2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2015,2016,2017,2018,2019 by the GROMACS development team.
5 * Copyright (c) 2020, by the GROMACS development team, led by
6 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
7 * and including many others, as listed in the AUTHORS file in the
8 * top-level source directory and at http://www.gromacs.org.
10 * GROMACS is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public License
12 * as published by the Free Software Foundation; either version 2.1
13 * of the License, or (at your option) any later version.
15 * GROMACS is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with GROMACS; if not, see
22 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
23 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
25 * If you want to redistribute modifications to GROMACS, please
26 * consider that scientific software is very special. Version
27 * control is crucial - bugs must be traceable. We will be happy to
28 * consider code for inclusion in the official distribution, but
29 * derived work must not be called official GROMACS. Details are found
30 * in the README & COPYING files - if they are missing, get the
31 * official version at http://www.gromacs.org.
33 * To help us fund GROMACS development, we humbly ask that you cite
34 * the research papers on the package. Check out http://www.gromacs.org.
37 * \brief Defines functionality for deciding whether tasks will run on GPUs.
39 * \author Mark Abraham <mark.j.abraham@gmail.com>
40 * \ingroup module_taskassignment
45 #include "decidegpuusage.h"
55 #include "gromacs/ewald/pme.h"
56 #include "gromacs/hardware/cpuinfo.h"
57 #include "gromacs/hardware/detecthardware.h"
58 #include "gromacs/hardware/hardwaretopology.h"
59 #include "gromacs/hardware/hw_info.h"
60 #include "gromacs/mdlib/gmx_omp_nthreads.h"
61 #include "gromacs/mdlib/update_constrain_gpu.h"
62 #include "gromacs/mdtypes/commrec.h"
63 #include "gromacs/mdtypes/inputrec.h"
64 #include "gromacs/mdtypes/md_enums.h"
65 #include "gromacs/mdtypes/mdrunoptions.h"
66 #include "gromacs/pulling/pull.h"
67 #include "gromacs/taskassignment/taskassignment.h"
68 #include "gromacs/topology/mtop_util.h"
69 #include "gromacs/topology/topology.h"
70 #include "gromacs/utility/baseversion.h"
71 #include "gromacs/utility/exceptions.h"
72 #include "gromacs/utility/fatalerror.h"
73 #include "gromacs/utility/gmxassert.h"
74 #include "gromacs/utility/logger.h"
75 #include "gromacs/utility/stringutil.h"
84 //! Helper variable to localise the text of an often repeated message.
85 const char* g_specifyEverythingFormatString =
86 "When you use mdrun -gputasks, %s must be set to non-default "
87 "values, so that the device IDs can be interpreted correctly."
88 #if GMX_GPU != GMX_GPU_NONE
89 " If you simply want to restrict which GPUs are used, then it is "
90 "better to use mdrun -gpu_id. Otherwise, setting the "
91 # if GMX_GPU == GMX_GPU_CUDA
92 "CUDA_VISIBLE_DEVICES"
93 # elif GMX_GPU == GMX_GPU_OPENCL
94 // Technically there is no portable way to do this offered by the
95 // OpenCL standard, but the only current relevant case for GROMACS
96 // is AMD OpenCL, which offers this variable.
99 # error "Unreachable branch"
101 " environment variable in your bash profile or job "
102 "script may be more convenient."
108 bool decideWhetherToUseGpusForNonbondedWithThreadMpi(const TaskTarget nonbondedTarget,
109 const std::vector<int>& gpuIdsToUse,
110 const std::vector<int>& userGpuTaskAssignment,
111 const EmulateGpuNonbonded emulateGpuNonbonded,
112 const bool buildSupportsNonbondedOnGpu,
113 const bool nonbondedOnGpuIsUseful,
114 const int numRanksPerSimulation)
116 // First, exclude all cases where we can't run NB on GPUs.
117 if (nonbondedTarget == TaskTarget::Cpu || emulateGpuNonbonded == EmulateGpuNonbonded::Yes
118 || !nonbondedOnGpuIsUseful || !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 || numRanksPerSimulation < 1)
131 GMX_THROW(InconsistentInputError(
132 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.
