src/gromacs/taskassignment/usergpuids.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2017,2018,2019,2020, 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.
   8  *
   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.
  13  *
  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.
  18  *
  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.
  23  *
  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.
  31  *
  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.
  34  */
  35 /*! \defgroup module_taskassignment Assigning simulation tasks to hardware (taskassignment)
  36  * \ingroup group_mdrun
  37  * \brief Provides code that manages assignment of simulation tasks to hardware.
  38  */
  39 /*! \libinternal
  40  * \file
  41  * \brief Declares routines for handling user-specified GPU IDs.
  42  *
  43  * \author Mark Abraham <mark.j.abraham@gmail.com>
  44  * \ingroup module_taskassignment
  45  * \inlibraryapi
  46  */
  47 #ifndef GMX_TASKASSIGNMENT_USERGPUIDS_H
  48 #define GMX_TASKASSIGNMENT_USERGPUIDS_H
  49
  50 #include <cstddef>
  51
  52 #include <memory>
  53 #include <string>
  54 #include <vector>
  55
  56 #include "gromacs/utility/arrayref.h"
  57
  58 struct DeviceInformation;
  59
  60 namespace gmx
  61 {
  62
  63 /*! \brief Parse a GPU ID specifier string into a container describing device IDs exposed to the run.
  64  *
  65  * \param[in]   gpuIdString  String like "013" or "0,1,3" typically
  66  *                           supplied by the user to mdrun -gpu_id.
  67  *                           Must contain only unique decimal digits, or only decimal
  68  *                           digits separated by comma delimiters. A terminal
  69  *                           comma is accceptable (and required to specify a
  70  *                           single ID that is larger than 9).
  71  *
  72  * \returns  A vector of unique GPU IDs.
  73  *
  74  * \throws   std::bad_alloc     If out of memory.
  75  *           InvalidInputError  If an invalid character is found (ie not a digit or ',') or if
  76  *                              identifiers are duplicated in the specifier list.
  77  */
  78 std::vector<int> parseUserGpuIdString(const std::string& gpuIdString);
  79
  80 /*! \brief Implement GPU ID selection by returning the available GPU
  81  * IDs on this physical node that are compatible.
  82  *
  83  * If the string supplied by the user is empty, then return the IDs of
  84  * all compatible GPUs on this physical node. Otherwise, check the
  85  * user specified compatible GPUs and return their IDs.
  86  *
  87  * \param[in]  deviceInfoList         Information on the GPUs on this physical node.
  88  * \param[in]  gpuIdsAvailableString  String like "013" or "0,1,3" typically
  89  *                                    supplied by the user to mdrun -gpu_id.
  90  *                                    Must contain only unique decimal digits, or only decimal
  91  *                                    digits separated by comma delimiters. A terminal
  92  *                                    comma is accceptable (and required to specify a
  93  *                                    single ID that is larger than 9).
  94  *
  95  * \returns  A vector of unique compatible GPU IDs on this physical node.
  96  *
  97  * \throws   std::bad_alloc     If out of memory.
  98  *           InvalidInputError  If an invalid character is found (ie not a digit or ',') or if
  99  *                              identifiers are duplicated in the specifier list.
 100  *           InvalidInputError  If gpuIdsAvailableString specifies GPU IDs that are
 101  *                              not compatible.
 102  */
 103 std::vector<int> makeGpuIdsToUse(const std::vector<std::unique_ptr<DeviceInformation>>& deviceInfoList,
 104                                  const std::string& gpuIdsAvailableString);
 105
 106 /*! \brief Parse a GPU ID specifier string into a container describing device ID to task mapping.
 107  *
 108  * \param[in]   gpuIdString  String like "0011" or "0,0,1,1" typically
 109  *                           supplied by the user to mdrun -gputasks.
 110  *                           Must contain only decimal digits, or only decimal
 111  *                           digits separated by comma delimiters. A terminal
 112  *                           comma is accceptable (and required to specify a
 113  *                           single ID that is larger than 9).
 114  *
 115  * \returns  A vector of GPU IDs.
 116  *
 117  * \throws   std::bad_alloc     If out of memory.
 118  *           InvalidInputError  If an invalid character is found (ie not a digit or ',').
 119  */
 120 std::vector<int> parseUserTaskAssignmentString(const std::string& gpuIdString);
 121
 122
 123 /*! \brief Make a vector containing \c numGpuTasks IDs of the IDs found in \c compatibleGpus.
 124  *
 125  * \throws  std::bad_alloc          If out of memory
 126  *
 127  * \returns A sorted vector of IDs of compatible vectors, whose
 128  * length matches that of the number of GPU tasks required.
 129  */
 130 std::vector<int> makeGpuIds(ArrayRef<const int> compatibleGpus, size_t numGpuTasks);
 131
 132 /*! \brief Convert a container of GPU deviced IDs to a string that
 133  * can be used by gmx tune_pme as input to mdrun -gputasks.
 134  *
 135  * Produce a valid input for mdrun -gputasks that refers to the device
 136  * IDs in \c gpuIds but produces a mapping for \c
 137  * totalNumberOfTasks tasks. Note that gmx tune_pme does not
 138  * currently support filling mdrun -gputasks.
 139  *
 140  * \param[in]   gpuIds              Container of device IDs
 141  * \param[in]   totalNumberOfTasks  Total number of tasks for the output mapping produced by the returned string.
 142  *
 143  * \returns  A string that is suitable to pass to mdrun -gputasks.
 144  *
 145  * \throws   std::bad_alloc     If out of memory.
 146  */
 147 std::string makeGpuIdString(const std::vector<int>& gpuIds, int totalNumberOfTasks);
 148
 149 /*! \brief Check that all user-selected GPUs are compatible.
 150  *
 151  * Given the \c gpuIds and \c hardwareInfo, throw if
 152  * any selected GPUs is not compatible.
 153  *
 154  * The error is given with a suitable descriptive message, which will
 155  * have context if this check is done after the hardware detection
 156  * results have been reported to the user. However, note that only the
 157  * GPUs detected on the master rank are reported, because of the
 158  * existing limitations of that reporting.
 159  *
 160  * \todo Note that the selected GPUs can be different on each rank,
 161  * and the IDs of compatible GPUs can be different on each node, so
 162  * this routine ought to do communication to determine whether all
 163  * ranks are able to proceed. Currently this relies on the MPI runtime
 164  * to kill the other processes because GROMACS lacks the appropriate
 165  * infrastructure to do a good job of coordinating error messages and
 166  * behaviour across MPMD ranks and multiple simulations.
 167  *
 168  * \param[in]   deviceInfoList  Information on the GPUs on this physical node.
 169  * \param[in]   compatibleGpus  Vector of GPUs that are compatible
 170  * \param[in]   gpuIds          The GPU IDs selected by the user.
 171  *
 172  * \throws  std::bad_alloc          If out of memory
 173  *          InconsistentInputError  If the assigned GPUs are not valid
 174  */
 175 void checkUserGpuIds(const std::vector<std::unique_ptr<DeviceInformation>>& deviceInfoList,
 176                      const std::vector<int>&                                compatibleGpus,
 177                      const std::vector<int>&                                gpuIds);
 178
 179 } // namespace gmx
 180
 181 #endif