/*
* This file is part of the GROMACS molecular simulation package.
*
- * Copyright (c) 2015,2016,2017,2018,2019, by the GROMACS development team, led by
+ * Copyright (c) 2015,2016,2017,2018,2019 by the GROMACS development team.
+ * Copyright (c) 2020,2021, by the GROMACS development team, led by
* Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
* and including many others, as listed in the AUTHORS file in the
* top-level source directory and at http://www.gromacs.org.
#include "gromacs/hardware/detecthardware.h"
#include "gromacs/hardware/hardwaretopology.h"
#include "gromacs/hardware/hw_info.h"
+#include "gromacs/listed_forces/gpubonded.h"
#include "gromacs/mdlib/gmx_omp_nthreads.h"
-#include "gromacs/mdlib/update_constrain_cuda.h"
+#include "gromacs/mdlib/update_constrain_gpu.h"
#include "gromacs/mdtypes/commrec.h"
#include "gromacs/mdtypes/inputrec.h"
#include "gromacs/mdtypes/md_enums.h"
#include "gromacs/mdtypes/mdrunoptions.h"
+#include "gromacs/pulling/pull.h"
#include "gromacs/taskassignment/taskassignment.h"
#include "gromacs/topology/mtop_util.h"
#include "gromacs/topology/topology.h"
const char* g_specifyEverythingFormatString =
"When you use mdrun -gputasks, %s must be set to non-default "
"values, so that the device IDs can be interpreted correctly."
-#if GMX_GPU != GMX_GPU_NONE
+#if GMX_GPU
" If you simply want to restrict which GPUs are used, then it is "
"better to use mdrun -gpu_id. Otherwise, setting the "
-# if GMX_GPU == GMX_GPU_CUDA
+# if GMX_GPU_CUDA
"CUDA_VISIBLE_DEVICES"
-# elif GMX_GPU == GMX_GPU_OPENCL
+# elif GMX_GPU_OPENCL
// Technically there is no portable way to do this offered by the
// OpenCL standard, but the only current relevant case for GROMACS
// is AMD OpenCL, which offers this variable.
"GPU_DEVICE_ORDINAL"
+# elif GMX_GPU_SYCL
+ // As with OpenCL, there are no portable way to do it.
+ // Intel reference: https://github.com/intel/llvm/blob/sycl/sycl/doc/EnvironmentVariables.md
+ // While SYCL_DEVICE_FILTER is a better option, as of 2021.1-beta10 it is not yet supported.
+ "SYCL_DEVICE_ALLOWLIST"
# else
# error "Unreachable branch"
# endif
} // namespace
bool decideWhetherToUseGpusForNonbondedWithThreadMpi(const TaskTarget nonbondedTarget,
- const std::vector<int>& gpuIdsToUse,
+ const int numDevicesToUse,
const std::vector<int>& userGpuTaskAssignment,
const EmulateGpuNonbonded emulateGpuNonbonded,
const bool buildSupportsNonbondedOnGpu,
// Because this is thread-MPI, we already know about the GPUs that
// all potential ranks can use, and can use that in a global
// decision that will later be consistent.
- auto haveGpus = !gpuIdsToUse.empty();
-
// If we get here, then the user permitted or required GPUs.
- return haveGpus;
+ return (numDevicesToUse > 0);
}
bool decideWhetherToUseGpusForPmeWithThreadMpi(const bool useGpuForNonbonded,
const TaskTarget pmeTarget,
- const std::vector<int>& gpuIdsToUse,
+ const TaskTarget pmeFftTarget,
+ const int numDevicesToUse,
const std::vector<int>& userGpuTaskAssignment,
const gmx_hw_info_t& hardwareInfo,
const t_inputrec& inputrec,
- const gmx_mtop_t& mtop,
const int numRanksPerSimulation,
const int numPmeRanksPerSimulation)
{
// First, exclude all cases where we can't run PME on GPUs.
if ((pmeTarget == TaskTarget::Cpu) || !useGpuForNonbonded || !pme_gpu_supports_build(nullptr)
- || !pme_gpu_supports_hardware(hardwareInfo, nullptr)
- || !pme_gpu_supports_input(inputrec, mtop, nullptr))
+ || !pme_gpu_supports_hardware(hardwareInfo, nullptr) || !pme_gpu_supports_input(inputrec, nullptr))
+ {
+ // PME can't run on a GPU. If the user required that, we issue an error later.
+ return false;
+ }
+
+ if (pmeFftTarget == TaskTarget::Cpu && !pme_gpu_mixed_mode_supports_input(inputrec, nullptr))
{
- // PME can't run on a GPU. If the user required that, we issue
- // an error later.
