src/gromacs/gpu_utils/gpu_utils.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   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.
  10  *
  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.
  15  *
  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.
  20  *
  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.
  25  *
  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.
  33  *
  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.
  36  */
  37 /*! \libinternal \file
  38  *  \brief Declare functions for detection and initialization for GPU devices.
  39  *
  40  *  \author Szilard Pall <pall.szilard@gmail.com>
  41  *  \author Mark Abraham <mark.j.abraham@gmail.com>
  42  *
  43  *  \inlibraryapi
  44  */
  45 #ifndef GMX_GPU_UTILS_GPU_UTILS_H
  46 #define GMX_GPU_UTILS_GPU_UTILS_H
  47
  48 #include <cstdio>
  49
  50 #include <vector>
  51
  52 #include "gromacs/gpu_utils/gpu_macros.h"
  53 #include "gromacs/utility/basedefinitions.h"
  54
  55 struct gmx_device_info_t;
  56 struct gmx_gpu_info_t;
  57
  58 namespace gmx
  59 {
  60 class MDLogger;
  61 }
  62
  63 //! Enum which is only used to describe transfer calls at the moment
  64 enum class GpuApiCallBehavior
  65 {
  66     Sync,
  67     Async
  68 };
  69
  70 /*! \brief Detect all GPUs in the system.
  71  *
  72  *  Will detect every GPU supported by the device driver in use. Also
  73  *  check for the compatibility of each and fill the gpu_info->gpu_dev array
  74  *  with the required information on each the device: ID, device properties,
  75  *  status.
  76  *
  77  *  \param[in] gpu_info    pointer to structure holding GPU information.
  78  *  \param[out] err_str    The error message of any GPU API error that caused
  79  *                         the detection to fail (if there was any). The memory
  80  *                         the pointer points to should be managed externally.
  81  *  \returns               non-zero if the detection encountered a failure, zero otherwise.
  82  */
  83 GPU_FUNC_QUALIFIER
  84 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)
  85
  86 /*! \brief Return a container of the detected GPUs that are compatible.
  87  *
  88  * This function filters the result of the detection for compatible
  89  * GPUs, based on the previously run compatibility tests.
  90  *
  91  * \param[in]     gpu_info    Information detected about GPUs, including compatibility.
  92  * \return                    vector of IDs of GPUs already recorded as compatible */
  93 std::vector<int> getCompatibleGpus(const gmx_gpu_info_t &gpu_info);
  94
  95 /*! \brief Return a string describing how compatible the GPU with given \c index is.
  96  *
  97  * \param[in]   gpu_info    Information about detected GPUs
  98  * \param[in]   index       index of GPU to ask about
  99  * \returns                 A null-terminated C string describing the compatibility status, useful for error messages.
 100  */
 101 const char *getGpuCompatibilityDescription(const gmx_gpu_info_t &GPU_FUNC_ARGUMENT(gpu_info),
 102                                            int GPU_FUNC_ARGUMENT(index));
 103
 104 /*! \brief Frees the gpu_dev and dev_use array fields of \p gpu_info.
 105  *
 106  * \param[in]    gpu_info    pointer to structure holding GPU information
 107  */
 108 GPU_FUNC_QUALIFIER
 109 void free_gpu_info(const struct gmx_gpu_info_t *GPU_FUNC_ARGUMENT(gpu_info)) GPU_FUNC_TERM
 110
 111 /*! \brief Initializes the GPU described by \c deviceInfo.
 112  *
 113  * TODO Doxygen complains about these - probably a Doxygen bug, since
 114  * the patterns here are the same as elsewhere in this header.
 115  *
 116  *  param[in]    mdlog        log file to write to
 117  * \param[inout] deviceInfo   device info of the GPU to initialize
 118  *
 119  * Issues a fatal error for any critical errors that occur during
 120  * initialization.
 121  */
 122 GPU_FUNC_QUALIFIER
 123 void init_gpu(const gmx::MDLogger &GPU_FUNC_ARGUMENT(mdlog),
 124               gmx_device_info_t *GPU_FUNC_ARGUMENT(deviceInfo)) GPU_FUNC_TERM
 125
 126 /*! \brief Frees up the CUDA GPU used by the active context at the time of calling.
 127  *
 128  * The context is explicitly destroyed and therefore all data uploaded to the GPU
 129  * is lost. This should only be called when none of this data is required anymore.
 130  *
 131  * Calls gmx_warning upon errors.
 132  *
 133  * \param[in]  deviceInfo   device info of the GPU to clean up for
 134  *
 135  * \returns                 true if no error occurs during the freeing.
 136  */
 137 CUDA_FUNC_QUALIFIER
 138 void free_gpu(const gmx_device_info_t *CUDA_FUNC_ARGUMENT(deviceInfo)) CUDA_FUNC_TERM
 139
 140 /*! \brief Return a pointer to the device info for \c deviceId
 141  *
 142  * \param[in] gpu_info      GPU info of all detected devices in the system.
