*
* Copyright (c) 1991-2000, University of Groningen, The Netherlands.
* Copyright (c) 2001-2004, The GROMACS development team.
- * Copyright (c) 2011-2019,2020, by the GROMACS development team, led by
+ * Copyright (c) 2011-2019,2020,2021, 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/cpuinfo.h"
#include "gromacs/hardware/detecthardware.h"
#include "gromacs/hardware/device_management.h"
+#include "gromacs/hardware/hardwaretopology.h"
#include "gromacs/hardware/printhardware.h"
#include "gromacs/imd/imd.h"
#include "gromacs/listed_forces/disre.h"
newRunner.hw_opt = hw_opt;
newRunner.filenames = filenames;
+ newRunner.hwinfo_ = hwinfo_;
newRunner.oenv = oenv;
newRunner.mdrunOptions = mdrunOptions;
newRunner.domdecOptions = domdecOptions;
gmx_wallcycle_t wcycle;
gmx_walltime_accounting_t walltime_accounting = nullptr;
MembedHolder membedHolder(filenames.size(), filenames.data());
- gmx_hw_info_t* hwinfo = nullptr;
/* CAUTION: threads may be started later on in this function, so
cr doesn't reflect the final parallel state right now */
gmx::LoggerOwner logOwner(buildLogger(fplog, isSimulationMasterRank));
gmx::MDLogger mdlog(logOwner.logger());
- // TODO The thread-MPI master rank makes a working
- // PhysicalNodeCommunicator here, but it gets rebuilt by all ranks
- // after the threads have been launched. This works because no use
- // is made of that communicator until after the execution paths
- // have rejoined. But it is likely that we can improve the way
- // this is expressed, e.g. by expressly running detection only the
- // master rank for thread-MPI, rather than relying on the mutex
- // and reference count.
- PhysicalNodeCommunicator physicalNodeComm(libraryWorldCommunicator, gmx_physicalnode_id_hash());
- hwinfo = gmx_detect_hardware(mdlog, physicalNodeComm);
-
- gmx_print_detected_hardware(fplog, isSimulationMasterRank && isMasterSim(ms), mdlog, hwinfo);
+ gmx_print_detected_hardware(fplog, isSimulationMasterRank && isMasterSim(ms), mdlog, hwinfo_);
- std::vector<int> gpuIdsToUse = makeGpuIdsToUse(hwinfo->deviceInfoList, hw_opt.gpuIdsAvailable);
+ std::vector<int> gpuIdsToUse = makeGpuIdsToUse(hwinfo_->deviceInfoList, hw_opt.gpuIdsAvailable);
const int numDevicesToUse = gmx::ssize(gpuIdsToUse);
// Print citation requests after all software/hardware printing
gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, GMX_THREAD_MPI),
hw_opt.nthreads_tmpi);
useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi(
- useGpuForNonbonded, pmeTarget, numDevicesToUse, userGpuTaskAssignment, *hwinfo,
- *inputrec, hw_opt.nthreads_tmpi, domdecOptions.numPmeRanks);
+ useGpuForNonbonded, pmeTarget, pmeFftTarget, numDevicesToUse, userGpuTaskAssignment,
+ *hwinfo_, *inputrec, hw_opt.nthreads_tmpi, domdecOptions.numPmeRanks);
}
GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
* prevent any possible subsequent checks from working
* correctly. */
hw_opt.nthreads_tmpi =
- get_nthreads_mpi(hwinfo, &hw_opt, numDevicesToUse, useGpuForNonbonded, useGpuForPme,
+ get_nthreads_mpi(hwinfo_, &hw_opt, numDevicesToUse, useGpuForNonbonded, useGpuForPme,
inputrec.get(), &mtop, mdlog, membedHolder.doMembed());
// Now start the threads for thread MPI.
spawnThreads(hw_opt.nthreads_tmpi);
// The spawned threads enter mdrunner() and execution of
// master and spawned threads joins at the end of this block.
