Algorithm to reconstruct CaloCluster objects with GPU acceleration, providing interoperability for calling standard CPU algorithms before and after the GPU processing part. More...

#include <CaloGPUHybridClusterProcessor.h>

Inheritance diagram for CaloGPUHybridClusterProcessor:

Collaboration diagram for CaloGPUHybridClusterProcessor:

Classes
class	simple_GPU_pointer_holder
	A simple RAII wrapper to ensure proper allocation and deallocation of GPU memory in a `void` * for the temporaries. More...

Public Member Functions
	CaloGPUHybridClusterProcessor (const std::string &name, ISvcLocator *pSvcLocator)

virtual	~CaloGPUHybridClusterProcessor () override=default

virtual StatusCode	initialize () override

virtual StatusCode	initialize_non_CUDA () override
	Initialization that does not invoke CUDA functions. More...

virtual StatusCode	initialize_CUDA () override
	Initialization that invokes CUDA functions. More...

virtual StatusCode	execute (const EventContext &ctx) const override

virtual StatusCode	finalize () override

virtual StatusCode	sysInitialize () override
	Override sysInitialize. More...

virtual bool	isClonable () const override
	Specify if the algorithm is clonable. More...

virtual unsigned int	cardinality () const override
	Cardinality (Maximum number of clones that can exist) special value 0 means that algorithm is reentrant. More...

virtual StatusCode	sysExecute (const EventContext &ctx) override
	Execute an algorithm. More...

virtual const DataObjIDColl &	extraOutputDeps () const override
	Return the list of extra output dependencies. More...

virtual bool	filterPassed (const EventContext &ctx) const

virtual void	setFilterPassed (bool state, const EventContext &ctx) const

ServiceHandle< StoreGateSvc > &	evtStore ()
	The standard `StoreGateSvc` (event store) Returns (kind of) a pointer to the `StoreGateSvc`. More...

const ServiceHandle< StoreGateSvc > &	evtStore () const
	The standard `StoreGateSvc` (event store) Returns (kind of) a pointer to the `StoreGateSvc`. More...

const ServiceHandle< StoreGateSvc > &	detStore () const
	The standard `StoreGateSvc/DetectorStore` Returns (kind of) a pointer to the `StoreGateSvc`. More...

virtual StatusCode	sysStart () override
	Handle START transition. More...

virtual std::vector< Gaudi::DataHandle * >	inputHandles () const override
	Return this algorithm's input handles. More...

virtual std::vector< Gaudi::DataHandle * >	outputHandles () const override
	Return this algorithm's output handles. More...

Gaudi::Details::PropertyBase &	declareProperty (Gaudi::Property< T, V, H > &t)

Gaudi::Details::PropertyBase *	declareProperty (const std::string &name, SG::VarHandleKey &hndl, const std::string &doc, const SG::VarHandleKeyType &)
	Declare a new Gaudi property. More...

Gaudi::Details::PropertyBase *	declareProperty (const std::string &name, SG::VarHandleBase &hndl, const std::string &doc, const SG::VarHandleType &)
	Declare a new Gaudi property. More...

Gaudi::Details::PropertyBase *	declareProperty (const std::string &name, SG::VarHandleKeyArray &hndArr, const std::string &doc, const SG::VarHandleKeyArrayType &)

Gaudi::Details::PropertyBase *	declareProperty (const std::string &name, T &property, const std::string &doc, const SG::NotHandleType &)
	Declare a new Gaudi property. More...

Gaudi::Details::PropertyBase *	declareProperty (const std::string &name, T &property, const std::string &doc="none")
	Declare a new Gaudi property. More...

void	updateVHKA (Gaudi::Details::PropertyBase &)

MsgStream &	msg () const

MsgStream &	msg (const MSG::Level lvl) const

bool	msgLvl (const MSG::Level lvl) const

void	handle (const Incident &incident) override

Protected Member Functions
void	renounceArray (SG::VarHandleKeyArray &handlesArray)
	remove all handles from I/O resolution More...

std::enable_if_t< std::is_void_v< std::result_of_t< decltype(&T::renounce)(T)> > &&!std::is_base_of_v< SG::VarHandleKeyArray, T > &&std::is_base_of_v< Gaudi::DataHandle, T >, void >	renounce (T &h)

void	extraDeps_update_handler (Gaudi::Details::PropertyBase &ExtraDeps)
	Add StoreName to extra input/output deps as needed. More...

void	record_times (const size_t event_num, const std::vector< size_t > &times) const

template<class ... Args>
void	record_times (const size_t event_num, const size_t &value) const

template<class ... Args>
void	record_times (const size_t event_num, const size_t &value, Args &&... args) const

void	print_times (const std::string &header, const size_t time_size) const

Protected Attributes
std::shared_mutex	m_timeMutex
	Mutex that is locked when recording times. More...

Gaudi::Property< bool >	m_measureTimes
	If `true`, times are recorded to the file given by `m_timeFileName`. More...

