2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2015,2016,2017,2018, 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.
36 * \brief Defines functionality for deciding whether tasks will run on GPUs.
38 * \author Mark Abraham <mark.j.abraham@gmail.com>
39 * \ingroup module_taskassignment
44 #include "decidegpuusage.h"
54 #include "gromacs/hardware/cpuinfo.h"
55 #include "gromacs/hardware/detecthardware.h"
56 #include "gromacs/hardware/hardwaretopology.h"
57 #include "gromacs/hardware/hw_info.h"
58 #include "gromacs/mdlib/gmx_omp_nthreads.h"
59 #include "gromacs/mdlib/nb_verlet.h"
60 #include "gromacs/mdtypes/commrec.h"
61 #include "gromacs/mdtypes/inputrec.h"
62 #include "gromacs/mdtypes/md_enums.h"
63 #include "gromacs/taskassignment/taskassignment.h"
64 #include "gromacs/topology/topology.h"
65 #include "gromacs/utility/baseversion.h"
66 #include "gromacs/utility/exceptions.h"
67 #include "gromacs/utility/fatalerror.h"
68 #include "gromacs/utility/gmxassert.h"
69 #include "gromacs/utility/logger.h"
70 #include "gromacs/utility/stringutil.h"
79 //! Helper variable to localise the text of an often repeated message.
80 const char * g_specifyEverythingFormatString =
81 "When you use mdrun -gputasks, %s must be set to non-default "
82 "values, so that the device IDs can be interpreted correctly."
83 #if GMX_GPU != GMX_GPU_NONE
84 " If you simply want to restrict which GPUs are used, then it is "
85 "better to use mdrun -gpu_id. Otherwise, setting the "
86 # if GMX_GPU == GMX_GPU_CUDA
87 "CUDA_VISIBLE_DEVICES"
88 # elif GMX_GPU == GMX_GPU_OPENCL
89 // Technically there is no portable way to do this offered by the
90 // OpenCL standard, but the only current relevant case for GROMACS
91 // is AMD OpenCL, which offers this variable.
94 # error "Unreachable branch"
96 " environment variable in your bash profile or job "
97 "script may be more convenient."
104 decideWhetherToUseGpusForNonbondedWithThreadMpi(const TaskTarget nonbondedTarget,
105 const std::vector<int> &gpuIdsToUse,
106 const std::vector<int> &userGpuTaskAssignment,
107 const EmulateGpuNonbonded emulateGpuNonbonded,
108 const bool buildSupportsNonbondedOnGpu,
109 const bool usingVerletScheme,
110 const bool nonbondedOnGpuIsUseful,
111 const int numRanksPerSimulation)
113 // First, exclude all cases where we can't run NB on GPUs.
114 if (nonbondedTarget == TaskTarget::Cpu ||
115 emulateGpuNonbonded == EmulateGpuNonbonded::Yes ||
116 !usingVerletScheme ||
117 !nonbondedOnGpuIsUseful ||
118 !buildSupportsNonbondedOnGpu)
120 // If the user required NB on GPUs, we issue an error later.
124 // We now know that NB on GPUs makes sense, if we have any.
126 if (!userGpuTaskAssignment.empty())
128 // Specifying -gputasks requires specifying everything.
129 if (nonbondedTarget == TaskTarget::Auto ||
130 numRanksPerSimulation < 1)
132 GMX_THROW(InconsistentInputError(formatString(g_specifyEverythingFormatString, "-nb and -ntmpi")));
137 if (nonbondedTarget == TaskTarget::Gpu)
142 // Because this is thread-MPI, we already know about the GPUs that
143 // all potential ranks can use, and can use that in a global
144 // decision that will later be consistent.
145 auto haveGpus = !gpuIdsToUse.empty();
147 // If we get here, then the user permitted or required GPUs.
152 decideWhetherToUseGpusForPmeWithThreadMpi(const bool useGpuForNonbonded,
153 const TaskTarget pmeTarget,
154 const std::vector<int> &gpuIdsToUse,
155 const std::vector<int> &userGpuTaskAssignment,
156 const bool canUseGpuForPme,
157 const int numRanksPerSimulation,
158 const int numPmeRanksPerSimulation)
160 // First, exclude all cases where we can't run PME on GPUs.
161 if ((pmeTarget == TaskTarget::Cpu) ||
162 !useGpuForNonbonded ||
165 // PME can't run on a GPU. If the user required that, we issue
170 // We now know that PME on GPUs might make sense, if we have any.
172 if (!userGpuTaskAssignment.empty())
174 // Follow the user's choice of GPU task assignment, if we
175 // can. Checking that their IDs are for compatible GPUs comes
178 // Specifying -gputasks requires specifying everything.