- physicalNodeComm =
- PhysicalNodeCommunicator(libraryWorldCommunicator, gmx_physicalnode_id_hash());
}
GMX_RELEASE_ASSERT(ms || simulationCommunicator != MPI_COMM_NULL,
t_commrec* cr = crHandle.get();
GMX_RELEASE_ASSERT(cr != nullptr, "Must have valid commrec");
+ PhysicalNodeCommunicator physicalNodeComm(libraryWorldCommunicator, gmx_physicalnode_id_hash());
+
+ // If we detected the topology on this system, double-check that it makes sense
+ if (hwinfo_->hardwareTopology->isThisSystem())
+ {
+ hardwareTopologyDoubleCheckDetection(mdlog, *hwinfo_->hardwareTopology);
+ }
+
if (PAR(cr))
{
/* now broadcast everything to the non-master nodes/threads: */
bool useGpuForPme = false;
bool useGpuForBonded = false;
bool useGpuForUpdate = false;
- bool gpusWereDetected = hwinfo->ngpu_compatible_tot > 0;
+ bool gpusWereDetected = hwinfo_->ngpu_compatible_tot > 0;
try
{
// It's possible that there are different numbers of GPUs on
nonbondedTarget, userGpuTaskAssignment, emulateGpuNonbonded, canUseGpuForNonbonded,
gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, !GMX_THREAD_MPI), gpusWereDetected);
useGpuForPme = decideWhetherToUseGpusForPme(
- useGpuForNonbonded, pmeTarget, userGpuTaskAssignment, *hwinfo, *inputrec,
- cr->sizeOfDefaultCommunicator, domdecOptions.numPmeRanks, gpusWereDetected);
- auto canUseGpuForBonded = buildSupportsGpuBondeds(nullptr)
- && inputSupportsGpuBondeds(*inputrec, mtop, nullptr);
- useGpuForBonded = decideWhetherToUseGpusForBonded(
- useGpuForNonbonded, useGpuForPme, bondedTarget, canUseGpuForBonded,
- EVDW_PME(inputrec->vdwtype), EEL_PME_EWALD(inputrec->coulombtype),
- domdecOptions.numPmeRanks, gpusWereDetected);
+ useGpuForNonbonded, pmeTarget, pmeFftTarget, userGpuTaskAssignment, *hwinfo_,
+ *inputrec, cr->sizeOfDefaultCommunicator, domdecOptions.numPmeRanks, gpusWereDetected);
+ useGpuForBonded = decideWhetherToUseGpusForBonded(useGpuForNonbonded, useGpuForPme,
+ bondedTarget, *inputrec, mtop,
+ domdecOptions.numPmeRanks, gpusWereDetected);
}
GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
*/
prepare_verlet_scheme(fplog, cr, inputrec.get(), nstlist_cmdline, &mtop, box,
useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes),
- *hwinfo->cpuInfo);
+ *hwinfo_->cpuInfo);
// This builder is necessary while we have multi-part construction
// of DD. Before DD is constructed, we use the existence of
// Produce the task assignment for this rank - done after DD is constructed
GpuTaskAssignments gpuTaskAssignments = GpuTaskAssignmentsBuilder::build(
- gpuIdsToUse, userGpuTaskAssignment, *hwinfo, simulationCommunicator, physicalNodeComm,
+ gpuIdsToUse, userGpuTaskAssignment, *hwinfo_, simulationCommunicator, physicalNodeComm,
nonbondedTarget, pmeTarget, bondedTarget, updateTarget, useGpuForNonbonded,
useGpuForPme, thisRankHasDuty(cr, DUTY_PP),
// TODO cr->duty & DUTY_PME should imply that a PME
ddBuilder.reset(nullptr);
// Note that local state still does not exist yet.
}
+ // Ensure that all atoms within the same update group are in the
+ // same periodic image. Otherwise, a simulation that did not use
+ // update groups (e.g. a single-rank simulation) cannot always be
+ // correctly restarted in a way that does use update groups
+ // (e.g. a multi-rank simulation).
