2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2010, The GROMACS development team.
6 * Copyright (c) 2012,2013,2014,2015,2016,2017,2018,2019, by the GROMACS development team, led by
7 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
8 * and including many others, as listed in the AUTHORS file in the
9 * top-level source directory and at http://www.gromacs.org.
11 * GROMACS is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public License
13 * as published by the Free Software Foundation; either version 2.1
14 * of the License, or (at your option) any later version.
16 * GROMACS is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with GROMACS; if not, see
23 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
24 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
26 * If you want to redistribute modifications to GROMACS, please
27 * consider that scientific software is very special. Version
28 * control is crucial - bugs must be traceable. We will be happy to
29 * consider code for inclusion in the official distribution, but
30 * derived work must not be called official GROMACS. Details are found
31 * in the README & COPYING files - if they are missing, get the
32 * official version at http://www.gromacs.org.
34 * To help us fund GROMACS development, we humbly ask that you cite
35 * the research papers on the package. Check out http://www.gromacs.org.
37 /*! \libinternal \file
38 * \brief Declare functions for detection and initialization for GPU devices.
40 * \author Szilard Pall <pall.szilard@gmail.com>
41 * \author Mark Abraham <mark.j.abraham@gmail.com>
45 #ifndef GMX_GPU_UTILS_GPU_UTILS_H
46 #define GMX_GPU_UTILS_GPU_UTILS_H
53 #include "gromacs/gpu_utils/gpu_macros.h"
54 #include "gromacs/utility/basedefinitions.h"
56 struct gmx_device_info_t;
57 struct gmx_gpu_info_t;
63 //! Enum which is only used to describe transfer calls at the moment
64 enum class GpuApiCallBehavior
70 //! Types of actions associated to waiting or checking the completion of GPU tasks
71 enum class GpuTaskCompletion
73 Wait, /*<< Issue a blocking wait for the task */
74 Check /*<< Only check whether the task has completed */
77 /*! \brief Return whether GPUs can be detected
79 * Returns true when this is a build of \Gromacs configured to support
80 * GPU usage, and a valid device driver, ICD, and/or runtime was detected.
82 * This function is not intended to be called from build
83 * configurations that do not support GPUs, and there will be no
84 * descriptive message in that case.
86 * \param[out] errorMessage When returning false on a build configured with
87 * GPU support and non-nullptr was passed,
88 * the string contains a descriptive message about
89 * why GPUs cannot be detected.
93 bool canDetectGpus(std::string *GPU_FUNC_ARGUMENT(errorMessage)) GPU_FUNC_TERM_WITH_RETURN(false);
95 /*! \brief Find all GPUs in the system.
97 * Will detect every GPU supported by the device driver in use. Must
98 * only be called if canDetectGpus() has returned true. This routine
99 * also checks for the compatibility of each and fill the
100 * gpu_info->gpu_dev array with the required information on each the
101 * device: ID, device properties, status.
103 * Note that this function leaves the GPU runtime API error state clean;
104 * this is implemented ATM in the CUDA flavor.
105 * TODO: check if errors do propagate in OpenCL as they do in CUDA and
106 * whether there is a mechanism to "clear" them.
108 * \param[in] gpu_info pointer to structure holding GPU information.
110 * \throws InternalError if a GPU API returns an unexpected failure (because
111 * the call to canDetectGpus() should always prevent this occuring)
114 void findGpus(gmx_gpu_info_t *GPU_FUNC_ARGUMENT(gpu_info)) GPU_FUNC_TERM
116 /*! \brief Return a container of the detected GPUs that are compatible.
118 * This function filters the result of the detection for compatible
119 * GPUs, based on the previously run compatibility tests.
121 * \param[in] gpu_info Information detected about GPUs, including compatibility.
122 * \return vector of IDs of GPUs already recorded as compatible */
123 std::vector<int> getCompatibleGpus(const gmx_gpu_info_t &gpu_info);
125 /*! \brief Return a string describing how compatible the GPU with given \c index is.
127 * \param[in] gpu_info Information about detected GPUs
128 * \param[in] index index of GPU to ask about
129 * \returns A null-terminated C string describing the compatibility status, useful for error messages.
131 const char *getGpuCompatibilityDescription(const gmx_gpu_info_t &gpu_info,
134 /*! \brief Frees the gpu_dev and dev_use array fields of \p gpu_info.
136 * \param[in] gpu_info pointer to structure holding GPU information
138 void free_gpu_info(const gmx_gpu_info_t *gpu_info);
140 /*! \brief Initializes the GPU described by \c deviceInfo.
142 * TODO Doxygen complains about these - probably a Doxygen bug, since
143 * the patterns here are the same as elsewhere in this header.
145 * \param[in] deviceInfo device info of the GPU to initialize
147 * Issues a fatal error for any critical errors that occur during
151 void init_gpu(const gmx_device_info_t *GPU_FUNC_ARGUMENT(deviceInfo)) GPU_FUNC_TERM
153 /*! \brief Frees up the CUDA GPU used by the active context at the time of calling.