179 if (pmeTarget == TaskTarget::Auto ||
180 numRanksPerSimulation < 1)
182 GMX_THROW(InconsistentInputError(formatString(g_specifyEverythingFormatString, "all of -nb, -pme, and -ntmpi")));
185 // PME on GPUs is only supported in a single case
186 if (pmeTarget == TaskTarget::Gpu)
188 if (((numRanksPerSimulation > 1) && (numPmeRanksPerSimulation == 0)) ||
189 (numPmeRanksPerSimulation > 1))
191 GMX_THROW(InconsistentInputError
192 ("When you run mdrun -pme gpu -gputasks, you must supply a PME-enabled .tpr file and use a single PME rank."));
197 // pmeTarget == TaskTarget::Auto
198 return numRanksPerSimulation == 1;
201 // Because this is thread-MPI, we already know about the GPUs that
202 // all potential ranks can use, and can use that in a global
203 // decision that will later be consistent.
205 if (pmeTarget == TaskTarget::Gpu)
207 if (((numRanksPerSimulation > 1) && (numPmeRanksPerSimulation == 0)) ||
208 (numPmeRanksPerSimulation > 1))
210 GMX_THROW(NotImplementedError
211 ("PME tasks were required to run on GPUs, but that is not implemented with "
212 "more than one PME rank. Use a single rank simulation, or a separate PME rank, "
213 "or permit PME tasks to be assigned to the CPU."));
218 if (numRanksPerSimulation == 1)
220 // PME can run well on a GPU shared with NB, and we permit
221 // mdrun to default to try that.
222 return !gpuIdsToUse.empty();
225 if (numRanksPerSimulation < 1)
227 // Full automated mode for thread-MPI (the default). PME can
228 // run well on a GPU shared with NB, and we permit mdrun to
229 // default to it if there is only one GPU available.
230 return (gpuIdsToUse.size() == 1);
233 // Not enough support for PME on GPUs for anything else
237 bool decideWhetherToUseGpusForNonbonded(const TaskTarget nonbondedTarget,
238 const std::vector<int> &userGpuTaskAssignment,
239 const EmulateGpuNonbonded emulateGpuNonbonded,
240 const bool buildSupportsNonbondedOnGpu,
241 const bool usingVerletScheme,
242 const bool nonbondedOnGpuIsUseful,
243 const bool gpusWereDetected)
245 if (nonbondedTarget == TaskTarget::Cpu)
247 if (!userGpuTaskAssignment.empty())
249 GMX_THROW(InconsistentInputError
250 ("A GPU task assignment was specified, but nonbonded interactions were "
251 "assigned to the CPU. Make no more than one of these choices."));
257 if (!buildSupportsNonbondedOnGpu && nonbondedTarget == TaskTarget::Gpu)
259 GMX_THROW(InconsistentInputError
260 ("Nonbonded interactions on the GPU were requested with -nb gpu, "
261 "but the GROMACS binary has been built without GPU support. "
262 "Either run without selecting GPU options, or recompile GROMACS "
263 "with GPU support enabled"));
266 // TODO refactor all these TaskTarget::Gpu checks into one place?
267 // e.g. use a subfunction that handles only the cases where
268 // TaskTargets are not Cpu?
269 if (emulateGpuNonbonded == EmulateGpuNonbonded::Yes)
271 if (nonbondedTarget == TaskTarget::Gpu)
273 GMX_THROW(InconsistentInputError
274 ("Nonbonded interactions on the GPU were required, which is inconsistent "
275 "with choosing emulation. Make no more than one of these choices."));
277 if (!userGpuTaskAssignment.empty())
279 GMX_THROW(InconsistentInputError
280 ("GPU ID usage was specified, as was GPU emulation. Make no more than one of these choices."));
286 if (!usingVerletScheme)
288 if (nonbondedTarget == TaskTarget::Gpu)
290 GMX_THROW(InconsistentInputError
291 ("Nonbonded interactions on the GPU were required, which requires using "
292 "the Verlet scheme. Either use the Verlet scheme, or do not require using GPUs."));
298 if (!nonbondedOnGpuIsUseful)
300 if (nonbondedTarget == TaskTarget::Gpu)
302 GMX_THROW(InconsistentInputError
303 ("Nonbonded interactions on the GPU were required, but not supported for these "
304 "simulation settings. Change your settings, or do not require using GPUs."));
310 if (!userGpuTaskAssignment.empty())
312 // Specifying -gputasks requires specifying everything.
313 if (nonbondedTarget == TaskTarget::Auto)
315 GMX_THROW(InconsistentInputError(formatString(g_specifyEverythingFormatString, "-nb and -ntmpi")));
321 if (nonbondedTarget == TaskTarget::Gpu)
323 // We still don't know whether it is an error if no GPUs are found
324 // because we don't know the duty of this rank, yet. For example,
325 // a node with only PME ranks and -pme cpu is OK if there are not
330 // If we get here, then the user permitted GPUs, which we should
331 // use for nonbonded interactions.