+ /* User requested PME FFT on CPU, but the current system is not compatible with Mixed mode,
+ * so we don't use GPUs at all.
+ * If the user had -pme gpu, we issue an error later. */
return false;
}
{
// PME can run well on a GPU shared with NB, and we permit
// mdrun to default to try that.
- return !gpuIdsToUse.empty();
+ return numDevicesToUse > 0;
}
if (numRanksPerSimulation < 1)
// Full automated mode for thread-MPI (the default). PME can
// run well on a GPU shared with NB, and we permit mdrun to
// default to it if there is only one GPU available.
- return (gpuIdsToUse.size() == 1);
+ return (numDevicesToUse == 1);
}
// Not enough support for PME on GPUs for anything else
// If we get here, then the user permitted GPUs, which we should
// use for nonbonded interactions.
- return gpusWereDetected;
+ return buildSupportsNonbondedOnGpu && gpusWereDetected;
}
bool decideWhetherToUseGpusForPme(const bool useGpuForNonbonded,
const TaskTarget pmeTarget,
+ const TaskTarget pmeFftTarget,
const std::vector<int>& userGpuTaskAssignment,
const gmx_hw_info_t& hardwareInfo,
const t_inputrec& inputrec,
- const gmx_mtop_t& mtop,
const int numRanksPerSimulation,
const int numPmeRanksPerSimulation,
const bool gpusWereDetected)
}
return false;
}
- if (!pme_gpu_supports_input(inputrec, mtop, &message))
+ if (!pme_gpu_supports_input(inputrec, &message))
{
if (pmeTarget == TaskTarget::Gpu)
{
}
return false;
}
+ if (pmeFftTarget == TaskTarget::Cpu && !pme_gpu_mixed_mode_supports_input(inputrec, &message))
+ {
+ /* User requested PME FFT on CPU, but the current system is not compatible with Mixed mode,
+ * so we don't use GPUs at all. */
+ if (pmeTarget == TaskTarget::Gpu)
+ {
+ GMX_THROW(NotImplementedError("Cannot compute PME interactions in Mixed mode, because " + message));
+ }
+ return false;
+ }
if (pmeTarget == TaskTarget::Cpu)
{
return false;
}
-bool decideWhetherToUseGpusForBonded(const bool useGpuForNonbonded,
- const bool useGpuForPme,
- const TaskTarget bondedTarget,
- const bool canUseGpuForBonded,
- const bool usingLJPme,
- const bool usingElecPmeOrEwald,
- const int numPmeRanksPerSimulation,
- const bool gpusWereDetected)
+
+PmeRunMode determinePmeRunMode(const bool useGpuForPme, const TaskTarget& pmeFftTarget, const t_inputrec& inputrec)
+{
+ if (!EEL_PME(inputrec.coulombtype))
+ {
+ return PmeRunMode::None;
+ }
+
+ if (useGpuForPme)
+ {
+ if (pmeFftTarget == TaskTarget::Cpu)
+ {
+ return PmeRunMode::Mixed;
+ }
+ else
+ {
+ return PmeRunMode::GPU;
+ }
+ }
+ else
+ {
+ if (pmeFftTarget == TaskTarget::Gpu)
+ {
+ gmx_fatal(FARGS,
+ "Assigning FFTs to GPU requires PME to be assigned to GPU as well. With PME "
+ "on CPU you should not be using -pmefft.");
+ }
+ return PmeRunMode::CPU;
+ }
+}
+
+bool decideWhetherToUseGpusForBonded(bool useGpuForNonbonded,
+ bool useGpuForPme,
+ TaskTarget bondedTarget,
+ const t_inputrec& inputrec,
+ const gmx_mtop_t& mtop,
+ int numPmeRanksPerSimulation,
+ bool gpusWereDetected)
{
if (bondedTarget == TaskTarget::Cpu)
{
return false;
}
- if (!canUseGpuForBonded)
+ std::string errorMessage;
+
+ if (!buildSupportsGpuBondeds(&errorMessage))
{
if (bondedTarget == TaskTarget::Gpu)
{
- GMX_THROW(InconsistentInputError(
- "Bonded interactions on the GPU were required, but not supported for these "
- "simulation settings. Change your settings, or do not require using GPUs."));
+ GMX_THROW(InconsistentInputError(errorMessage.c_str()));
+ }
+
+ return false;
+ }
+
+ if (!inputSupportsGpuBondeds(inputrec, mtop, &errorMessage))
+ {
+ if (bondedTarget == TaskTarget::Gpu)
+ {
+ GMX_THROW(InconsistentInputError(errorMessage.c_str()));
}
return false;
// Note that here we assume that the auto setting of PME ranks will not
// choose seperate PME ranks when nonBonded are assigned to the GPU.