Gaudi::Property< std::string >	m_timeFileName
	File to which times should be saved. More...

Private Types
typedef ServiceHandle< StoreGateSvc >	StoreGateSvc_t

Private Member Functions
Gaudi::Details::PropertyBase &	declareGaudiProperty (Gaudi::Property< T, V, H > &hndl, const SG::VarHandleKeyType &)
	specialization for handling Gaudi::Property<SG::VarHandleKey> More...

Gaudi::Details::PropertyBase &	declareGaudiProperty (Gaudi::Property< T, V, H > &hndl, const SG::VarHandleKeyArrayType &)
	specialization for handling Gaudi::Property<SG::VarHandleKeyArray> More...

Gaudi::Details::PropertyBase &	declareGaudiProperty (Gaudi::Property< T, V, H > &hndl, const SG::VarHandleType &)
	specialization for handling Gaudi::Property<SG::VarHandleBase> More...

Gaudi::Details::PropertyBase &	declareGaudiProperty (Gaudi::Property< T, V, H > &t, const SG::NotHandleType &)
	specialization for handling everything that's not a Gaudi::Property<SG::VarHandleKey> or a <SG::VarHandleKeyArray> More...

void	record_times_helper (const size_t) const

template<class Arg >
void	record_times_helper (const size_t index, Arg &&arg) const

template<class ... Args>
void	record_times_helper (size_t index, Args &&... args) const

Private Attributes
ToolHandle< ICaloClusterGPUConstantTransformer >	m_transformConstantData {this, "ConstantDataToGPUTool", "", "Tool for transforming the constant data and sending it to the GPU"}
	The tool that will convert the constant data from the CPU to the GPU. More...

ToolHandleArray< CaloClusterCollectionProcessor >	m_preGPUoperations {this, "BeforeGPUTools", {}, "Tools to be applied to the clusters on the CPU before processing them on the GPU"}
	Tools to be applied to the clusters before being sent to the GPU for processing. More...

ToolHandle< ICaloClusterGPUInputTransformer >	m_transformForGPU {this, "EventDataToGPUTool", "", "Tool for transforming the event data and sending it to the GPU"}
	The tool that will actually convert the data from the CPU to the GPU. More...

ToolHandleArray< CaloClusterGPUProcessor >	m_GPUoperations {this, "GPUTools", {}, "Tools to be applied to the clusters on the GPU"}
	Tools to be applied to the clusters on the GPU. More...

ToolHandle< ICaloClusterGPUOutputTransformer >	m_transformBackToCPU {this, "GPUToEventDataTool", {}, "Tool for getting the data from the GPU back to the CPU Athena data structures"}
	The tool that will convert the data from the GPU back to the CPU. More...

ToolHandleArray< CaloClusterCollectionProcessor >	m_postGPUoperations {this, "AfterGPUTools", {}, "Tools to be applied to the clusters on the CPU after returning from the GPU"}
	Tools to be applied to the clusters after returning from the GPU. More...

Gaudi::Property< bool >	m_doPlots {this, "DoPlots", false, "Do plots based on the plotter tool optionally provided."}
	If `true`, calls the plotter specified by `m_plotterTool` at every tool execution. More...

ToolHandle< ICaloClusterGPUPlotter >	m_plotterTool {this, "PlotterTool", "", "An optional plotter, for testing and/or debugging purposes"}
	An optional plotter, for testing and/or debugging purposes. More...

Gaudi::Property< bool >	m_writeTriggerSpecificInfo {this, "WriteTriggerSpecificInfo", false, "Write some trigger-specific decorations and use the trigger auxiliary container."}
	If `true`, writes some trigger-specific decorations. More...

SG::WriteDecorHandleKey< xAOD::CaloClusterContainer >	m_mDecor_ncells {this, "Decor_ncells", "nCells", "Decorator containing the number of cells associated to a cluster"}
	Key to the handle for writing the number of cells as a decoration. More...

Gaudi::Property< size_t >	m_numPreAllocatedGPUData {this, "NumPreAllocatedDataHolders", 0, "Number of event data holders to pre-allocate on GPU memory"}
	Number of events for which to pre-allocate space on GPU memory (should ideally be set to the expected number of threads to be run with). More...

SG::WriteHandleKey< xAOD::CaloClusterContainer >	m_clusterOutput {this, "ClustersOutputName", "", "The name of the key in StoreGate for the output CaloClusterContainer"}
	The name of the key in StoreGate for the output CaloClusterContainer. More...

SG::WriteHandleKey< CaloClusterCellLinkContainer >	m_clusterCellLinkOutput {this, "ClusterCellLinksOutputName", "", "The name of the key in StoreGate for the output CaloClusterCellLinkContainer"}
	The name of the key in StoreGate for the output CaloClusterCellLinkContainer. More...

