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_OCL_H
36 #define GMX_GPU_UTILS_DEVICEBUFFER_OCL_H
38 /*! \libinternal \file
39 * \brief Implements the DeviceBuffer type and routines for OpenCL.
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/devicebuffer_datatype.h"
49 #include "gromacs/gpu_utils/gpu_utils.h" //only for GpuApiCallBehavior
50 #include "gromacs/gpu_utils/gputraits_ocl.h"
51 #include "gromacs/gpu_utils/oclutils.h"
52 #include "gromacs/utility/gmxassert.h"
53 #include "gromacs/utility/stringutil.h"
55 /*! \libinternal \brief
56 * Allocates a device-side buffer.
57 * It is currently a caller's responsibility to call it only on not-yet allocated buffers.
59 * \tparam ValueType Raw value type of the \p buffer.
60 * \param[in,out] buffer Pointer to the device-side buffer.
61 * \param[in] numValues Number of values to accomodate.
62 * \param[in] deviceContext The buffer's device context-to-be.
64 template<typename ValueType>
65 void allocateDeviceBuffer(DeviceBuffer<ValueType>* buffer, size_t numValues, DeviceContext deviceContext)
67 GMX_ASSERT(buffer, "needs a buffer pointer");
68 void* hostPtr = nullptr;
70 *buffer = clCreateBuffer(deviceContext, CL_MEM_READ_WRITE, numValues * sizeof(ValueType),
72 GMX_RELEASE_ASSERT(clError == CL_SUCCESS,
73 gmx::formatString("clCreateBuffer failure (OpenCL error %d: %s)", clError,
74 ocl_get_error_string(clError).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 cl_int clError = clReleaseMemObject(*buffer);
93 GMX_RELEASE_ASSERT(clError == CL_SUCCESS,
94 gmx::formatString("clReleaseMemObject failed (OpenCL error %d: %s)",
95 clError, ocl_get_error_string(clError).c_str())
101 * Performs the host-to-device data copy, synchronous or asynchronously on request.
103 * TODO: This is meant to gradually replace cu/ocl_copy_h2d.
105 * \tparam ValueType Raw value type of the \p buffer.
106 * \param[in,out] buffer Pointer to the device-side buffer
107 * \param[in] hostBuffer Pointer to the raw host-side memory, also typed \p ValueType
108 * \param[in] startingOffset Offset (in values) at the device-side buffer to copy into.
109 * \param[in] numValues Number of values to copy.
110 * \param[in] stream GPU stream to perform asynchronous copy in.
111 * \param[in] transferKind Copy type: synchronous or asynchronous.
112 * \param[out] timingEvent A pointer to the H2D copy timing event to be filled in.
113 * If the pointer is not null, the event can further be used
114 * to queue a wait for this operation or to query profiling information.
116 template<typename ValueType>
117 void copyToDeviceBuffer(DeviceBuffer<ValueType>* buffer,
118 const ValueType* hostBuffer,
119 size_t startingOffset,
121 CommandStream stream,
122 GpuApiCallBehavior transferKind,
123 CommandEvent* timingEvent)
127 return; // such calls are actually made with empty domains
129 GMX_ASSERT(buffer, "needs a buffer pointer");
130 GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
132 const size_t offset = startingOffset * sizeof(ValueType);
133 const size_t bytes = numValues * sizeof(ValueType);
134 switch (transferKind)
136 case GpuApiCallBehavior::Async:
137 clError = clEnqueueWriteBuffer(stream, *buffer, CL_FALSE, offset, bytes, hostBuffer, 0,
138 nullptr, timingEvent);
140 clError == CL_SUCCESS,
141 gmx::formatString("Asynchronous H2D copy failed (OpenCL error %d: %s)", clError,
142 ocl_get_error_string(clError).c_str())
146 case GpuApiCallBehavior::Sync:
147 clError = clEnqueueWriteBuffer(stream, *buffer, CL_TRUE, offset, bytes, hostBuffer, 0,
148 nullptr, timingEvent);
150 clError == CL_SUCCESS,
151 gmx::formatString("Synchronous H2D copy failed (OpenCL error %d: %s)", clError,
152 ocl_get_error_string(clError).c_str())
161 * Performs the device-to-host data copy, synchronous or asynchronously on request.
163 * TODO: This is meant to gradually replace cu/ocl_copy_d2h.