bool usingOurCpuForPmeOrEwald =
- (usingLJPme || (usingElecPmeOrEwald && !useGpuForPme && numPmeRanksPerSimulation <= 0));
+ (EVDW_PME(inputrec.vdwtype)
+ || (EEL_PME_EWALD(inputrec.coulombtype) && !useGpuForPme && numPmeRanksPerSimulation <= 0));
return gpusWereDetected && usingOurCpuForPmeOrEwald;
}
-bool decideWhetherToUseGpuForUpdate(const bool isDomainDecomposition,
- const bool useGpuForPme,
- const bool useGpuForNonbonded,
- const TaskTarget updateTarget,
- const bool gpusWereDetected,
- const t_inputrec& inputrec,
- const gmx_mtop_t& mtop,
- const bool useEssentialDynamics,
- const bool doOrientationRestraints,
- const bool useReplicaExchange)
+bool decideWhetherToUseGpuForUpdate(const bool isDomainDecomposition,
+ const bool useUpdateGroups,
+ const PmeRunMode pmeRunMode,
+ const bool havePmeOnlyRank,
+ const bool useGpuForNonbonded,
+ const TaskTarget updateTarget,
+ const bool gpusWereDetected,
+ const t_inputrec& inputrec,
+ const gmx_mtop_t& mtop,
+ const bool useEssentialDynamics,
+ const bool doOrientationRestraints,
+ const bool useReplicaExchange,
+ const bool haveFrozenAtoms,
+ const bool doRerun,
+ const DevelopmentFeatureFlags& devFlags,
+ const gmx::MDLogger& mdlog)
{
- if (updateTarget == TaskTarget::Cpu)
+ // '-update cpu' overrides the environment variable, '-update auto' does not
+ if (updateTarget == TaskTarget::Cpu
+ || (updateTarget == TaskTarget::Auto && !devFlags.forceGpuUpdateDefault))
{
return false;
}
+ const bool hasAnyConstraints = gmx_mtop_interaction_count(mtop, IF_CONSTRAINT) > 0;
+ const bool pmeUsesCpu = (pmeRunMode == PmeRunMode::CPU || pmeRunMode == PmeRunMode::Mixed);
+
std::string errorMessage;
if (isDomainDecomposition)
{
- errorMessage += "Domain decomposition is not supported.\n";
+ if (hasAnyConstraints && !useUpdateGroups)
+ {
+ errorMessage +=
+ "Domain decomposition is only supported with constraints when update "
+ "groups "
+ "are used. This means constraining all bonds is not supported, except for "
+ "small molecules, and box sizes close to half the pair-list cutoff are not "
+ "supported.\n ";
+ }
+ }
+
+ if (havePmeOnlyRank)
+ {
+ if (pmeUsesCpu)
+ {
+ errorMessage += "With separate PME rank(s), PME must run fully on the GPU.\n";
+ }
+ }
+
+ if (inputrec.useMts)
+ {
+ errorMessage += "Multiple time stepping is not supported.\n";
+ }
+
+ if (inputrec.eConstrAlg == econtSHAKE && hasAnyConstraints && gmx_mtop_ftype_count(mtop, F_CONSTR) > 0)
+ {
+ errorMessage += "SHAKE constraints are not supported.\n";
}
// Using the GPU-version of update if:
- // 1. PME is on the GPU (there should be a copy of coordinates on GPU for PME spread), or
+ // 1. PME is on the GPU (there should be a copy of coordinates on GPU for PME spread) or inactive, or
// 2. Non-bonded interactions are on the GPU.