Gaudi::Property< bool >	m_deferConstantDataToFirstEvent {this, "DeferConstantDataPreparationToFirstEvent", true, "Convert and send event data on first event instead of during initialize (needed for exporting geometry and noise properly)?"}
	If `true`, the constant data is only converted and sent to the GPU on the first event, in case not all the necessary information is available during the `initialize` phase. More...

Gaudi::Property< bool >	m_skipConversions {this, "SkipConversions", false, "If true, skip converting CPU to GPU data (useful if only instanting CPU tools)"}
	If `true`, both constant and event data conversion is skipped. More...

CaloRecGPU::Helpers::separate_thread_holder< CaloRecGPU::EventDataHolder > m_eventDataThreadedHolder	ATLAS_THREAD_SAFE
	A way to reduce allocations over multiple threads by keeping a cache of previously allocated objects that get assigned to the threads as they need them. More...

CaloRecGPU::Helpers::separate_thread_holder< simple_GPU_pointer_holder > m_temporariesThreadedHolder	ATLAS_THREAD_SAFE
	A way to reduce allocations over multiple threads by keeping a cache of previously allocated objects that get assigned to the threads as they need them. More...

size_t	m_temporariesSize
	The size of the temporary buffer to allocate for the GPU tools that will be called. More...

CaloRecGPU::ConstantDataHolder m_constantData	ATLAS_THREAD_SAFE
	Constant data, common for all events and persisted throughout the run. More...

std::atomic< bool >	m_constantDataSent
	A flag to signal that the constant data has been adequately sent to the GPU. More...

std::mutex	m_mutex
	This mutex is locked when sending the constant data on the first event to ensure thread safety. More...

DataObjIDColl	m_extendedExtraObjects
	Extra output dependency collection, extended by AthAlgorithmDHUpdate to add symlinks. More...

StoreGateSvc_t	m_evtStore
	Pointer to StoreGate (event store by default) More...

StoreGateSvc_t	m_detStore
	Pointer to StoreGate (detector store by default) More...

std::vector< SG::VarHandleKeyArray * >	m_vhka

bool	m_varHandleArraysDeclared

Detailed Description

Algorithm to reconstruct CaloCluster objects with GPU acceleration, providing interoperability for calling standard CPU algorithms before and after the GPU processing part.

Author: Nuno Fernandes nuno..nosp@m.dos..nosp@m.santo.nosp@m.s.fe.nosp@m.rnand.nosp@m.es@c.nosp@m.ern.c.nosp@m.h

Date: 27 May 2022 This class is meant as a replacement for CaloClusterMaker in that it creates a CaloClusterCollection and runs several tools over it. The main addition is the fact that, besides CPU-based tools, GPU-accelerated versions of the standard tools can be run, with adequate memory sharing between them to minimize transfers and (re-)conversions from and to the GPU-friendly data representation.

Definition at line 50 of file CaloGPUHybridClusterProcessor.h.

Member Typedef Documentation

◆ StoreGateSvc_t

typedef ServiceHandle<StoreGateSvc> AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::StoreGateSvc_t

privateinherited

Definition at line 388 of file AthCommonDataStore.h.

Constructor & Destructor Documentation

◆ CaloGPUHybridClusterProcessor()

CaloGPUHybridClusterProcessor::CaloGPUHybridClusterProcessor	(	const std::string &	name,
		ISvcLocator *	pSvcLocator
	)

Definition at line 31 of file CaloGPUHybridClusterProcessor.cxx.

                                                                                                              :
   AthReentrantAlgorithm(name, pSvcLocator),
   CaloGPUTimed(this),
   m_temporariesSize(0),
   m_constantDataSent(false)
 {
  
 }

◆ ~CaloGPUHybridClusterProcessor()

virtual CaloGPUHybridClusterProcessor::~CaloGPUHybridClusterProcessor ( )

overridevirtualdefault

Member Function Documentation

◆ cardinality()

unsigned int AthCommonReentrantAlgorithm< Gaudi::Algorithm >::cardinality

overridevirtualinherited

Cardinality (Maximum number of clones that can exist) special value 0 means that algorithm is reentrant.

Override this to return 0 for reentrant algorithms.

Definition at line 75 of file AthCommonReentrantAlgorithm.cxx.

 {
   return 0;
 }

◆ declareGaudiProperty() [1/4]

Gaudi::Details::PropertyBase& AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::declareGaudiProperty	(	Gaudi::Property< T, V, H > &	hndl,
		const SG::VarHandleKeyArrayType &
	)

inlineprivateinherited

specialization for handling Gaudi::Property<SG::VarHandleKeyArray>

Definition at line 170 of file AthCommonDataStore.h.

   {
     return *AthCommonDataStore<PBASE>::declareProperty(hndl.name(),
                                                        hndl.value(), 
                                                        hndl.documentation());
  
   }

◆ declareGaudiProperty() [2/4]

Gaudi::Details::PropertyBase& AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::declareGaudiProperty	(	Gaudi::Property< T, V, H > &	hndl,
		const SG::VarHandleKeyType &
	)

inlineprivateinherited

specialization for handling Gaudi::Property<SG::VarHandleKey>

Definition at line 156 of file AthCommonDataStore.h.