332 return gpusWereDetected;
335 bool decideWhetherToUseGpusForPme(const bool useGpuForNonbonded,
336 const TaskTarget pmeTarget,
337 const std::vector<int> &userGpuTaskAssignment,
338 const bool canUseGpuForPme,
339 const int numRanksPerSimulation,
340 const int numPmeRanksPerSimulation,
341 const bool gpusWereDetected)
343 if (pmeTarget == TaskTarget::Cpu)
348 if (!useGpuForNonbonded)
350 if (pmeTarget == TaskTarget::Gpu)
352 GMX_THROW(NotImplementedError
353 ("The PME on the GPU is only supported when nonbonded interactions run on GPUs also."));
358 if (!canUseGpuForPme)
360 if (pmeTarget == TaskTarget::Gpu)
362 // TODO Pass in the inputrec so we can give more help here?
363 GMX_THROW(NotImplementedError
364 ("The input simulation did not use PME in a way that is supported on the GPU."));
369 if (pmeTarget == TaskTarget::Cpu)
371 if (!userGpuTaskAssignment.empty())
373 GMX_THROW(InconsistentInputError
374 ("A GPU task assignment was specified, but PME interactions were "
375 "assigned to the CPU. Make no more than one of these choices."));
381 if (!userGpuTaskAssignment.empty())
383 // Specifying -gputasks requires specifying everything.
384 if (pmeTarget == TaskTarget::Auto)
386 GMX_THROW(InconsistentInputError(formatString(g_specifyEverythingFormatString, "all of -nb, -pme, and -ntmpi")));
392 // We still don't know whether it is an error if no GPUs are found
393 // because we don't know the duty of this rank, yet. For example,
394 // a node with only PME ranks and -pme cpu is OK if there are not
397 if (pmeTarget == TaskTarget::Gpu)
399 if (((numRanksPerSimulation > 1) && (numPmeRanksPerSimulation == 0)) ||
400 (numPmeRanksPerSimulation > 1))
402 GMX_THROW(NotImplementedError
403 ("PME tasks were required to run on GPUs, but that is not implemented with "
404 "more than one PME rank. Use a single rank simulation, or a separate PME rank, "
405 "or permit PME tasks to be assigned to the CPU."));
410 // If we get here, then the user permitted GPUs.
411 if (numRanksPerSimulation == 1)
413 // PME can run well on a single GPU shared with NB when there
414 // is one rank, so we permit mdrun to try that if we have
416 return gpusWereDetected;
419 // Not enough support for PME on GPUs for anything else
423 bool decideWhetherToUseGpusForBonded(const bool useGpuForNonbonded,
424 const bool useGpuForPme,
425 const bool usingVerletScheme,
426 const TaskTarget bondedTarget,
427 const bool canUseGpuForBonded,
428 const bool usingLJPme,
429 const bool usingElecPmeOrEwald,
430 const int numPmeRanksPerSimulation,
431 const bool gpusWereDetected)
433 if (bondedTarget == TaskTarget::Cpu)
438 if (!usingVerletScheme)
440 if (bondedTarget == TaskTarget::Gpu)
442 GMX_THROW(InconsistentInputError
443 ("Bonded interactions on the GPU were required, which requires using "
444 "the Verlet scheme. Either use the Verlet scheme, or do not require using GPUs."));
450 if (!canUseGpuForBonded)
452 if (bondedTarget == TaskTarget::Gpu)
454 GMX_THROW(InconsistentInputError
455 ("Bonded interactions on the GPU were required, but not supported for these "
456 "simulation settings. Change your settings, or do not require using GPUs."));
462 if (!useGpuForNonbonded)
464 if (bondedTarget == TaskTarget::Gpu)
466 GMX_THROW(InconsistentInputError
467 ("Bonded interactions on the GPU were required, but this requires that "
468 "short-ranged non-bonded interactions are also run on the GPU. Change "
469 "your settings, or do not require using GPUs."));
475 // TODO If the bonded kernels do not get fused, then performance
476 // overheads might suggest alternative choices here.
478 if (bondedTarget == TaskTarget::Gpu)
480 // We still don't know whether it is an error if no GPUs are
485 // If we get here, then the user permitted GPUs, which we should
486 // use for bonded interactions if any were detected and the CPU
487 // is busy, for which we currently only check PME or Ewald.
488 // (It would be better to dynamically assign bondeds based on timings)
489 // Note that here we assume that the auto setting of PME ranks will not
490 // choose seperate PME ranks when nonBonded are assigned to the GPU.
491 bool usingOurCpuForPmeOrEwald = (usingLJPme || (usingElecPmeOrEwald && !useGpuForPme && numPmeRanksPerSimulation <= 0));
493 return gpusWereDetected && usingOurCpuForPmeOrEwald;