- if (!(useGpuForPme || useGpuForNonbonded))
+ if ((pmeRunMode == PmeRunMode::CPU || pmeRunMode == PmeRunMode::None) && !useGpuForNonbonded)
{
errorMessage +=
"Either PME or short-ranged non-bonded interaction tasks must run on the GPU.\n";
}
+
if (!gpusWereDetected)
{
errorMessage += "Compatible GPUs must have been found.\n";
}
- if (GMX_GPU != GMX_GPU_CUDA)
+ if (!GMX_GPU_CUDA)
{
errorMessage += "Only a CUDA build is supported.\n";
}
{
errorMessage += "Nose-Hoover temperature coupling is not supported.\n";
}
- if (!(inputrec.epc == epcNO || inputrec.epc == epcPARRINELLORAHMAN || inputrec.epc == epcBERENDSEN))
+ if (!(inputrec.epc == epcNO || inputrec.epc == epcPARRINELLORAHMAN
+ || inputrec.epc == epcBERENDSEN || inputrec.epc == epcCRESCALE))
{
- errorMessage += "Only Parrinello-Rahman and Berendsen pressure coupling are supported.\n";
+ errorMessage +=
+ "Only Parrinello-Rahman, Berendsen, and C-rescale pressure coupling are "
+ "supported.\n";
}
if (EEL_PME_EWALD(inputrec.coulombtype) && inputrec.epsilon_surface != 0)
{
{
errorMessage += "Essential dynamics is not supported.\n";
}
- if (inputrec.bPull || inputrec.pull)
+ if (inputrec.bPull && pull_have_constraint(*inputrec.pull))
{
- // Pull potentials are actually supported, but constraint pulling is not
- errorMessage += "Pulling is not supported.\n";
+ errorMessage += "Constraints pulling is not supported.\n";
}
if (doOrientationRestraints)
{
// The graph is needed, but not supported
errorMessage += "Orientation restraints are not supported.\n";
}
- if (inputrec.efep != efepNO)
+ if (inputrec.efep != efepNO && (haveFepPerturbedMasses(mtop) || havePerturbedConstraints(mtop)))
+ {
+ errorMessage += "Free energy perturbation for mass and constraints are not supported.\n";
+ }
+ const auto particleTypes = gmx_mtop_particletype_count(mtop);
+ if (particleTypes[eptShell] > 0)
{
- // Actually all free-energy options except for mass and constraint perturbation are supported
- errorMessage += "Free energy perturbations are not supported.\n";
+ errorMessage += "Shells are not supported.\n";
}
if (useReplicaExchange)
{
{
errorMessage += "Swapping the coordinates is not supported.\n";
}
+ if (doRerun)
+ {
+ errorMessage += "Re-run is not supported.\n";
+ }
// TODO: F_CONSTRNC is only unsupported, because isNumCoupledConstraintsSupported()
// does not support it, the actual CUDA LINCS code does support it
if (gmx_mtop_ftype_count(mtop, F_CONSTRNC) > 0)
{
- errorMessage += "Non-connecting constraints are not supported";
+ errorMessage += "Non-connecting constraints are not supported\n";
}
- if (!UpdateConstrainCuda::isNumCoupledConstraintsSupported(mtop))
+ if (!UpdateConstrainGpu::isNumCoupledConstraintsSupported(mtop))
{
errorMessage +=
- "The number of coupled constraints is higher than supported in the CUDA LINCS "
+ "The number of coupled constraints is higher than supported in the GPU LINCS "
"code.\n";
}
+ if (haveFrozenAtoms)
+ {
+ // There is a known bug with frozen atoms and GPU update, see Issue #3920.
+ errorMessage += "Frozen atoms not supported.\n";
+ }
if (!errorMessage.empty())
{
- if (updateTarget == TaskTarget::Gpu)
+ if (updateTarget == TaskTarget::Auto && devFlags.forceGpuUpdateDefault)
+ {
+ GMX_LOG(mdlog.warning)
+ .asParagraph()
+ .appendText(
+ "Update task on the GPU was required, by the "
+ "GMX_FORCE_UPDATE_DEFAULT_GPU environment variable, but the following "
+ "condition(s) were not satisfied:");
+ GMX_LOG(mdlog.warning).asParagraph().appendText(errorMessage.c_str());
+ GMX_LOG(mdlog.warning).asParagraph().appendText("Will use CPU version of update.");
+ }
+ else if (updateTarget == TaskTarget::Gpu)
{
std::string prefix = gmx::formatString(
"Update task on the GPU was required,\n"
return false;
}
- return true;
+ return (updateTarget == TaskTarget::Gpu
+ || (updateTarget == TaskTarget::Auto && devFlags.forceGpuUpdateDefault));
+}
+
+bool decideWhetherToUseGpuForHalo(const DevelopmentFeatureFlags& devFlags,
+ bool havePPDomainDecomposition,
+ bool useGpuForNonbonded,
+ bool useModularSimulator,
+ bool doRerun,
+ bool haveEnergyMinimization)
+{
+ return havePPDomainDecomposition && devFlags.enableGpuHaloExchange && useGpuForNonbonded
+ && !useModularSimulator && !doRerun && !haveEnergyMinimization;
}
} // namespace gmx