   {
     return *AthCommonDataStore<PBASE>::declareProperty(hndl.name(),
                                                        hndl.value(), 
                                                        hndl.documentation());
  
   }

◆ declareGaudiProperty() [3/4]

Gaudi::Details::PropertyBase& AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::declareGaudiProperty	(	Gaudi::Property< T, V, H > &	hndl,
		const SG::VarHandleType &
	)

inlineprivateinherited

specialization for handling Gaudi::Property<SG::VarHandleBase>

Definition at line 184 of file AthCommonDataStore.h.

   {
     return *AthCommonDataStore<PBASE>::declareProperty(hndl.name(),
                                                        hndl.value(), 
                                                        hndl.documentation());
   }

◆ declareGaudiProperty() [4/4]

Gaudi::Details::PropertyBase& AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::declareGaudiProperty	(	Gaudi::Property< T, V, H > &	t,
		const SG::NotHandleType &
	)

inlineprivateinherited

specialization for handling everything that's not a Gaudi::Property<SG::VarHandleKey> or a <SG::VarHandleKeyArray>

Definition at line 199 of file AthCommonDataStore.h.

   {
     return PBASE::declareProperty(t);
   }

◆ declareProperty() [1/6]

Gaudi::Details::PropertyBase* AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::declareProperty	(	const std::string &	name,
		SG::VarHandleBase &	hndl,
		const std::string &	doc,
		const SG::VarHandleType &
	)

inlineinherited

Declare a new Gaudi property.

Parameters

name	Name of the property.
hndl	Object holding the property value.
doc	Documentation string for the property.

This is the version for types that derive from SG::VarHandleBase. The property value object is put on the input and output lists as appropriate; then we forward to the base class.

Definition at line 245 of file AthCommonDataStore.h.

   {
     this->declare(hndl.vhKey());
     hndl.vhKey().setOwner(this);
  
     return PBASE::declareProperty(name,hndl,doc);
   }

◆ declareProperty() [2/6]

Gaudi::Details::PropertyBase* AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::declareProperty	(	const std::string &	name,
		SG::VarHandleKey &	hndl,
		const std::string &	doc,
		const SG::VarHandleKeyType &
	)

inlineinherited

Declare a new Gaudi property.

Parameters

name	Name of the property.
hndl	Object holding the property value.
doc	Documentation string for the property.

This is the version for types that derive from SG::VarHandleKey. The property value object is put on the input and output lists as appropriate; then we forward to the base class.

Definition at line 221 of file AthCommonDataStore.h.

   {
     this->declare(hndl);
     hndl.setOwner(this);
  
     return PBASE::declareProperty(name,hndl,doc);
   }

◆ declareProperty() [3/6]

Gaudi::Details::PropertyBase* AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::declareProperty	(	const std::string &	name,
		SG::VarHandleKeyArray &	hndArr,
		const std::string &	doc,
		const SG::VarHandleKeyArrayType &
	)

inlineinherited

Definition at line 259 of file AthCommonDataStore.h.

   {
  
     // std::ostringstream ost;
     // ost << Algorithm::name() << " VHKA declareProp: " << name 
     //     << " size: " << hndArr.keys().size() 
     //     << " mode: " << hndArr.mode() 
     //     << "  vhka size: " << m_vhka.size()
     //     << "\n";
     // debug() << ost.str() << endmsg;
  
     hndArr.setOwner(this);
     m_vhka.push_back(&hndArr);
  
     Gaudi::Details::PropertyBase* p =  PBASE::declareProperty(name, hndArr, doc);
     if (p != 0) {
       p->declareUpdateHandler(&AthCommonDataStore<PBASE>::updateVHKA, this);
     } else {
       ATH_MSG_ERROR("unable to call declareProperty on VarHandleKeyArray " 
                     << name);
     }
  
     return p;
  
   }

◆ declareProperty() [4/6]

Gaudi::Details::PropertyBase* AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::declareProperty	(	const std::string &	name,
		T &	property,
		const std::string &	doc,
		const SG::NotHandleType &
	)

inlineinherited

Declare a new Gaudi property.

Parameters

name	Name of the property.
property	Object holding the property value.
doc	Documentation string for the property.

This is the generic version, for types that do not derive from SG::VarHandleKey. It just forwards to the base class version of declareProperty.

Definition at line 333 of file AthCommonDataStore.h.

   {
     return PBASE::declareProperty(name, property, doc);
   }

◆ declareProperty() [5/6]

Gaudi::Details::PropertyBase* AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::declareProperty	(	const std::string &	name,
		T &	property,
		const std::string &	doc = `"none"`
	)

inlineinherited

Declare a new Gaudi property.

Parameters

name	Name of the property.
property	Object holding the property value.
doc	Documentation string for the property.

