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, 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;
64 //! Enum which is only used to describe transfer calls at the moment
65 enum class GpuApiCallBehavior
71 //! Types of actions associated to waiting or checking the completion of GPU tasks
72 enum class GpuTaskCompletion
74 Wait, /*<< Issue a blocking wait for the task */
75 Check /*<< Only check whether the task has completed */
78 /*! \brief Return whether GPUs can be detected
80 * Returns true when this is a build of \Gromacs configured to support
81 * GPU usage, and a valid device driver, ICD, and/or runtime was detected.
83 * \param[out] errorMessage When returning false and non-nullptr was passed,
84 * the string contains a descriptive message about
85 * why GPUs cannot be detected.
89 bool canDetectGpus(std::string *GPU_FUNC_ARGUMENT(errorMessage)) GPU_FUNC_TERM_WITH_RETURN(false);
91 /*! \brief Find all GPUs in the system.
93 * Will detect every GPU supported by the device driver in use. Must
94 * only be called if canDetectGpus() has returned true. This routine
95 * also checks for the compatibility of each and fill the
96 * gpu_info->gpu_dev array with the required information on each the
97 * device: ID, device properties, status.
99 * Note that this function leaves the GPU runtime API error state clean;
100 * this is implemented ATM in the CUDA flavor.
101 * TODO: check if errors do propagate in OpenCL as they do in CUDA and
102 * whether there is a mechanism to "clear" them.
104 * \param[in] gpu_info pointer to structure holding GPU information.
106 * \throws InternalError if a GPU API returns an unexpected failure (because
107 * the call to canDetectGpus() should always prevent this occuring)
110 void findGpus(struct gmx_gpu_info_t *GPU_FUNC_ARGUMENT(gpu_info)) GPU_FUNC_TERM
112 /*! \brief Return a container of the detected GPUs that are compatible.
114 * This function filters the result of the detection for compatible
115 * GPUs, based on the previously run compatibility tests.
117 * \param[in] gpu_info Information detected about GPUs, including compatibility.
118 * \return vector of IDs of GPUs already recorded as compatible */
119 std::vector<int> getCompatibleGpus(const gmx_gpu_info_t &gpu_info);
121 /*! \brief Return a string describing how compatible the GPU with given \c index is.
123 * \param[in] gpu_info Information about detected GPUs
124 * \param[in] index index of GPU to ask about
125 * \returns A null-terminated C string describing the compatibility status, useful for error messages.
127 const char *getGpuCompatibilityDescription(const gmx_gpu_info_t &GPU_FUNC_ARGUMENT(gpu_info),
128 int GPU_FUNC_ARGUMENT(index));
130 /*! \brief Frees the gpu_dev and dev_use array fields of \p gpu_info.
132 * \param[in] gpu_info pointer to structure holding GPU information
135 void free_gpu_info(const struct gmx_gpu_info_t *GPU_FUNC_ARGUMENT(gpu_info)) GPU_FUNC_TERM
137 /*! \brief Initializes the GPU described by \c deviceInfo.
139 * TODO Doxygen complains about these - probably a Doxygen bug, since
140 * the patterns here are the same as elsewhere in this header.
142 * param[in] mdlog log file to write to
143 * \param[inout] deviceInfo device info of the GPU to initialize
145 * Issues a fatal error for any critical errors that occur during
149 void init_gpu(const gmx::MDLogger &GPU_FUNC_ARGUMENT(mdlog),
150 gmx_device_info_t *GPU_FUNC_ARGUMENT(deviceInfo)) GPU_FUNC_TERM
152 /*! \brief Frees up the CUDA GPU used by the active context at the time of calling.
154 * If \c deviceInfo is nullptr, then it is understood that no device
155 * was selected so no context is active to be freed. Otherwise, the
156 * context is explicitly destroyed and therefore all data uploaded to
157 * the GPU is lost. This must only be called when none of this data is
158 * required anymore, because subsequent attempts to free memory
159 * associated with the context will otherwise fail.
161 * Calls gmx_warning upon errors.
163 * \param[in] deviceInfo device info of the GPU to clean up for
165 * \returns true if no error occurs during the freeing.
