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, 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
52 #include "gromacs/gpu_utils/gpu_macros.h"
53 #include "gromacs/utility/basedefinitions.h"
55 struct gmx_device_info_t;
56 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 Detect all GPUs in the system.
79 * Will detect every GPU supported by the device driver in use. Also
80 * check for the compatibility of each and fill the gpu_info->gpu_dev array
81 * with the required information on each the device: ID, device properties,
84 * \param[in] gpu_info pointer to structure holding GPU information.
85 * \param[out] err_str The error message of any GPU API error that caused
86 * the detection to fail (if there was any). The memory
87 * the pointer points to should be managed externally.
88 * \returns non-zero if the detection encountered a failure, zero otherwise.
91 int detect_gpus(struct gmx_gpu_info_t *GPU_FUNC_ARGUMENT(gpu_info), char *GPU_FUNC_ARGUMENT(err_str)) GPU_FUNC_TERM_WITH_RETURN(-1)
93 /*! \brief Return a container of the detected GPUs that are compatible.
95 * This function filters the result of the detection for compatible
96 * GPUs, based on the previously run compatibility tests.
98 * \param[in] gpu_info Information detected about GPUs, including compatibility.
99 * \return vector of IDs of GPUs already recorded as compatible */
100 std::vector<int> getCompatibleGpus(const gmx_gpu_info_t &gpu_info);
102 /*! \brief Return a string describing how compatible the GPU with given \c index is.
104 * \param[in] gpu_info Information about detected GPUs
105 * \param[in] index index of GPU to ask about
106 * \returns A null-terminated C string describing the compatibility status, useful for error messages.
108 const char *getGpuCompatibilityDescription(const gmx_gpu_info_t &GPU_FUNC_ARGUMENT(gpu_info),
109 int GPU_FUNC_ARGUMENT(index));
111 /*! \brief Frees the gpu_dev and dev_use array fields of \p gpu_info.
113 * \param[in] gpu_info pointer to structure holding GPU information
116 void free_gpu_info(const struct gmx_gpu_info_t *GPU_FUNC_ARGUMENT(gpu_info)) GPU_FUNC_TERM
118 /*! \brief Initializes the GPU described by \c deviceInfo.
120 * TODO Doxygen complains about these - probably a Doxygen bug, since
121 * the patterns here are the same as elsewhere in this header.
123 * param[in] mdlog log file to write to
124 * \param[inout] deviceInfo device info of the GPU to initialize
126 * Issues a fatal error for any critical errors that occur during
130 void init_gpu(const gmx::MDLogger &GPU_FUNC_ARGUMENT(mdlog),
131 gmx_device_info_t *GPU_FUNC_ARGUMENT(deviceInfo)) GPU_FUNC_TERM
133 /*! \brief Frees up the CUDA GPU used by the active context at the time of calling.
135 * If \c deviceInfo is nullptr, then it is understood that no device
136 * was selected so no context is active to be freed. Otherwise, the
137 * context is explicitly destroyed and therefore all data uploaded to
138 * the GPU is lost. This must only be called when none of this data is
139 * required anymore, because subsequent attempts to free memory
140 * associated with the context will otherwise fail.
142 * Calls gmx_warning upon errors.
144 * \param[in] deviceInfo device info of the GPU to clean up for
146 * \returns true if no error occurs during the freeing.
149 void free_gpu(const gmx_device_info_t *CUDA_FUNC_ARGUMENT(deviceInfo)) CUDA_FUNC_TERM
151 /*! \brief Return a pointer to the device info for \c deviceId
153 * \param[in] gpu_info GPU info of all detected devices in the system.
154 * \param[in] deviceId ID for the GPU device requested.
156 * \returns Pointer to the device info for \c deviceId.
159 gmx_device_info_t *getDeviceInfo(const gmx_gpu_info_t &GPU_FUNC_ARGUMENT(gpu_info),
160 int GPU_FUNC_ARGUMENT(deviceId)) GPU_FUNC_TERM_WITH_RETURN(NULL)
162 /*! \brief Returns the device ID of the CUDA GPU currently in use.
164 * The GPU used is the one that is active at the time of the call in the active context.
166 * \returns device ID of the GPU in use at the time of the call
169 int get_current_cuda_gpu_device_id(void) CUDA_FUNC_TERM_WITH_RETURN(-1)
171 /*! \brief Formats and returns a device information string for a given GPU.
173 * Given an index *directly* into the array of available GPUs (gpu_dev)
174 * returns a formatted info string for the respective GPU which includes
175 * ID, name, compute capability, and detection status.
177 * \param[out] s pointer to output string (has to be allocated externally)
178 * \param[in] gpu_info Information about detected GPUs
179 * \param[in] index an index *directly* into the array of available GPUs
182 void get_gpu_device_info_string(char *GPU_FUNC_ARGUMENT(s),
183 const struct gmx_gpu_info_t &GPU_FUNC_ARGUMENT(gpu_info),
184 int GPU_FUNC_ARGUMENT(index)) GPU_FUNC_TERM
186 /*! \brief Returns the size of the gpu_dev_info struct.
188 * The size of gpu_dev_info can be used for allocation and communication.
190 * \returns size in bytes of gpu_dev_info
193 size_t sizeof_gpu_dev_info(void) GPU_FUNC_TERM_WITH_RETURN(0)
195 /*! \brief Returns a pointer *ptr to page-locked memory of size nbytes.
197 * The allocated memory is suitable to be used for data transfers between host
199 * Error handling should be done within this function.
201 typedef void gmx_host_alloc_t (void **ptr, size_t nbytes);
203 /*! \brief Frees page-locked memory pointed to by *ptr.
205 * NULL should not be passed to this function.
207 typedef void gmx_host_free_t (void *ptr);
209 /*! \brief Set page-locked memory allocation functions used by the GPU host. */
210 void gpu_set_host_malloc_and_free(bool bUseGpuKernels,
211 gmx_host_alloc_t **nb_alloc,
212 gmx_host_free_t **nb_free);
216 /*! \brief Starts the GPU profiler if mdrun is being profiled.
218 * When a profiler run is in progress (based on the presence of the NVPROF_ID
219 * env. var.), the profiler is started to begin collecting data during the
220 * rest of the run (or until stopGpuProfiler is called).
222 * Note that this is implemented only for the CUDA API.
225 void startGpuProfiler(void) CUDA_FUNC_TERM
228 /*! \brief Resets the GPU profiler if mdrun is being profiled.
230 * When a profiler run is in progress (based on the presence of the NVPROF_ID
231 * env. var.), the profiler data is restet in order to eliminate the data collected
232 * from the preceding part fo the run.
234 * This function should typically be called at the mdrun counter reset time.
236 * Note that this is implemented only for the CUDA API.
239 void resetGpuProfiler(void) CUDA_FUNC_TERM
242 /*! \brief Stops the CUDA profiler if mdrun is being profiled.
244 * This function can be called at cleanup when skipping recording
245 * recording subsequent API calls from being traces/profiled is desired,
246 * e.g. before uninitialization.
248 * Note that this is implemented only for the CUDA API.
251 void stopGpuProfiler(void) CUDA_FUNC_TERM
253 //! Tells whether the host buffer was pinned for non-blocking transfers. Only implemented for CUDA.
255 bool isHostMemoryPinned(void *CUDA_FUNC_ARGUMENT(h_ptr)) CUDA_FUNC_TERM_WITH_RETURN(false)