Asynchronous H2D copy failed. " + gmx::getDeviceErrorString(stat)).c_str());
145 case GpuApiCallBehavior::Sync:
147 // NOLINTNEXTLINE(google-readability-casting)
148 *((ValueType**)buffer) + startingOffset,
151 cudaMemcpyHostToDevice);
154 ("Synchronous H2D copy failed. " + gmx::getDeviceErrorString(stat)).c_str());
162 * Performs the device-to-host data copy, synchronous or asynchronously on request.
164 * \tparam ValueType Raw value type of the \p buffer.
165 * \param[in,out] hostBuffer Pointer to the raw host-side memory, also typed \p ValueType
166 * \param[in] buffer Pointer to the device-side buffer
167 * \param[in] startingOffset Offset (in values) at the device-side buffer to copy from.
168 * \param[in] numValues Number of values to copy.
169 * \param[in] deviceStream GPU stream to perform asynchronous copy in.
170 * \param[in] transferKind Copy type: synchronous or asynchronous.
171 * \param[out] timingEvent A dummy pointer to the H2D copy timing event to be filled in.
172 * Not used in CUDA implementation.
174 template<typename ValueType>
175 void copyFromDeviceBuffer(ValueType* hostBuffer,
176 DeviceBuffer<ValueType>* buffer,
177 size_t startingOffset,
179 const DeviceStream& deviceStream,
180 GpuApiCallBehavior transferKind,
181 CommandEvent* /*timingEvent*/)
187 GMX_ASSERT(buffer, "needs a buffer pointer");
188 GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
191 const size_t bytes = numValues * sizeof(ValueType);
192 switch (transferKind)
194 case GpuApiCallBehavior::Async:
195 GMX_ASSERT(isHostMemoryPinned(hostBuffer),
196 "Destination host buffer was not pinned for CUDA");
197 stat = cudaMemcpyAsync(hostBuffer,
198 // NOLINTNEXTLINE(google-readability-casting)
199 *((ValueType**)buffer) + startingOffset,
201 cudaMemcpyDeviceToHost,
202 deviceStream.stream());
205 ("Asynchronous D2H copy failed. " + gmx::getDeviceErrorString(stat)).c_str());
208 case GpuApiCallBehavior::Sync:
209 stat = cudaMemcpy(hostBuffer,
210 // NOLINTNEXTLINE(google-readability-casting)
211 *((ValueType**)buffer) + startingOffset,
213 cudaMemcpyDeviceToHost);
216 ("Synchronous D2H copy failed. " + gmx::getDeviceErrorString(stat)).c_str());
224 * Performs the device-to-device data copy, synchronous or asynchronously on request.
226 * \tparam ValueType Raw value type of the \p buffer.
227 * \param[in,out] destinationDeviceBuffer Device-side buffer to copy to
228 * \param[in] sourceDeviceBuffer Device-side buffer to copy from
229 * \param[in] numValues Number of values to copy.
230 * \param[in] deviceStream GPU stream to perform asynchronous copy in.
231 * \param[in] transferKind Copy type: synchronous or asynchronous.
232 * \param[out] timingEvent A dummy pointer to the D2D copy timing event to be filled
233 * in. Not used in CUDA implementation.
235 template<typename ValueType>
236 void copyBetweenDeviceBuffers(DeviceBuffer<ValueType>* destinationDeviceBuffer,
237 DeviceBuffer<ValueType>* sourceDeviceBuffer,
239 const DeviceStream& deviceStream,
240 GpuApiCallBehavior transferKind,
241 CommandEvent* /*timingEvent*/)
247 GMX_ASSERT(destinationDeviceBuffer, "needs a destination buffer pointer");
248 GMX_ASSERT(sourceDeviceBuffer, "needs a source buffer pointer");
251 const size_t bytes = numValues * sizeof(ValueType);
252 switch (transferKind)
254 case GpuApiCallBehavior::Async:
255 stat = cudaMemcpyAsync(*destinationDeviceBuffer,
258 cudaMemcpyDeviceToDevice,
259 deviceStream.stream());
262 ("Asynchronous D2D copy failed. " + gmx::getDeviceErrorString(stat)).c_str());
265 case GpuApiCallBehavior::Sync:
266 stat = cudaMemcpy(*destinationDeviceBuffer, *sourceDeviceBuffer, bytes, cudaMemcpyDeviceToDevice);
269 ("Synchronous D2D copy failed. " + gmx::getDeviceErrorString(stat)).c_str());
277 * Clears the device buffer asynchronously.