151 bool decideWhetherToUseGpusForPmeWithThreadMpi(const bool useGpuForNonbonded,
152 const TaskTarget pmeTarget,
153 const std::vector<int>& gpuIdsToUse,
154 const std::vector<int>& userGpuTaskAssignment,
155 const gmx_hw_info_t& hardwareInfo,
156 const t_inputrec& inputrec,
157 const gmx_mtop_t& mtop,
158 const int numRanksPerSimulation,
159 const int numPmeRanksPerSimulation)
161 // First, exclude all cases where we can't run PME on GPUs.
162 if ((pmeTarget == TaskTarget::Cpu) || !useGpuForNonbonded || !pme_gpu_supports_build(nullptr)
163 || !pme_gpu_supports_hardware(hardwareInfo, nullptr)
164 || !pme_gpu_supports_input(inputrec, mtop, nullptr))
166 // PME can't run on a GPU. If the user required that, we issue
171 // We now know that PME on GPUs might make sense, if we have any.
173 if (!userGpuTaskAssignment.empty())
175 // Follow the user's choice of GPU task assignment, if we
176 // can. Checking that their IDs are for compatible GPUs comes
179 // Specifying -gputasks requires specifying everything.
180 if (pmeTarget == TaskTarget::Auto || numRanksPerSimulation < 1)
182 GMX_THROW(InconsistentInputError(
183 formatString(g_specifyEverythingFormatString, "all of -nb, -pme, and -ntmpi")));
186 // PME on GPUs is only supported in a single case
187 if (pmeTarget == TaskTarget::Gpu)
189 if (((numRanksPerSimulation > 1) && (numPmeRanksPerSimulation == 0))
190 || (numPmeRanksPerSimulation > 1))
192 GMX_THROW(InconsistentInputError(
193 "When you run mdrun -pme gpu -gputasks, you must supply a PME-enabled .tpr "
194 "file and use a single PME rank."));
199 // pmeTarget == TaskTarget::Auto
200 return numRanksPerSimulation == 1;
203 // Because this is thread-MPI, we already know about the GPUs that
204 // all potential ranks can use, and can use that in a global
205 // decision that will later be consistent.
207 if (pmeTarget == TaskTarget::Gpu)
209 if (((numRanksPerSimulation > 1) && (numPmeRanksPerSimulation == 0))
210 || (numPmeRanksPerSimulation > 1))
212 GMX_THROW(NotImplementedError(
213 "PME tasks were required to run on GPUs, but that is not implemented with "
214 "more than one PME rank. Use a single rank simulation, or a separate PME rank, "
215 "or permit PME tasks to be assigned to the CPU."));
220 if (numRanksPerSimulation == 1)
222 // PME can run well on a GPU shared with NB, and we permit
223 // mdrun to default to try that.
224 return !gpuIdsToUse.empty();
227 if (numRanksPerSimulation < 1)
229 // Full automated mode for thread-MPI (the default). PME can
230 // run well on a GPU shared with NB, and we permit mdrun to
231 // default to it if there is only one GPU available.
232 return (gpuIdsToUse.size() == 1);
235 // Not enough support for PME on GPUs for anything else
239 bool decideWhetherToUseGpusForNonbonded(const TaskTarget nonbondedTarget,
240 const std::vector<int>& userGpuTaskAssignment,
241 const EmulateGpuNonbonded emulateGpuNonbonded,
242 const bool buildSupportsNonbondedOnGpu,
243 const bool nonbondedOnGpuIsUseful,
244 const bool gpusWereDetected)
246 if (nonbondedTarget == TaskTarget::Cpu)
248 if (!userGpuTaskAssignment.empty())
250 GMX_THROW(InconsistentInputError(
251 "A GPU task assignment was specified, but nonbonded interactions were "
252 "assigned to the CPU. Make no more than one of these choices."));
258 if (!buildSupportsNonbondedOnGpu && nonbondedTarget == TaskTarget::Gpu)
260 GMX_THROW(InconsistentInputError(
261 "Nonbonded interactions on the GPU were requested with -nb gpu, "
262 "but the GROMACS binary has been built without GPU support. "
263 "Either run without selecting GPU options, or recompile GROMACS "
264 "with GPU support enabled"));
267 // TODO refactor all these TaskTarget::Gpu checks into one place?
268 // e.g. use a subfunction that handles only the cases where
269 // TaskTargets are not Cpu?
