2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2017,2018,2019,2020,2021, 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.
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.
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.
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.
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.
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.
35 /*! \libinternal \file
36 * \brief Declares functionality for deciding whether tasks will run on GPUs.
38 * \author Mark Abraham <mark.j.abraham@gmail.com>
39 * \ingroup module_taskassignment
43 #ifndef GMX_TASKASSIGNMENT_DECIDEGPUUSAGE_H
44 #define GMX_TASKASSIGNMENT_DECIDEGPUUSAGE_H
51 enum class PmeRunMode;
58 //! Record where a compute task is targetted.
59 enum class TaskTarget : int
66 //! Help pass GPU-emulation parameters with type safety.
67 enum class EmulateGpuNonbonded : bool
69 //! Do not emulate GPUs.
76 * \brief Structure that holds boolean flags corresponding to the development
77 * features present enabled through environment variables.
80 struct DevelopmentFeatureFlags
82 //! True if the Buffer ops development feature is enabled
83 // TODO: when the trigger of the buffer ops offload is fully automated this should go away
84 bool enableGpuBufferOps = false;
85 //! If true, forces 'mdrun -update auto' default to 'gpu'
86 bool forceGpuUpdateDefault = false;
87 //! True if the GPU halo exchange development feature is enabled
88 bool enableGpuHaloExchange = false;
89 //! True if the PME PP direct communication GPU development feature is enabled
90 bool enableGpuPmePPComm = false;
96 /*! \brief Decide whether this thread-MPI simulation will run
97 * nonbonded tasks on GPUs.
99 * The number of GPU tasks and devices influences both the choice of
100 * the number of ranks, and checks upon any such choice made by the
101 * user. So we need to consider this before any automated choice of
102 * the number of thread-MPI ranks.
104 * \param[in] nonbondedTarget The user's choice for mdrun -nb for where to assign
105 * short-ranged nonbonded interaction tasks.
106 * \param[in] numDevicesToUse Number of compatible GPUs that the user permitted
108 * \param[in] userGpuTaskAssignment The user-specified assignment of GPU tasks to device IDs.
109 * \param[in] emulateGpuNonbonded Whether we will emulate GPU calculation of nonbonded
111 * \param[in] buildSupportsNonbondedOnGpu Whether GROMACS was built with GPU support.
112 * \param[in] nonbondedOnGpuIsUseful Whether computing nonbonded interactions on a GPU is
113 * useful for this calculation.
114 * \param[in] numRanksPerSimulation The number of ranks in each simulation.
116 * \returns Whether the simulation will run nonbonded tasks on GPUs.
118 * \throws std::bad_alloc If out of memory
119 * InconsistentInputError If the user requirements are inconsistent. */
120 bool decideWhetherToUseGpusForNonbondedWithThreadMpi(TaskTarget nonbondedTarget,
122 const std::vector<int>& userGpuTaskAssignment,
123 EmulateGpuNonbonded emulateGpuNonbonded,
124 bool buildSupportsNonbondedOnGpu,
125 bool nonbondedOnGpuIsUseful,
126 int numRanksPerSimulation);
128 /*! \brief Decide whether this thread-MPI simulation will run
131 * The number of GPU tasks and devices influences both the choice of
132 * the number of ranks, and checks upon any such choice made by the
133 * user. So we need to consider this before any automated choice of
134 * the number of thread-MPI ranks.
136 * \param[in] useGpuForNonbonded Whether GPUs will be used for nonbonded interactions.
137 * \param[in] pmeTarget The user's choice for mdrun -pme for where to assign
138 * long-ranged PME nonbonded interaction tasks.
139 * \param[in] pmeFftTarget The user's choice for mdrun -pmefft for where to run FFT.
140 * \param[in] numDevicesToUse The number of compatible GPUs that the user permitted us to use.
141 * \param[in] userGpuTaskAssignment The user-specified assignment of GPU tasks to device IDs.
142 * \param[in] hardwareInfo Hardware information
143 * \param[in] inputrec The user input
144 * \param[in] numRanksPerSimulation The number of ranks in each simulation.
145 * \param[in] numPmeRanksPerSimulation The number of PME ranks in each simulation.
147 * \returns Whether the simulation will run PME tasks on GPUs.