279 * \tparam ValueType Raw value type of the \p buffer.
280 * \param[in,out] buffer Pointer to the device-side buffer
281 * \param[in] startingOffset Offset (in values) at the device-side buffer to start clearing at.
282 * \param[in] numValues Number of values to clear.
283 * \param[in] deviceStream GPU stream.
285 template<typename ValueType>
286 void clearDeviceBufferAsync(DeviceBuffer<ValueType>* buffer,
287 size_t startingOffset,
289 const DeviceStream& deviceStream)
295 GMX_ASSERT(buffer, "needs a buffer pointer");
296 const size_t bytes = numValues * sizeof(ValueType);
297 const char pattern = 0;
299 cudaError_t stat = cudaMemsetAsync(
300 // NOLINTNEXTLINE(google-readability-casting)
301 *((ValueType**)buffer) + startingOffset,
304 deviceStream.stream());
305 GMX_RELEASE_ASSERT(stat == cudaSuccess,
306 ("Couldn't clear the device buffer. " + gmx::getDeviceErrorString(stat)).c_str());
309 /*! \brief Check the validity of the device buffer.
311 * Checks if the buffer is not nullptr.
313 * \todo Add checks on the buffer size when it will be possible.
315 * \param[in] buffer Device buffer to be checked.
316 * \param[in] requiredSize Number of elements that the buffer will have to accommodate.
318 * \returns Whether the device buffer can be set.
321 gmx_unused static bool checkDeviceBuffer(DeviceBuffer<T> buffer, gmx_unused int requiredSize)
323 GMX_ASSERT(buffer != nullptr, "The device pointer is nullptr");
324 return buffer != nullptr;
327 //! Device texture wrapper.
328 using DeviceTexture = cudaTextureObject_t;
330 /*! \brief Create a texture object for an array of type ValueType.
332 * Creates the device buffer, copies data and binds texture object for an array of type ValueType.
334 * \todo Test if using textures is still relevant on modern hardware.
336 * \tparam ValueType Raw data type.
338 * \param[out] deviceBuffer Device buffer to store data in.
339 * \param[out] deviceTexture Device texture object to initialize.
340 * \param[in] hostBuffer Host buffer to get date from
341 * \param[in] numValues Number of elements in the buffer.
342 * \param[in] deviceContext GPU device context.
344 template<typename ValueType>
345 void initParamLookupTable(DeviceBuffer<ValueType>* deviceBuffer,
346 DeviceTexture* deviceTexture,
347 const ValueType* hostBuffer,
349 const DeviceContext& deviceContext)
355 GMX_ASSERT(hostBuffer, "Host buffer should be specified.");
357 allocateDeviceBuffer(deviceBuffer, numValues, deviceContext);
359 const size_t sizeInBytes = numValues * sizeof(ValueType);
362 // NOLINTNEXTLINE(google-readability-casting)
363 cudaMemcpy(*((ValueType**)deviceBuffer), hostBuffer, sizeInBytes, cudaMemcpyHostToDevice);
365 GMX_RELEASE_ASSERT(stat == cudaSuccess,
366 ("Synchronous H2D copy failed. " + gmx::getDeviceErrorString(stat)).c_str());
368 if (!c_disableCudaTextures)
373 memset(&rd, 0, sizeof(rd));
374 rd.resType = cudaResourceTypeLinear;
375 rd.res.linear.devPtr = *deviceBuffer;
376 rd.res.linear.desc = cudaCreateChannelDesc<ValueType>();
377 rd.res.linear.sizeInBytes = sizeInBytes;
379 memset(&td, 0, sizeof(td));
380 td.readMode = cudaReadModeElementType;
381 stat = cudaCreateTextureObject(deviceTexture, &rd, &td, nullptr);
384 ("Binding of the texture object failed. " + gmx::getDeviceErrorString(stat)).c_str());
388 /*! \brief Unbind the texture and release the CUDA texture object.
390 * \tparam ValueType Raw data type
392 * \param[in,out] deviceBuffer Device buffer to store data in.
393 * \param[in,out] deviceTexture Device texture object to unbind.
395 template<typename ValueType>
396 void destroyParamLookupTable(DeviceBuffer<ValueType>* deviceBuffer, const DeviceTexture* deviceTexture)
398 if (!c_disableCudaTextures && deviceTexture && deviceBuffer)
400 cudaError_t stat = cudaDestroyTextureObject(*deviceTexture);
403 ("Destruction of the texture object failed. " + gmx::getDeviceErrorString(stat)).c_str());
405 freeDeviceBuffer(deviceBuffer);