270 if (emulateGpuNonbonded == EmulateGpuNonbonded::Yes)
272 if (nonbondedTarget == TaskTarget::Gpu)
274 GMX_THROW(InconsistentInputError(
275 "Nonbonded interactions on the GPU were required, which is inconsistent "
276 "with choosing emulation. Make no more than one of these choices."));
278 if (!userGpuTaskAssignment.empty())
281 InconsistentInputError("GPU ID usage was specified, as was GPU emulation. Make "
282 "no more than one of these choices."));
288 if (!nonbondedOnGpuIsUseful)
290 if (nonbondedTarget == TaskTarget::Gpu)
292 GMX_THROW(InconsistentInputError(
293 "Nonbonded interactions on the GPU were required, but not supported for these "
294 "simulation settings. Change your settings, or do not require using GPUs."));
300 if (!userGpuTaskAssignment.empty())
302 // Specifying -gputasks requires specifying everything.
303 if (nonbondedTarget == TaskTarget::Auto)
305 GMX_THROW(InconsistentInputError(
306 formatString(g_specifyEverythingFormatString, "-nb and -ntmpi")));
312 if (nonbondedTarget == TaskTarget::Gpu)
314 // We still don't know whether it is an error if no GPUs are found
315 // because we don't know the duty of this rank, yet. For example,
316 // a node with only PME ranks and -pme cpu is OK if there are not
321 // If we get here, then the user permitted GPUs, which we should
322 // use for nonbonded interactions.
323 return gpusWereDetected;
326 bool decideWhetherToUseGpusForPme(const bool useGpuForNonbonded,
327 const TaskTarget pmeTarget,
328 const std::vector<int>& userGpuTaskAssignment,
329 const gmx_hw_info_t& hardwareInfo,
330 const t_inputrec& inputrec,
331 const gmx_mtop_t& mtop,
332 const int numRanksPerSimulation,
333 const int numPmeRanksPerSimulation,
334 const bool gpusWereDetected)
336 if (pmeTarget == TaskTarget::Cpu)
341 if (!useGpuForNonbonded)
343 if (pmeTarget == TaskTarget::Gpu)
345 GMX_THROW(NotImplementedError(
346 "PME on GPUs is only supported when nonbonded interactions run on GPUs also."));
352 if (!pme_gpu_supports_build(&message))
354 if (pmeTarget == TaskTarget::Gpu)
356 GMX_THROW(NotImplementedError("Cannot compute PME interactions on a GPU, because " + message));
360 if (!pme_gpu_supports_hardware(hardwareInfo, &message))
362 if (pmeTarget == TaskTarget::Gpu)
364 GMX_THROW(NotImplementedError("Cannot compute PME interactions on a GPU, because " + message));
368 if (!pme_gpu_supports_input(inputrec, mtop, &message))
370 if (pmeTarget == TaskTarget::Gpu)
372 GMX_THROW(NotImplementedError("Cannot compute PME interactions on a GPU, because " + message));
377 if (pmeTarget == TaskTarget::Cpu)
379 if (!userGpuTaskAssignment.empty())
381 GMX_THROW(InconsistentInputError(
382 "A GPU task assignment was specified, but PME interactions were "
383 "assigned to the CPU. Make no more than one of these choices."));
389 if (!userGpuTaskAssignment.empty())
391 // Specifying -gputasks requires specifying everything.
392 if (pmeTarget == TaskTarget::Auto)
394 GMX_THROW(InconsistentInputError(formatString(
395 g_specifyEverythingFormatString, "all of -nb, -pme, and -ntmpi"))); // TODO ntmpi?
401 // We still don't know whether it is an error if no GPUs are found
402 // because we don't know the duty of this rank, yet. For example,
403 // a node with only PME ranks and -pme cpu is OK if there are not
406 if (pmeTarget == TaskTarget::Gpu)
408 if (((numRanksPerSimulation > 1) && (numPmeRanksPerSimulation == 0))
409 || (numPmeRanksPerSimulation > 1))
411 GMX_THROW(NotImplementedError(
412 "PME tasks were required to run on GPUs, but that is not implemented with "
413 "more than one PME rank. Use a single rank simulation, or a separate PME rank, "
414 "or permit PME tasks to be assigned to the CPU."));
419 // If we get here, then the user permitted GPUs.