165 * \tparam ValueType Raw value type of the \p buffer.
166 * \param[in,out] hostBuffer Pointer to the raw host-side memory, also typed \p ValueType
167 * \param[in] buffer Pointer to the device-side buffer
168 * \param[in] startingOffset Offset (in values) at the device-side buffer to copy from.
169 * \param[in] numValues Number of values to copy.
170 * \param[in] stream GPU stream to perform asynchronous copy in.
171 * \param[in] transferKind Copy type: synchronous or asynchronous.
172 * \param[out] timingEvent A pointer to the H2D copy timing event to be filled in.
173 * If the pointer is not null, the event can further be used
174 * to queue a wait for this operation or to query profiling information.
176 template<typename ValueType>
177 void copyFromDeviceBuffer(ValueType* hostBuffer,
178 DeviceBuffer<ValueType>* buffer,
179 size_t startingOffset,
181 CommandStream stream,
182 GpuApiCallBehavior transferKind,
183 CommandEvent* timingEvent)
185 GMX_ASSERT(buffer, "needs a buffer pointer");
186 GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
188 const size_t offset = startingOffset * sizeof(ValueType);
189 const size_t bytes = numValues * sizeof(ValueType);
190 switch (transferKind)
192 case GpuApiCallBehavior::Async:
193 clError = clEnqueueReadBuffer(stream, *buffer, CL_FALSE, offset, bytes, hostBuffer, 0,
194 nullptr, timingEvent);
196 clError == CL_SUCCESS,
197 gmx::formatString("Asynchronous D2H copy failed (OpenCL error %d: %s)", clError,
198 ocl_get_error_string(clError).c_str())
202 case GpuApiCallBehavior::Sync:
203 clError = clEnqueueReadBuffer(stream, *buffer, CL_TRUE, offset, bytes, hostBuffer, 0,
204 nullptr, timingEvent);
206 clError == CL_SUCCESS,
207 gmx::formatString("Synchronous D2H copy failed (OpenCL error %d: %s)", clError,
208 ocl_get_error_string(clError).c_str())
217 * Clears the device buffer asynchronously.
219 * \tparam ValueType Raw value type of the \p buffer.
220 * \param[in,out] buffer Pointer to the device-side buffer
221 * \param[in] startingOffset Offset (in values) at the device-side buffer to start clearing at.
222 * \param[in] numValues Number of values to clear.
223 * \param[in] stream GPU stream.
225 template<typename ValueType>
226 void clearDeviceBufferAsync(DeviceBuffer<ValueType>* buffer, size_t startingOffset, size_t numValues, CommandStream stream)
228 GMX_ASSERT(buffer, "needs a buffer pointer");
229 const size_t offset = startingOffset * sizeof(ValueType);
230 const size_t bytes = numValues * sizeof(ValueType);
231 const int pattern = 0;
232 const cl_uint numWaitEvents = 0;
233 const cl_event* waitEvents = nullptr;
234 cl_event commandEvent;
235 cl_int clError = clEnqueueFillBuffer(stream, *buffer, &pattern, sizeof(pattern), offset, bytes,
236 numWaitEvents, waitEvents, &commandEvent);
237 GMX_RELEASE_ASSERT(clError == CL_SUCCESS,
238 gmx::formatString("Couldn't clear the device buffer (OpenCL error %d: %s)",
239 clError, ocl_get_error_string(clError).c_str())
243 /*! \brief Check the validity of the device buffer.
245 * Checks if the buffer is not nullptr and if its allocation is big enough.
247 * \param[in] buffer Device buffer to be checked.
248 * \param[in] requiredSize Number of elements that the buffer will have to accommodate.
250 * \returns Whether the device buffer can be set.
253 static bool checkDeviceBuffer(DeviceBuffer<T> buffer, int requiredSize)
256 int retval = clGetMemObjectInfo(buffer, CL_MEM_SIZE, sizeof(size), &size, nullptr);
257 GMX_ASSERT(retval == CL_SUCCESS,
258 gmx::formatString("clGetMemObjectInfo failed with error code #%d", retval).c_str());
259 GMX_ASSERT(static_cast<int>(size) >= requiredSize,
260 "Number of atoms in device buffer is smaller then required size.");
261 return retval == CL_SUCCESS && static_cast<int>(size) >= requiredSize;