This dispatches to either the generic declareProperty or the one for VarHandle/Key/KeyArray.

Definition at line 352 of file AthCommonDataStore.h.

   {
     typedef typename SG::HandleClassifier<T>::type htype;
     return declareProperty (name, property, doc, htype());
   }

◆ declareProperty() [6/6]

Gaudi::Details::PropertyBase& AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::declareProperty ( Gaudi::Property< T, V, H > & t )

inlineinherited

Definition at line 145 of file AthCommonDataStore.h.

                                                                      {
     typedef typename SG::HandleClassifier<T>::type htype;
     return AthCommonDataStore<PBASE>::declareGaudiProperty(t, htype());
   }

◆ detStore()

const ServiceHandle<StoreGateSvc>& AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::detStore ( ) const

inlineinherited

The standard StoreGateSvc/DetectorStore Returns (kind of) a pointer to the StoreGateSvc.

Definition at line 95 of file AthCommonDataStore.h.

95 { return m_detStore; }

◆ evtStore() [1/2]

ServiceHandle<StoreGateSvc>& AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::evtStore ( )

inlineinherited

The standard StoreGateSvc (event store) Returns (kind of) a pointer to the StoreGateSvc.

Definition at line 85 of file AthCommonDataStore.h.

85 { return m_evtStore; }

◆ evtStore() [2/2]

const ServiceHandle<StoreGateSvc>& AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::evtStore ( ) const

inlineinherited

The standard StoreGateSvc (event store) Returns (kind of) a pointer to the StoreGateSvc.

Definition at line 90 of file AthCommonDataStore.h.

90 { return m_evtStore; }

◆ execute()

StatusCode CaloGPUHybridClusterProcessor::execute ( const EventContext & ctx ) const

overridevirtual

Definition at line 164 of file CaloGPUHybridClusterProcessor.cxx.

 {
   SG::WriteHandle<xAOD::CaloClusterContainer> cluster_collection (m_clusterOutput, ctx);
  
   if (m_writeTriggerSpecificInfo)
     {
       ATH_CHECK( cluster_collection.record (std::make_unique<xAOD::CaloClusterContainer>(),  std::make_unique<xAOD::CaloClusterTrigAuxContainer> () ));
     }
   else
     {
       ATH_CHECK(CaloClusterStoreHelper::AddContainerWriteHandle(cluster_collection));
     }
  
   //ATH_CHECK(CaloClusterStoreHelper::AddContainerWriteHandle(&(*evtStore()), cluster_collection, msg()));
  
   xAOD::CaloClusterContainer * cluster_collection_ptr = cluster_collection.ptr();
  
   if (m_deferConstantDataToFirstEvent && !m_skipConversions && !m_constantDataSent.load())
     {
       std::lock_guard<std::mutex> lock_guard(m_mutex);
       if (!m_constantDataSent.load())
         {
           ConstantDataHolder * cdh_ptr ATLAS_THREAD_SAFE = &m_constantData;
           ATH_CHECK( m_transformConstantData->convert(ctx, *cdh_ptr, m_doPlots) );
           m_constantDataSent.store(true);
         }
     }
  
   EventDataHolder * event_data_ptr = nullptr;
  
   Helpers::separate_thread_accessor<EventDataHolder> sep_th_acc_1(m_eventDataThreadedHolder, event_data_ptr);
   //This is a RAII wrapper to access an object held by Helpers::separate_thread_holder,
   //to ensure the event data is appropriately released when we are done processing.
  
   if (event_data_ptr == nullptr && !m_skipConversions)
     {
       ATH_MSG_ERROR("Could not get valid Event Data Holder! Event: " << ctx.evt() );
       return StatusCode::FAILURE;
     }
  
   if (!m_skipConversions)
     {
       event_data_ptr->allocate(true);
       //No-op if already allocated.
     }
  
   simple_GPU_pointer_holder * temporaries_data_ptr_holder = nullptr;
  
   Helpers::separate_thread_accessor<simple_GPU_pointer_holder> sep_th_acc_2(m_temporariesThreadedHolder, temporaries_data_ptr_holder);
   if (!temporaries_data_ptr_holder)
     {
       ATH_MSG_ERROR("temporaries_data_ptr_holder is null in CaloGPUHybridClusterProcessor::execute" );
       return StatusCode::FAILURE;
     }
   temporaries_data_ptr_holder->allocate(m_temporariesSize);
   //This will not perform any allocations if they've already been done.
  
   if ( (temporaries_data_ptr_holder->get_pointer() == nullptr) && !m_skipConversions && m_temporariesSize > 0 )
     {
       ATH_MSG_ERROR("Could not get valid temporary buffer holder! Event: " << ctx.evt() );
       return StatusCode::FAILURE;
     }
  
   const ConstantDataHolder & constant_data_holder ATLAS_THREAD_SAFE = m_constantData;
   //Just to shut up the checker. We know what we are doing...
  
