{
DevelopmentFeatureFlags devFlags;
- devFlags.enableGpuBufferOps =
- GMX_GPU_CUDA && useGpuForNonbonded && (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
+ devFlags.enableGpuBufferOps = (GMX_GPU_CUDA || GMX_GPU_SYCL) && useGpuForNonbonded
+ && (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
devFlags.enableGpuHaloExchange = GMX_MPI && GMX_GPU_CUDA && getenv("GMX_GPU_DD_COMMS") != nullptr;
devFlags.forceGpuUpdateDefault = (getenv("GMX_FORCE_UPDATE_DEFAULT_GPU") != nullptr) || GMX_FAHCORE;
devFlags.enableGpuPmePPComm = GMX_MPI && GMX_GPU_CUDA && getenv("GMX_GPU_PME_PP_COMMS") != nullptr;
print_flop(fplog, nrnb_tot, &nbfs, &mflop);
}
- if (thisRankHasDuty(cr, DUTY_PP) && DOMAINDECOMP(cr))
+ if (thisRankHasDuty(cr, DUTY_PP) && haveDDAtomOrdering(*cr))
{
print_dd_statistics(cr, inputrec, fplog);
}
hw_opt.nthreads_tmpi);
useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi(useGpuForNonbonded,
pmeTarget,
+ pmeFftTarget,
numAvailableDevices,
userGpuTaskAssignment,
*hwinfo_,
// master and spawned threads joins at the end of this block.
}
- GMX_RELEASE_ASSERT(ms || simulationCommunicator != MPI_COMM_NULL,
+ GMX_RELEASE_ASSERT(!GMX_MPI || ms || simulationCommunicator != MPI_COMM_NULL,
"Must have valid communicator unless running a multi-simulation");
CommrecHandle crHandle = init_commrec(simulationCommunicator);
t_commrec* cr = crHandle.get();
GMX_RELEASE_ASSERT(inputrec != nullptr, "All ranks should have a valid inputrec now");
partialDeserializedTpr.reset(nullptr);
- GMX_RELEASE_ASSERT(
- !inputrec->useConstantAcceleration,
- "Linear acceleration has been removed in GROMACS 2022, and was broken for many years "
- "before that. Use GROMACS 4.5 or earlier if you need this feature.");
-
// Now the number of ranks is known to all ranks, and each knows
// the inputrec read by the master rank. The ranks can now all run
// the task-deciding functions and will agree on the result
// without needing to communicate.
+ // The LBFGS minimizer, test-particle insertion, normal modes and shell dynamics don't support DD
const bool useDomainDecomposition =
- (PAR(cr) && !(EI_TPI(inputrec->eI) || inputrec->eI == IntegrationAlgorithm::NM));
+ !(inputrec->eI == IntegrationAlgorithm::LBFGS || EI_TPI(inputrec->eI)
+ || inputrec->eI == IntegrationAlgorithm::NM
+ || gmx_mtop_particletype_count(mtop)[ParticleType::Shell] > 0);
// Note that these variables describe only their own node.
//
gpusWereDetected);
useGpuForPme = decideWhetherToUseGpusForPme(useGpuForNonbonded,
pmeTarget,
+ pmeFftTarget,
userGpuTaskAssignment,
*hwinfo_,
*inputrec,
globalState = std::make_unique<t_state>();
}
broadcastStateWithoutDynamics(
- cr->mpiDefaultCommunicator, DOMAINDECOMP(cr), PAR(cr), globalState.get());
+ cr->mpiDefaultCommunicator, haveDDAtomOrdering(*cr), PAR(cr), globalState.get());
}
/* A parallel command line option consistency check that we can
int deviceId = -1;
DeviceInformation* deviceInfo = gpuTaskAssignments.initDevice(&deviceId);
- // timing enabling - TODO put this in gpu_utils (even though generally this is just option handling?)
- bool useTiming = true;
-
- if (GMX_GPU_CUDA)
- {
- /* WARNING: CUDA timings are incorrect with multiple streams.
- * This is the main reason why they are disabled by default.
- */
- // TODO: Consider turning on by default when we can detect nr of streams.
- useTiming = (getenv("GMX_ENABLE_GPU_TIMING") != nullptr);
- }
- else if (GMX_GPU_OPENCL)
- {
- useTiming = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
- }
-
// TODO Currently this is always built, yet DD partition code
// checks if it is built before using it. Probably it should
// become an MDModule that is made only when another module
if (deviceInfo != nullptr)
{
- if (DOMAINDECOMP(cr) && thisRankHasDuty(cr, DUTY_PP))
+ if (runScheduleWork.simulationWork.havePpDomainDecomposition && thisRankHasDuty(cr, DUTY_PP))
{
dd_setup_dlb_resource_sharing(cr, deviceId);
}
- deviceStreamManager = std::make_unique<DeviceStreamManager>(
- *deviceInfo, havePPDomainDecomposition(cr), runScheduleWork.simulationWork, useTiming);
+ const bool useGpuTiming = decideGpuTimingsUsage();
+ deviceStreamManager = std::make_unique<DeviceStreamManager>(
+ *deviceInfo, havePPDomainDecomposition(cr), runScheduleWork.simulationWork, useGpuTiming);
}
// If the user chose a task assignment, give them some hints
// Enable Peer access between GPUs where available
// Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
// any of the GPU communication features are active.
- if (DOMAINDECOMP(cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
+ if (haveDDAtomOrdering(*cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
&& (runScheduleWork.simulationWork.useGpuHaloExchange
|| runScheduleWork.simulationWork.useGpuPmePpCommunication))
{
}
}
// Make the DD reverse topology, now that any vsites that are present are available
- if (DOMAINDECOMP(cr))
+ if (haveDDAtomOrdering(*cr))
{
dd_make_reverse_top(fplog, cr->dd, mtop, vsite.get(), *inputrec, domdecOptions.ddBondedChecking);
}
ms,
&nrnb,
wcycle.get(),
- fr->bMolPBC);
+ fr->bMolPBC,
+ &observablesReducerBuilder);
/* Energy terms and groups */
gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(),
mdrunOptions.imdOptions,
startingBehavior);
- if (DOMAINDECOMP(cr))
+ if (haveDDAtomOrdering(*cr))
{
GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
/* This call is not included in init_domain_decomposition
makeBondedLinks(cr->dd, mtop, fr->atomInfoForEachMoleculeBlock);
}
- if (runScheduleWork.simulationWork.useGpuBufferOps)
+ if (runScheduleWork.simulationWork.useGpuFBufferOps)
{
fr->gpuForceReduction[gmx::AtomLocality::Local] = std::make_unique<gmx::GpuForceReduction>(
deviceStreamManager->context(),
std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
if (gpusWereDetected
&& ((runScheduleWork.simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME))
- || runScheduleWork.simulationWork.useGpuBufferOps))
+ || runScheduleWork.simulationWork.useGpuXBufferOps))
{
GpuApiCallBehavior transferKind =
(inputrec->eI == IntegrationAlgorithm::MD && !doRerun && !useModularSimulator)
biod.pnpi.spb.ru / alexxy / gromacs.git / blobdiff