+ if (isSimulationMasterRank)
+ {
+ const bool useUpdateGroups = cr->dd ? ddUsesUpdateGroups(*cr->dd) : false;
+ if (useUpdateGroups)
+ {
+ putUpdateGroupAtomsInSamePeriodicImage(*cr->dd, mtop, globalState->box, globalState->x);
+ }
+ }
// The GPU update is decided here because we need to know whether the constraints or
// SETTLEs can span accross the domain borders (i.e. whether or not update groups are
try
{
const bool useUpdateGroups = cr->dd ? ddUsesUpdateGroups(*cr->dd) : false;
+ const bool haveFrozenAtoms = inputrecFrozenAtoms(inputrec.get());
useGpuForUpdate = decideWhetherToUseGpuForUpdate(
useDomainDecomposition, useUpdateGroups, pmeRunMode, domdecOptions.numPmeRanks > 0,
useGpuForNonbonded, updateTarget, gpusWereDetected, *inputrec, mtop,
doEssentialDynamics, gmx_mtop_ftype_count(mtop, F_ORIRES) > 0,
- replExParams.exchangeInterval > 0, doRerun, devFlags, mdlog);
+ replExParams.exchangeInterval > 0, haveFrozenAtoms, doRerun, devFlags, mdlog);
}
GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
// that existing affinity setting was from OpenMP or something
// else, so we run this code both before and after we initialize
// the OpenMP support.
- gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo->nthreads_hw_avail, FALSE);
+ gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, FALSE);
/* Check and update the number of OpenMP threads requested */
- checkAndUpdateRequestedNumOpenmpThreads(&hw_opt, *hwinfo, cr, ms, physicalNodeComm.size_,
+ checkAndUpdateRequestedNumOpenmpThreads(&hw_opt, *hwinfo_, cr, ms, physicalNodeComm.size_,
pmeRunMode, mtop, *inputrec);
- gmx_omp_nthreads_init(mdlog, cr, hwinfo->nthreads_hw_avail, physicalNodeComm.size_,
+ gmx_omp_nthreads_init(mdlog, cr, hwinfo_->nthreads_hw_avail, physicalNodeComm.size_,
hw_opt.nthreads_omp, hw_opt.nthreads_omp_pme, !thisRankHasDuty(cr, DUTY_PP));
// Enable FP exception detection, but not in
}
/* Now that we know the setup is consistent, check for efficiency */
- check_resource_division_efficiency(hwinfo, gpuTaskAssignments.thisRankHasAnyGpuTask(),
+ check_resource_division_efficiency(hwinfo_, gpuTaskAssignments.thisRankHasAnyGpuTask(),
mdrunOptions.ntompOptionIsSet, cr, mdlog);
/* getting number of PP/PME threads on this MPI / tMPI rank.
*/
const int numThreadsOnThisRank = thisRankHasDuty(cr, DUTY_PP) ? gmx_omp_nthreads_get(emntNonbonded)
: gmx_omp_nthreads_get(emntPME);
- checkHardwareOversubscription(numThreadsOnThisRank, cr->nodeid, *hwinfo->hardwareTopology,
+ checkHardwareOversubscription(numThreadsOnThisRank, cr->nodeid, *hwinfo_->hardwareTopology,
physicalNodeComm, mdlog);
// Enable Peer access between GPUs where available
* - which indicates that probably the OpenMP library has changed it
* since we first checked).