420 if (numRanksPerSimulation == 1)
422 // PME can run well on a single GPU shared with NB when there
423 // is one rank, so we permit mdrun to try that if we have
425 return gpusWereDetected;
428 // Not enough support for PME on GPUs for anything else
433 PmeRunMode determinePmeRunMode(const bool useGpuForPme, const TaskTarget& pmeFftTarget, const t_inputrec& inputrec)
435 if (!EEL_PME(inputrec.coulombtype))
437 return PmeRunMode::None;
442 if (pmeFftTarget == TaskTarget::Cpu)
444 return PmeRunMode::Mixed;
448 return PmeRunMode::GPU;
453 if (pmeFftTarget == TaskTarget::Gpu)
456 "Assigning FFTs to GPU requires PME to be assigned to GPU as well. With PME "
457 "on CPU you should not be using -pmefft.");
459 return PmeRunMode::CPU;
463 bool decideWhetherToUseGpusForBonded(const bool useGpuForNonbonded,
464 const bool useGpuForPme,
465 const TaskTarget bondedTarget,
466 const bool canUseGpuForBonded,
467 const bool usingLJPme,
468 const bool usingElecPmeOrEwald,
469 const int numPmeRanksPerSimulation,
470 const bool gpusWereDetected)
472 if (bondedTarget == TaskTarget::Cpu)
477 if (!canUseGpuForBonded)
479 if (bondedTarget == TaskTarget::Gpu)
481 GMX_THROW(InconsistentInputError(
482 "Bonded interactions on the GPU were required, but not supported for these "
483 "simulation settings. Change your settings, or do not require using GPUs."));
489 if (!useGpuForNonbonded)
491 if (bondedTarget == TaskTarget::Gpu)
493 GMX_THROW(InconsistentInputError(
494 "Bonded interactions on the GPU were required, but this requires that "
495 "short-ranged non-bonded interactions are also run on the GPU. Change "
496 "your settings, or do not require using GPUs."));
502 // TODO If the bonded kernels do not get fused, then performance
503 // overheads might suggest alternative choices here.
505 if (bondedTarget == TaskTarget::Gpu)
507 // We still don't know whether it is an error if no GPUs are
512 // If we get here, then the user permitted GPUs, which we should
513 // use for bonded interactions if any were detected and the CPU
514 // is busy, for which we currently only check PME or Ewald.
515 // (It would be better to dynamically assign bondeds based on timings)
516 // Note that here we assume that the auto setting of PME ranks will not
517 // choose seperate PME ranks when nonBonded are assigned to the GPU.
518 bool usingOurCpuForPmeOrEwald =
519 (usingLJPme || (usingElecPmeOrEwald && !useGpuForPme && numPmeRanksPerSimulation <= 0));
521 return gpusWereDetected && usingOurCpuForPmeOrEwald;
524 bool decideWhetherToUseGpuForUpdate(const bool forceGpuUpdateDefault,
525 const bool isDomainDecomposition,
526 const bool useUpdateGroups,
527 const PmeRunMode pmeRunMode,
528 const bool havePmeOnlyRank,
529 const bool useGpuForNonbonded,
530 const TaskTarget updateTarget,
531 const bool gpusWereDetected,
532 const t_inputrec& inputrec,
533 const gmx_mtop_t& mtop,
534 const bool useEssentialDynamics,
535 const bool doOrientationRestraints,
536 const bool useReplicaExchange,
538 const gmx::MDLogger& mdlog)
541 // '-update cpu' overrides the environment variable, '-update auto' does not
542 if (updateTarget == TaskTarget::Cpu || (updateTarget == TaskTarget::Auto && !forceGpuUpdateDefault))
547 const bool hasAnyConstraints = gmx_mtop_interaction_count(mtop, IF_CONSTRAINT) > 0;
549 std::string errorMessage;
551 if (isDomainDecomposition)
553 if (!forceGpuUpdateDefault)
555 errorMessage += "Domain decomposition is not supported.\n ";
557 else if (hasAnyConstraints && !useUpdateGroups)
560 "Domain decomposition is only supported with constraints when update groups "
561 "are used. This means constraining all bonds is not supported, except for "
562 "small molecules, and box sizes close to half the pair-list cutoff are not "
566 if (inputrec.eConstrAlg == econtSHAKE && hasAnyConstraints && gmx_mtop_ftype_count(mtop, F_CONSTR) > 0)
568 errorMessage += "SHAKE constraints are not supported.\n";
570 // Using the GPU-version of update if:
571 // 1. PME is on the GPU (there should be a copy of coordinates on GPU for PME spread), or
572 // 2. Non-bonded interactions are on the GPU.