   using clock_type = boost::chrono::thread_clock;
   auto time_cast = [](const auto & before, const auto & after)
   {
     return boost::chrono::duration_cast<boost::chrono::microseconds>(after - before).count();
   };
  
   std::vector<size_t> times;
  
   size_t plot_time = 0;
  
   if (m_measureTimes)
     {
       const size_t time_size = m_preGPUoperations.size() + m_GPUoperations.size() + m_postGPUoperations.size() + m_doPlots + 2 * !m_skipConversions;
       //+2 for the conversions
       //+1 for the plotter (only added at the end)
       times.reserve(time_size);
     }
  
   if (m_doPlots)
     {
       auto t1 = clock_type::now();
       ATH_CHECK( m_plotterTool->update_plots_start(ctx, constant_data_holder, cluster_collection_ptr) );
       auto t2 = clock_type::now();
       if (m_measureTimes)
         {
           plot_time += time_cast(t1, t2);
         }
     }
   
   for (const auto & pre_GPU_tool : m_preGPUoperations)
     {
       auto t1 = clock_type::now();
       ATH_CHECK( pre_GPU_tool->execute(ctx, cluster_collection_ptr) );
       auto t2 = clock_type::now();
       if (m_measureTimes)
         {
           times.push_back(time_cast(t1, t2));
         }
       if (m_doPlots)
         {
           auto t3 = clock_type::now();
           ATH_CHECK( m_plotterTool->update_plots(ctx, constant_data_holder, cluster_collection_ptr, pre_GPU_tool.get()) );
           auto t4 = clock_type::now();
           if (m_measureTimes)
             {
               plot_time += time_cast(t3, t4);
             }
         }
     }
  
   if (!m_skipConversions)
     {
       auto t3 = clock_type::now();
       ATH_CHECK( m_transformForGPU->convert(ctx, constant_data_holder, cluster_collection_ptr, *event_data_ptr) );
       auto t4 = clock_type::now();
       if (m_measureTimes)
         {
           times.push_back(time_cast(t3, t4));
         }
     }
  
   if (m_doPlots)
     {
       auto t1 = clock_type::now();
       ATH_CHECK( m_plotterTool->update_plots(ctx, constant_data_holder, cluster_collection_ptr, *event_data_ptr, m_transformForGPU.get()) );
       auto t2 = clock_type::now();
       if (m_measureTimes)
         {
           plot_time += time_cast(t1, t2);
         }
     }
  
   for (const auto & GPU_tool : m_GPUoperations)
     {
       auto t5 = clock_type::now();
       ATH_CHECK( GPU_tool->execute(ctx, constant_data_holder, *event_data_ptr, temporaries_data_ptr_holder->get_pointer()) );
       auto t6 = clock_type::now();
       if (m_measureTimes)
         {
           times.push_back(time_cast(t5, t6));
         }
       if (m_doPlots)
         {
           auto t3 = clock_type::now();
           ATH_CHECK( m_plotterTool->update_plots(ctx, constant_data_holder, cluster_collection_ptr, *event_data_ptr, GPU_tool.get()) );
           auto t4 = clock_type::now();
           if (m_measureTimes)
             {
               plot_time += time_cast(t3, t4);
             }
         }
     }
  
   if (!m_skipConversions)
     {
       auto t7 = clock_type::now();
       ATH_CHECK( m_transformBackToCPU->convert(ctx, constant_data_holder, *event_data_ptr, cluster_collection_ptr) );
       auto t8 = clock_type::now();
       if (m_measureTimes)
         {
           times.push_back(time_cast(t7, t8));
         }
     }
  
   if (m_doPlots)
     {
       auto t1 = clock_type::now();
       ATH_CHECK( m_plotterTool->update_plots(ctx, constant_data_holder, cluster_collection_ptr, *event_data_ptr, m_transformBackToCPU.get()) );
       auto t2 = clock_type::now();
       if (m_measureTimes)
         {
           plot_time += time_cast(t1, t2);
         }
     }
  
   for (const auto & post_GPU_tool : m_postGPUoperations)
     {
       auto t9 = clock_type::now();
       ATH_CHECK( post_GPU_tool->execute(ctx, cluster_collection_ptr) );
       auto t10 = clock_type::now();
       if (m_measureTimes)
         {
           times.push_back(time_cast(t9, t10));
         }
       if (m_doPlots)
         {
           auto t3 = clock_type::now();
           ATH_CHECK( m_plotterTool->update_plots(ctx, constant_data_holder, cluster_collection_ptr, post_GPU_tool.get()) );
           auto t4 = clock_type::now();
           if (m_measureTimes)
             {
               plot_time += time_cast(t3, t4);
             }
         }
     }
  
   if (m_doPlots)
     {
       auto t1 = clock_type::now();
       ATH_CHECK( m_plotterTool->update_plots_end(ctx, constant_data_holder, cluster_collection_ptr) );
       auto t2 = clock_type::now();
       if (m_measureTimes)
         {
           plot_time += time_cast(t1, t2);
         }
     }
  
   if (m_writeTriggerSpecificInfo)
     {
       SG::WriteDecorHandle<xAOD::CaloClusterContainer, int> decor_handle(m_mDecor_ncells, ctx);
  
       for (const xAOD::CaloCluster * cl : *cluster_collection_ptr)
         {
           const CaloClusterCellLink * cell_links = cl->getCellLinks();
           if (!cell_links)
             {
               decor_handle(*cl) = 0;
             }
           else
             {
               decor_handle(*cl) = cell_links->size();
             }
         }
     }
  