149 * \throws std::bad_alloc If out of memory
150 * InconsistentInputError If the user requirements are inconsistent. */
151 bool decideWhetherToUseGpusForPmeWithThreadMpi(bool useGpuForNonbonded,
152 TaskTarget pmeTarget,
153 TaskTarget pmeFftTarget,
155 const std::vector<int>& userGpuTaskAssignment,
156 const gmx_hw_info_t& hardwareInfo,
157 const t_inputrec& inputrec,
158 int numRanksPerSimulation,
159 int numPmeRanksPerSimulation);
161 /*! \brief Decide whether the simulation will try to run nonbonded
164 * The final decision cannot be made until after the duty of the rank
165 * is known. But we need to know if nonbonded will run on GPUs for
166 * setting up DD (particularly rlist) and determining duty. If the
167 * user requires GPUs for the tasks of that duty, then it will be an
168 * error when none are found.
170 * With thread-MPI, calls have been made to
171 * decideWhetherToUseGpusForNonbondedWithThreadMpi() and
172 * decideWhetherToUseGpusForPmeWithThreadMpi() to help determine
173 * the number of ranks and run some checks, but the final
174 * decision is made in this routine, along with many more
175 * consistency checks.
177 * \param[in] nonbondedTarget The user's choice for mdrun -nb for where to assign short-ranged nonbonded interaction tasks.
178 * \param[in] userGpuTaskAssignment The user-specified assignment of GPU tasks to device IDs.
179 * \param[in] emulateGpuNonbonded Whether we will emulate GPU calculation of nonbonded interactions.
180 * \param[in] buildSupportsNonbondedOnGpu Whether GROMACS was build with GPU support.
181 * \param[in] nonbondedOnGpuIsUseful Whether computing nonbonded interactions on a GPU is useful for this calculation.
182 * \param[in] gpusWereDetected Whether compatible GPUs were detected on any node.
184 * \returns Whether the simulation will run nonbonded and PME tasks, respectively, on GPUs.
186 * \throws std::bad_alloc If out of memory
187 * InconsistentInputError If the user requirements are inconsistent. */
188 bool decideWhetherToUseGpusForNonbonded(TaskTarget nonbondedTarget,
189 const std::vector<int>& userGpuTaskAssignment,
190 EmulateGpuNonbonded emulateGpuNonbonded,
191 bool buildSupportsNonbondedOnGpu,
192 bool nonbondedOnGpuIsUseful,
193 bool gpusWereDetected);
195 /*! \brief Decide whether the simulation will try to run tasks of
196 * different types on GPUs.
198 * The final decision cannot be made until after the duty of the rank
199 * is known. But we need to know if nonbonded will run on GPUs for
200 * setting up DD (particularly rlist) and determining duty. If the
201 * user requires GPUs for the tasks of that duty, then it will be an
202 * error when none are found.
204 * With thread-MPI, calls have been made to
205 * decideWhetherToUseGpusForNonbondedWithThreadMpi() and
206 * decideWhetherToUseGpusForPmeWithThreadMpi() to help determine
207 * the number of ranks and run some checks, but the final
208 * decision is made in this routine, along with many more
209 * consistency checks.
211 * \param[in] useGpuForNonbonded Whether GPUs will be used for nonbonded interactions.
212 * \param[in] pmeTarget The user's choice for mdrun -pme for where to assign long-ranged PME nonbonded interaction tasks.
213 * \param[in] pmeFftTarget The user's choice for mdrun -pmefft for where to do FFT for PME.
214 * \param[in] userGpuTaskAssignment The user-specified assignment of GPU tasks to device IDs.
215 * \param[in] hardwareInfo Hardware information
216 * \param[in] inputrec The user input
217 * \param[in] numRanksPerSimulation The number of ranks in each simulation.
218 * \param[in] numPmeRanksPerSimulation The number of PME ranks in each simulation.
219 * \param[in] gpusWereDetected Whether compatible GPUs were detected on any node.
221 * \returns Whether the simulation will run nonbonded and PME tasks, respectively, on GPUs.
223 * \throws std::bad_alloc If out of memory
224 * InconsistentInputError If the user requirements are inconsistent. */
225 bool decideWhetherToUseGpusForPme(bool useGpuForNonbonded,
226 TaskTarget pmeTarget,
227 TaskTarget pmeFftTarget,
228 const std::vector<int>& userGpuTaskAssignment,
229 const gmx_hw_info_t& hardwareInfo,
230 const t_inputrec& inputrec,
231 int numRanksPerSimulation,
232 int numPmeRanksPerSimulation,
233 bool gpusWereDetected);
235 /*! \brief Determine PME run mode.
237 * Given the PME task assignment in \p useGpuForPme and the user-provided
238 * FFT task target in \p pmeFftTarget, returns a PME run mode for the
239 * current run. It also checks the compatibility of the two.