168 void free_gpu(const gmx_device_info_t *CUDA_FUNC_ARGUMENT(deviceInfo)) CUDA_FUNC_TERM
170 /*! \brief Return a pointer to the device info for \c deviceId
172 * \param[in] gpu_info GPU info of all detected devices in the system.
173 * \param[in] deviceId ID for the GPU device requested.
175 * \returns Pointer to the device info for \c deviceId.
178 gmx_device_info_t *getDeviceInfo(const gmx_gpu_info_t &GPU_FUNC_ARGUMENT(gpu_info),
179 int GPU_FUNC_ARGUMENT(deviceId)) GPU_FUNC_TERM_WITH_RETURN(NULL)
181 /*! \brief Returns the device ID of the CUDA GPU currently in use.
183 * The GPU used is the one that is active at the time of the call in the active context.
185 * \returns device ID of the GPU in use at the time of the call
188 int get_current_cuda_gpu_device_id(void) CUDA_FUNC_TERM_WITH_RETURN(-1)
190 /*! \brief Formats and returns a device information string for a given GPU.
192 * Given an index *directly* into the array of available GPUs (gpu_dev)
193 * returns a formatted info string for the respective GPU which includes
194 * ID, name, compute capability, and detection status.
196 * \param[out] s pointer to output string (has to be allocated externally)
197 * \param[in] gpu_info Information about detected GPUs
198 * \param[in] index an index *directly* into the array of available GPUs
201 void get_gpu_device_info_string(char *GPU_FUNC_ARGUMENT(s),
202 const struct gmx_gpu_info_t &GPU_FUNC_ARGUMENT(gpu_info),
203 int GPU_FUNC_ARGUMENT(index)) GPU_FUNC_TERM
205 /*! \brief Returns the size of the gpu_dev_info struct.
207 * The size of gpu_dev_info can be used for allocation and communication.
209 * \returns size in bytes of gpu_dev_info
212 size_t sizeof_gpu_dev_info(void) GPU_FUNC_TERM_WITH_RETURN(0)
214 /*! \brief Returns a pointer *ptr to page-locked memory of size nbytes.
216 * The allocated memory is suitable to be used for data transfers between host
218 * Error handling should be done within this function.
220 typedef void gmx_host_alloc_t (void **ptr, size_t nbytes);
222 /*! \brief Frees page-locked memory pointed to by *ptr.
224 * NULL should not be passed to this function.
226 typedef void gmx_host_free_t (void *ptr);
228 /*! \brief Set page-locked memory allocation functions used by the GPU host. */
229 void gpu_set_host_malloc_and_free(bool bUseGpuKernels,
230 gmx_host_alloc_t **nb_alloc,
231 gmx_host_free_t **nb_free);
235 /*! \brief Starts the GPU profiler if mdrun is being profiled.
237 * When a profiler run is in progress (based on the presence of the NVPROF_ID
238 * env. var.), the profiler is started to begin collecting data during the
239 * rest of the run (or until stopGpuProfiler is called).
241 * Note that this is implemented only for the CUDA API.
244 void startGpuProfiler(void) CUDA_FUNC_TERM
247 /*! \brief Resets the GPU profiler if mdrun is being profiled.
249 * When a profiler run is in progress (based on the presence of the NVPROF_ID
250 * env. var.), the profiler data is restet in order to eliminate the data collected
251 * from the preceding part fo the run.
253 * This function should typically be called at the mdrun counter reset time.
255 * Note that this is implemented only for the CUDA API.
258 void resetGpuProfiler(void) CUDA_FUNC_TERM
261 /*! \brief Stops the CUDA profiler if mdrun is being profiled.
263 * This function can be called at cleanup when skipping recording
264 * recording subsequent API calls from being traces/profiled is desired,
265 * e.g. before uninitialization.
267 * Note that this is implemented only for the CUDA API.
270 void stopGpuProfiler(void) CUDA_FUNC_TERM
272 //! Tells whether the host buffer was pinned for non-blocking transfers. Only implemented for CUDA.
274 bool isHostMemoryPinned(const void *CUDA_FUNC_ARGUMENT(h_ptr)) CUDA_FUNC_TERM_WITH_RETURN(false)