   ATH_MSG_DEBUG("Created cluster container with " << cluster_collection->size() << " clusters");
  
   SG::WriteHandle<CaloClusterCellLinkContainer> cell_links(m_clusterCellLinkOutput, ctx);
  
   ATH_CHECK( CaloClusterStoreHelper::finalizeClusters(cell_links, cluster_collection.ptr()) );
  
   if (m_measureTimes)
     {
       if (m_doPlots)
         {
           times.push_back(plot_time);
         }
       record_times(ctx.evt(), times);
     }
  
   return StatusCode::SUCCESS;
 }

◆ extraDeps_update_handler()

void AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::extraDeps_update_handler ( Gaudi::Details::PropertyBase & ExtraDeps )

protectedinherited

Add StoreName to extra input/output deps as needed.

use the logic of the VarHandleKey to parse the DataObjID keys supplied via the ExtraInputs and ExtraOuputs Properties to add the StoreName if it's not explicitly given

◆ extraOutputDeps()

const DataObjIDColl & AthCommonReentrantAlgorithm< Gaudi::Algorithm >::extraOutputDeps

overridevirtualinherited

Return the list of extra output dependencies.

This list is extended to include symlinks implied by inheritance relations.

Definition at line 94 of file AthCommonReentrantAlgorithm.cxx.

 {
   // If we didn't find any symlinks to add, just return the collection
   // from the base class.  Otherwise, return the extended collection.
   if (!m_extendedExtraObjects.empty()) {
     return m_extendedExtraObjects;
   }
   return BaseAlg::extraOutputDeps();
 }

◆ filterPassed()

virtual bool AthCommonReentrantAlgorithm< Gaudi::Algorithm >::filterPassed ( const EventContext & ctx ) const

inlinevirtualinherited

Definition at line 96 of file AthCommonReentrantAlgorithm.h.

                                                            {
     return execState( ctx ).filterPassed();
   }

◆ finalize()

StatusCode CaloGPUHybridClusterProcessor::finalize ( )

overridevirtual

Definition at line 413 of file CaloGPUHybridClusterProcessor.cxx.

 {
   if (m_measureTimes)
     {
       std::string header_string;
  
       auto add_name_to_string = [&](const auto & obj)
       {
         std::string rep = obj->name();
         std::replace(rep.begin(), rep.end(), ' ', '_');
         header_string += rep + " ";
       };
  
       for (const auto & pre_GPU_tool : m_preGPUoperations)
         {
           add_name_to_string(pre_GPU_tool);
         }
  
       if (!m_skipConversions)
         {
           add_name_to_string(m_transformForGPU);
         }
  
       for (const auto & GPU_tool : m_GPUoperations)
         {
           add_name_to_string(GPU_tool);
         }
  
       if (!m_skipConversions)
         {
           add_name_to_string(m_transformBackToCPU);
         }
  
       for (const auto & post_GPU_tool : m_postGPUoperations)
         {
           add_name_to_string(post_GPU_tool);
         }
  
       if (m_doPlots)
         {
           add_name_to_string(m_plotterTool);
         }
  
       print_times(header_string, m_preGPUoperations.size() + m_GPUoperations.size() + m_postGPUoperations.size() + 2 * !m_skipConversions + m_doPlots);
     }
  
   if (m_doPlots)
     {
       ATH_CHECK(m_plotterTool->finalize_plots());
     }
  
   return StatusCode::SUCCESS;
 }

◆ handle()

void CaloGPUCUDAInitialization::handle ( const Incident & incident )

inlineoverrideinherited

Definition at line 66 of file CaloGPUCUDAInitialization.h.

   {
     const bool is_multiprocess = (Gaudi::Concurrency::ConcurrencyFlags::numProcs() > 0);
     if (is_multiprocess && incident.type() == AthenaInterprocess::UpdateAfterFork::type())
       {
         if (!this->initialize_CUDA().isSuccess())
         {
           throw GaudiException("Failed to perform the CUDA initialization!",
                                "CaloGPUCUDAInitialization::handle",
                                StatusCode::FAILURE);
         }
       }
   }

◆ initialize()

virtual StatusCode CaloGPUHybridClusterProcessor::initialize ( )

inlineoverridevirtual

Reimplemented from CaloGPUCUDAInitialization.