241 * \note Aborts the run upon incompatible values of \p useGpuForPme and \p pmeFftTarget.
243 * \param[in] useGpuForPme PME task assignment, true if PME task is mapped to the GPU.
244 * \param[in] pmeFftTarget The user's choice for -pmefft for where to assign the FFT
245 * work of the PME task. \param[in] inputrec The user input record
247 PmeRunMode determinePmeRunMode(bool useGpuForPme, const TaskTarget& pmeFftTarget, const t_inputrec& inputrec);
249 /*! \brief Decide whether the simulation will try to run bonded tasks on GPUs.
251 * \param[in] useGpuForNonbonded Whether GPUs will be used for nonbonded interactions.
252 * \param[in] useGpuForPme Whether GPUs will be used for PME interactions.
253 * \param[in] bondedTarget The user's choice for mdrun -bonded for where to assign tasks.
254 * \param[in] inputrec The user input.
255 * \param[in] mtop The global topology.
256 * \param[in] numPmeRanksPerSimulation The number of PME ranks in each simulation, can be -1 for auto.
257 * \param[in] gpusWereDetected Whether compatible GPUs were detected on any node.
259 * \returns Whether the simulation will run bondeded tasks on GPUs.
261 * \throws std::bad_alloc If out of memory
262 * InconsistentInputError If the user requirements are inconsistent. */
263 bool decideWhetherToUseGpusForBonded(bool useGpuForNonbonded,
265 TaskTarget bondedTarget,
266 const t_inputrec& inputrec,
267 const gmx_mtop_t& mtop,
268 int numPmeRanksPerSimulation,
269 bool gpusWereDetected);
271 /*! \brief Decide whether to use GPU for update.
273 * \param[in] isDomainDecomposition Whether there more than one domain.
274 * \param[in] useUpdateGroups If the constraints can be split across domains.
275 * \param[in] pmeRunMode PME running mode: CPU, GPU or mixed.
276 * \param[in] havePmeOnlyRank If there is a PME-only rank in the simulation.
277 * \param[in] useGpuForNonbonded Whether GPUs will be used for nonbonded interactions.
278 * \param[in] updateTarget User choice for running simulation on GPU.
279 * \param[in] gpusWereDetected Whether compatible GPUs were detected on any node.
280 * \param[in] inputrec The user input.
281 * \param[in] mtop The global topology.
282 * \param[in] useEssentialDynamics If essential dynamics is active.
283 * \param[in] doOrientationRestraints If orientation restraints are enabled.
284 * \param[in] useReplicaExchange If this is a REMD simulation.
285 * \param[in] haveFrozenAtoms If this simulation has frozen atoms (see Issue #3920).
286 * \param[in] doRerun It this is a rerun.
287 * \param[in] devFlags GPU development / experimental feature flags.
288 * \param[in] mdlog MD logger.
290 * \returns Whether complete simulation can be run on GPU.
291 * \throws std::bad_alloc If out of memory
292 * InconsistentInputError If the user requirements are inconsistent.
294 bool decideWhetherToUseGpuForUpdate(bool isDomainDecomposition,
295 bool useUpdateGroups,
296 PmeRunMode pmeRunMode,
297 bool havePmeOnlyRank,
298 bool useGpuForNonbonded,
299 TaskTarget updateTarget,
300 bool gpusWereDetected,
301 const t_inputrec& inputrec,
302 const gmx_mtop_t& mtop,
303 bool useEssentialDynamics,
304 bool doOrientationRestraints,
305 bool useReplicaExchange,
306 bool haveFrozenAtoms,
308 const DevelopmentFeatureFlags& devFlags,
309 const gmx::MDLogger& mdlog);
312 /*! \brief Decide whether to use GPU for halo exchange.
314 * \param[in] devFlags GPU development / experimental feature flags.
315 * \param[in] havePPDomainDecomposition Whether PP domain decomposition is in use.
316 * \param[in] useGpuForNonbonded Whether GPUs will be used for nonbonded interactions.
317 * \param[in] useModularSimulator Whether modularsimulator is in use.
318 * \param[in] doRerun Whether this is a rerun.
319 * \param[in] haveEnergyMinimization Whether energy minimization is in use.
321 * \returns Whether halo exchange can be run on GPU.
323 bool decideWhetherToUseGpuForHalo(const DevelopmentFeatureFlags& devFlags,
324 bool havePPDomainDecomposition,
325 bool useGpuForNonbonded,
326 bool useModularSimulator,
328 bool haveEnergyMinimization);