573 if (pmeRunMode == PmeRunMode::CPU && !useGpuForNonbonded)
576 "Either PME or short-ranged non-bonded interaction tasks must run on the GPU.\n";
579 // If PME is active (i.e. not PmeRunMode::None), then GPU update requires
580 // either a single-rank run, or that PME runs fully on the GPU.
581 const bool pmeRunningOnCpu = (pmeRunMode == PmeRunMode::CPU || pmeRunMode == PmeRunMode::Mixed);
582 if (pmeRunningOnCpu && isDomainDecomposition)
584 errorMessage += "With domain decomposition, PME must run fully on the GPU.\n";
586 if (pmeRunningOnCpu && havePmeOnlyRank)
588 errorMessage += "With separate PME rank(s), PME must run fully on the GPU.\n";
591 if (!gpusWereDetected)
593 errorMessage += "Compatible GPUs must have been found.\n";
595 if (GMX_GPU != GMX_GPU_CUDA)
597 errorMessage += "Only a CUDA build is supported.\n";
599 if (inputrec.eI != eiMD)
601 errorMessage += "Only the md integrator is supported.\n";
603 if (inputrec.etc == etcNOSEHOOVER)
605 errorMessage += "Nose-Hoover temperature coupling is not supported.\n";
607 if (!(inputrec.epc == epcNO || inputrec.epc == epcPARRINELLORAHMAN || inputrec.epc == epcBERENDSEN))
609 errorMessage += "Only Parrinello-Rahman and Berendsen pressure coupling are supported.\n";
611 if (EEL_PME_EWALD(inputrec.coulombtype) && inputrec.epsilon_surface != 0)
613 // The graph is needed, but not supported
614 errorMessage += "Ewald surface correction is not supported.\n";
616 if (gmx_mtop_interaction_count(mtop, IF_VSITE) > 0)
618 errorMessage += "Virtual sites are not supported.\n";
620 if (useEssentialDynamics)
622 errorMessage += "Essential dynamics is not supported.\n";
624 if (inputrec.bPull && pull_have_constraint(inputrec.pull))
626 errorMessage += "Constraints pulling is not supported.\n";
628 if (doOrientationRestraints)
630 // The graph is needed, but not supported
631 errorMessage += "Orientation restraints are not supported.\n";
633 if (inputrec.efep != efepNO
634 && (haveFreeEnergyType(inputrec, efptBONDED) || haveFreeEnergyType(inputrec, efptMASS)))
636 errorMessage += "Free energy perturbation for mass and constraints are not supported.\n";
638 const auto particleTypes = gmx_mtop_particletype_count(mtop);
639 if (particleTypes[eptShell] > 0)
641 errorMessage += "Shells are not supported.\n";
643 if (useReplicaExchange)
645 errorMessage += "Replica exchange simulations are not supported.\n";
647 if (inputrec.eSwapCoords != eswapNO)
649 errorMessage += "Swapping the coordinates is not supported.\n";
653 errorMessage += "Re-run is not supported.\n";
656 // TODO: F_CONSTRNC is only unsupported, because isNumCoupledConstraintsSupported()
657 // does not support it, the actual CUDA LINCS code does support it
658 if (gmx_mtop_ftype_count(mtop, F_CONSTRNC) > 0)
660 errorMessage += "Non-connecting constraints are not supported";
662 if (!UpdateConstrainGpu::isNumCoupledConstraintsSupported(mtop))
665 "The number of coupled constraints is higher than supported in the GPU LINCS "
669 if (!errorMessage.empty())
671 if (updateTarget != TaskTarget::Gpu && forceGpuUpdateDefault)
673 GMX_LOG(mdlog.warning)
676 "Update task on the GPU was required, by the "
677 "GMX_FORCE_UPDATE_DEFAULT_GPU environment variable, but the following "
678 "condition(s) were not satisfied:");
679 GMX_LOG(mdlog.warning).asParagraph().appendText(errorMessage.c_str());
680 GMX_LOG(mdlog.warning).asParagraph().appendText("Will use CPU version of update.");
682 else if (updateTarget == TaskTarget::Gpu)
684 std::string prefix = gmx::formatString(
685 "Update task on the GPU was required,\n"
686 "but the following condition(s) were not satisfied:\n");
687 GMX_THROW(InconsistentInputError((prefix + errorMessage).c_str()));
692 if (isDomainDecomposition)
694 return forceGpuUpdateDefault;
698 return (updateTarget == TaskTarget::Gpu || forceGpuUpdateDefault);