 143  * \param[in] deviceId      ID for the GPU device requested.
 144  *
 145  * \returns                 Pointer to the device info for \c deviceId.
 146  */
 147 GPU_FUNC_QUALIFIER
 148 gmx_device_info_t *getDeviceInfo(const gmx_gpu_info_t &GPU_FUNC_ARGUMENT(gpu_info),
 149                                  int GPU_FUNC_ARGUMENT(deviceId)) GPU_FUNC_TERM_WITH_RETURN(NULL)
 150
 151 /*! \brief Returns the device ID of the CUDA GPU currently in use.
 152  *
 153  * The GPU used is the one that is active at the time of the call in the active context.
 154  *
 155  * \returns                 device ID of the GPU in use at the time of the call
 156  */
 157 CUDA_FUNC_QUALIFIER
 158 int get_current_cuda_gpu_device_id(void) CUDA_FUNC_TERM_WITH_RETURN(-1)
 159
 160 /*! \brief Formats and returns a device information string for a given GPU.
 161  *
 162  * Given an index *directly* into the array of available GPUs (gpu_dev)
 163  * returns a formatted info string for the respective GPU which includes
 164  * ID, name, compute capability, and detection status.
 165  *
 166  * \param[out]  s           pointer to output string (has to be allocated externally)
 167  * \param[in]   gpu_info    Information about detected GPUs
 168  * \param[in]   index       an index *directly* into the array of available GPUs
 169  */
 170 GPU_FUNC_QUALIFIER
 171 void get_gpu_device_info_string(char *GPU_FUNC_ARGUMENT(s),
 172                                 const struct gmx_gpu_info_t &GPU_FUNC_ARGUMENT(gpu_info),
 173                                 int GPU_FUNC_ARGUMENT(index)) GPU_FUNC_TERM
 174
 175 /*! \brief Returns the size of the gpu_dev_info struct.
 176  *
 177  * The size of gpu_dev_info can be used for allocation and communication.
 178  *
 179  * \returns                 size in bytes of gpu_dev_info
 180  */
 181 GPU_FUNC_QUALIFIER
 182 size_t sizeof_gpu_dev_info(void) GPU_FUNC_TERM_WITH_RETURN(0)
 183
 184 /*! \brief Returns a pointer *ptr to page-locked memory of size nbytes.
 185  *
 186  * The allocated memory is suitable to be used for data transfers between host
 187  * and GPU.
 188  * Error handling should be done within this function.
 189  */
 190 typedef void gmx_host_alloc_t (void **ptr, size_t nbytes);
 191
 192 /*! \brief Frees page-locked memory pointed to by *ptr.
 193  *
 194  * NULL should not be passed to this function.
 195  */
 196 typedef void gmx_host_free_t (void *ptr);
 197
 198 /*! \brief Set page-locked memory allocation functions used by the GPU host. */
 199 void gpu_set_host_malloc_and_free(bool               bUseGpuKernels,
 200                                   gmx_host_alloc_t **nb_alloc,
 201                                   gmx_host_free_t  **nb_free);
 202
 203
 204
 205 /*! \brief Starts the GPU profiler if mdrun is being profiled.
 206  *
 207  *  When a profiler run is in progress (based on the presence of the NVPROF_ID
 208  *  env. var.), the profiler is started to begin collecting data during the
 209  *  rest of the run (or until stopGpuProfiler is called).
 210  *
 211  *  Note that this is implemented only for the CUDA API.
 212  */
 213 CUDA_FUNC_QUALIFIER
 214 void startGpuProfiler(void) CUDA_FUNC_TERM
 215
 216
 217 /*! \brief Resets the GPU profiler if mdrun is being profiled.
 218  *
 219  * When a profiler run is in progress (based on the presence of the NVPROF_ID
 220  * env. var.), the profiler data is restet in order to eliminate the data collected
 221  * from the preceding part fo the run.
 222  *
 223  * This function should typically be called at the mdrun counter reset time.
 224  *
 225  * Note that this is implemented only for the CUDA API.
 226  */
 227 CUDA_FUNC_QUALIFIER
 228 void resetGpuProfiler(void) CUDA_FUNC_TERM
 229
 230
 231 /*! \brief Stops the CUDA profiler if mdrun is being profiled.
 232  *
 233  *  This function can be called at cleanup when skipping recording
 234  *  recording subsequent API calls from being traces/profiled is desired,
 235  *  e.g. before uninitialization.
 236  *
 237  *  Note that this is implemented only for the CUDA API.
 238  */
 239 CUDA_FUNC_QUALIFIER
 240 void stopGpuProfiler(void) CUDA_FUNC_TERM
 241
 242 //! Tells whether the host buffer was pinned for non-blocking transfers. Only implemented for CUDA.
 243 CUDA_FUNC_QUALIFIER
 244 bool isHostMemoryPinned(void *CUDA_FUNC_ARGUMENT(h_ptr)) CUDA_FUNC_TERM_WITH_RETURN(false)
 245
 246 #endif