155 * If \c deviceInfo is nullptr, then it is understood that no device
156 * was selected so no context is active to be freed. Otherwise, the
157 * context is explicitly destroyed and therefore all data uploaded to
158 * the GPU is lost. This must only be called when none of this data is
159 * required anymore, because subsequent attempts to free memory
160 * associated with the context will otherwise fail.
162 * Calls gmx_warning upon errors.
164 * \param[in] deviceInfo device info of the GPU to clean up for
166 * \returns true if no error occurs during the freeing.
169 void free_gpu(const gmx_device_info_t *CUDA_FUNC_ARGUMENT(deviceInfo)) CUDA_FUNC_TERM
171 /*! \brief Return a pointer to the device info for \c deviceId
173 * \param[in] gpu_info GPU info of all detected devices in the system.
174 * \param[in] deviceId ID for the GPU device requested.
176 * \returns Pointer to the device info for \c deviceId.
179 gmx_device_info_t *getDeviceInfo(const gmx_gpu_info_t &GPU_FUNC_ARGUMENT(gpu_info),
180 int GPU_FUNC_ARGUMENT(deviceId)) GPU_FUNC_TERM_WITH_RETURN(nullptr)
182 /*! \brief Returns the device ID of the CUDA GPU currently in use.
184 * The GPU used is the one that is active at the time of the call in the active context.
186 * \returns device ID of the GPU in use at the time of the call
189 int get_current_cuda_gpu_device_id(void) CUDA_FUNC_TERM_WITH_RETURN(-1)
191 /*! \brief Formats and returns a device information string for a given GPU.
193 * Given an index *directly* into the array of available GPUs (gpu_dev)
194 * returns a formatted info string for the respective GPU which includes
195 * ID, name, compute capability, and detection status.
197 * \param[out] s pointer to output string (has to be allocated externally)
198 * \param[in] gpu_info Information about detected GPUs
199 * \param[in] index an index *directly* into the array of available GPUs
202 void get_gpu_device_info_string(char *GPU_FUNC_ARGUMENT(s),
203 const gmx_gpu_info_t &GPU_FUNC_ARGUMENT(gpu_info),
204 int GPU_FUNC_ARGUMENT(index)) GPU_FUNC_TERM
206 /*! \brief Returns whether all compatible OpenCL devices are from AMD.
208 * This is currently the most useful and best tested platform for
209 * supported OpenCL devices, so some modules may need to check what
210 * degree of support they should offer.
212 * \todo An enumeration visible in the hardware module would make such
213 * checks more configurable, if we discover other needs in future.
215 * \returns whether all detected compatible devices have AMD for the vendor.
217 OPENCL_FUNC_QUALIFIER
218 bool areAllGpuDevicesFromAmd(const gmx_gpu_info_t &OPENCL_FUNC_ARGUMENT(gpuInfo))
219 OPENCL_FUNC_TERM_WITH_RETURN(false)
221 /*! \brief Returns the size of the gpu_dev_info struct.
223 * The size of gpu_dev_info can be used for allocation and communication.
225 * \returns size in bytes of gpu_dev_info
228 size_t sizeof_gpu_dev_info() GPU_FUNC_TERM_WITH_RETURN(0)
230 //! Get status of device with specified index
231 int gpu_info_get_stat(const gmx_gpu_info_t &info, int index);
233 /*! \brief Check if GROMACS has been built with GPU support.
235 * \param[in] error Pointer to error string or nullptr.
236 * \todo Move this to NB module once it exists.
238 bool buildSupportsNonbondedOnGpu(std::string *error);
240 /*! \brief Starts the GPU profiler if mdrun is being profiled.
242 * When a profiler run is in progress (based on the presence of the NVPROF_ID
243 * env. var.), the profiler is started to begin collecting data during the
244 * rest of the run (or until stopGpuProfiler is called).
246 * Note that this is implemented only for the CUDA API.
249 void startGpuProfiler(void) CUDA_FUNC_TERM
252 /*! \brief Resets the GPU profiler if mdrun is being profiled.
254 * When a profiler run is in progress (based on the presence of the NVPROF_ID
255 * env. var.), the profiler data is restet in order to eliminate the data collected
256 * from the preceding part fo the run.
258 * This function should typically be called at the mdrun counter reset time.
260 * Note that this is implemented only for the CUDA API.
263 void resetGpuProfiler(void) CUDA_FUNC_TERM
266 /*! \brief Stops the CUDA profiler if mdrun is being profiled.
268 * This function can be called at cleanup when skipping recording
269 * recording subsequent API calls from being traces/profiled is desired,
270 * e.g. before uninitialization.
272 * Note that this is implemented only for the CUDA API.
275 void stopGpuProfiler(void) CUDA_FUNC_TERM
277 //! Tells whether the host buffer was pinned for non-blocking transfers. Only implemented for CUDA.
279 bool isHostMemoryPinned(const void *CUDA_FUNC_ARGUMENT(h_ptr)) CUDA_FUNC_TERM_WITH_RETURN(false)