*/
- gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo->nthreads_hw_avail, TRUE);
+ gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, TRUE);
int numThreadsOnThisNode, intraNodeThreadOffset;
analyzeThreadsOnThisNode(physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode,
&intraNodeThreadOffset);
/* Set the CPU affinity */
- gmx_set_thread_affinity(mdlog, cr, &hw_opt, *hwinfo->hardwareTopology, numThreadsOnThisRank,
+ gmx_set_thread_affinity(mdlog, cr, &hw_opt, *hwinfo_->hardwareTopology, numThreadsOnThisRank,
numThreadsOnThisNode, intraNodeThreadOffset, nullptr);
}
deviceStreamManager->stream(DeviceStreamType::PmePpTransfer));
}
- fr->nbv = Nbnxm::init_nb_verlet(mdlog, inputrec.get(), fr, cr, *hwinfo,
+ fr->nbv = Nbnxm::init_nb_verlet(mdlog, inputrec.get(), fr, cr, *hwinfo_,
runScheduleWork.simulationWork.useGpuNonbonded,
deviceStreamManager.get(), &mtop, box, wcycle);
// TODO: Move the logic below to a GPU bonded builder
gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(),
inputrec->fepvals->n_lambda);
+ // cos acceleration is only supported by md, but older tpr
+ // files might still combine it with other integrators
+ GMX_RELEASE_ASSERT(inputrec->cos_accel == 0.0 || inputrec->eI == eiMD,
+ "cos_acceleration is only supported by integrator=md");
+
/* Kinetic energy data */
gmx_ekindata_t ekind;
init_ekindata(fplog, &mtop, &(inputrec->opts), &ekind, inputrec->cos_accel);
{
fr->gpuForceReduction[gmx::AtomLocality::Local] = std::make_unique<gmx::GpuForceReduction>(
deviceStreamManager->context(),
- deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedLocal));
+ deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedLocal), wcycle);
fr->gpuForceReduction[gmx::AtomLocality::NonLocal] = std::make_unique<gmx::GpuForceReduction>(
deviceStreamManager->context(),
- deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedNonLocal));
+ deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedNonLocal), wcycle);
}
std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
sfree(disresdata);
sfree(oriresdata);
- if (!hwinfo->deviceInfoList.empty())
+ if (!hwinfo_->deviceInfoList.empty())
{
/* stop the GPU profiler (only CUDA) */
stopGpuProfiler();
Mdrunner::Mdrunner(Mdrunner&&) noexcept = default;
-Mdrunner& Mdrunner::operator=(Mdrunner&& /*handle*/) noexcept = default;
+//NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265 in CentOS 7
+Mdrunner& Mdrunner::operator=(Mdrunner&& /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
class Mdrunner::BuilderImplementation
{
real forceWarningThreshold,
StartingBehavior startingBehavior);
+ void addHardwareDetectionResult(const gmx_hw_info_t* hwinfo);
+
void addDomdec(const DomdecOptions& options);
void addInput(SimulationInputHandle inputHolder);
//! The modules that comprise the functionality of mdrun.
std::unique_ptr<MDModules> mdModules_;
+ //! Detected hardware.
+ const gmx_hw_info_t* hwinfo_ = nullptr;
+
//! \brief Parallelism information.
gmx_hw_opt_t hardwareOptions_;
// nullptr is a valid value for the multisim handle
newRunner.ms = multiSimulation_;
+ if (hwinfo_)
+ {
+ newRunner.hwinfo_ = hwinfo_;
+ }
+ else
+ {
+ GMX_THROW(gmx::APIError(
+ "MdrunnerBuilder::addHardwareDetectionResult() is required before build()"));
+ }
+
if (inputHolder_)
{
newRunner.inputHolder_ = std::move(inputHolder_);
return newRunner;
}
+void Mdrunner::BuilderImplementation::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
+{
+ hwinfo_ = hwinfo;
+}
+
void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
{
nbpu_opt_ = nbpu_opt;
MdrunnerBuilder::~MdrunnerBuilder() = default;
+MdrunnerBuilder& MdrunnerBuilder::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
+{
+ impl_->addHardwareDetectionResult(hwinfo);
+ return *this;
+}
+
MdrunnerBuilder& MdrunnerBuilder::addSimulationMethod(const MdrunOptions& options,
real forceWarningThreshold,
const StartingBehavior startingBehavior)
biod.pnpi.spb.ru / alexxy / gromacs.git / blobdiff