Definition at line 57 of file CaloGPUHybridClusterProcessor.h.

   {
     return CaloGPUCUDAInitialization::initialize();
   }

◆ initialize_CUDA()

StatusCode CaloGPUHybridClusterProcessor::initialize_CUDA ( )

overridevirtual

Initialization that invokes CUDA functions.

Reimplemented from CaloGPUCUDAInitialization.

Definition at line 126 of file CaloGPUHybridClusterProcessor.cxx.

 {
   if (!m_deferConstantDataToFirstEvent && !m_skipConversions)
     {
       ATH_CHECK( m_transformConstantData->initialize() );
       //Not sure if this is needed or the tool will get initialized by this point.
  
       ATH_CHECK( m_transformConstantData->convert(m_constantData, m_doPlots) );
       m_constantDataSent = true;
     }
  
   if (size_t(m_numPreAllocatedGPUData) > 0)
     {
       ATH_MSG_INFO("Pre-allocating event data and temporary buffer for " << size_t(m_numPreAllocatedGPUData) << " parellel events.");
  
       m_eventDataThreadedHolder.resize(m_numPreAllocatedGPUData);
       m_temporariesThreadedHolder.resize(m_numPreAllocatedGPUData);
       //This will allocate the object holders.
  
       m_eventDataThreadedHolder.operate_on_all( [&](EventDataHolder & edh)
       {
         edh.allocate(true);
       }
                                               );
       m_temporariesThreadedHolder.operate_on_all( [&](simple_GPU_pointer_holder & ph)
       {
         ph.allocate(m_temporariesSize);
       }
                                                 );
       //This will allocate all the memory at this point.
       //Also useful to prevent/debug potential allocation issues?
       //But the main point is really reducing the execute times...
     }
  
   return StatusCode::SUCCESS;
 }

◆ initialize_non_CUDA()

StatusCode CaloGPUHybridClusterProcessor::initialize_non_CUDA ( )

overridevirtual

Initialization that does not invoke CUDA functions.

Reimplemented from CaloGPUCUDAInitialization.

Definition at line 40 of file CaloGPUHybridClusterProcessor.cxx.

 {
   ATH_CHECK( m_clusterOutput.initialize() );
  
   if (m_clusterCellLinkOutput.key().empty())
     {
       m_clusterCellLinkOutput = m_clusterOutput.key() + "_links";
     }
   ATH_CHECK( m_clusterCellLinkOutput.initialize() );
  
  
   bool any_failed = false;
  
   auto retrieve_and_report = [&](auto & var, const auto & type, bool & falsify_if_empty)
   {
     if (var.empty())
       {
         falsify_if_empty = false;
         ATH_MSG_DEBUG("There is nothing to retrieve for " << type << ".");
       }
     else if (var.retrieve().isFailure())
       {
         ATH_MSG_ERROR("Failed to retrieve " << type << ": " << var);
         any_failed = true;
       }
     else
       {
         ATH_MSG_DEBUG("Successfully retrieved " << type << ": " << var);
       }
   };
   //A generic lambda to prevent code repetition.
  
  
   bool checker = true;
  
   retrieve_and_report(m_preGPUoperations, "pre-GPU operations", checker);
   retrieve_and_report(m_GPUoperations, "GPU operations", checker);
   retrieve_and_report(m_postGPUoperations, "post-GPU operations", checker);
  
   if (!m_skipConversions)
     {
       retrieve_and_report(m_transformConstantData, "constant data to GPU transformer", checker);
       retrieve_and_report(m_transformForGPU, "event data to GPU transformer", checker);
       retrieve_and_report(m_transformBackToCPU, "GPU to Athena transformer", checker);
     }
   else
     {
       m_transformConstantData.disable();
       m_transformForGPU.disable();
       m_transformBackToCPU.disable();
     }
  
   if (m_doPlots)
     {
       checker = true;
       retrieve_and_report(m_plotterTool, "plotter tool", checker);
       m_doPlots = checker;
     }
   else
     {
       m_plotterTool.disable();
     }
  
   if (any_failed)
     {
       return StatusCode::FAILURE;
     }
  
   m_temporariesSize = 0;
  
   for (const auto & tool : m_GPUoperations)
     {
       m_temporariesSize = std::max(m_temporariesSize, tool->size_of_temporaries());
     }
  
   if (m_writeTriggerSpecificInfo)
     {
       m_mDecor_ncells = m_clusterOutput.key() + "." + m_mDecor_ncells.key();
     }
  
   ATH_CHECK(m_mDecor_ncells.initialize(m_writeTriggerSpecificInfo));
  
   return StatusCode::SUCCESS;
 }

◆ inputHandles()

virtual std::vector<Gaudi::DataHandle*> AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::inputHandles ( ) const

overridevirtualinherited

Return this algorithm's input handles.

We override this to include handle instances from key arrays if they have not yet been declared. See comments on updateVHKA.

◆ isClonable()

bool AthCommonReentrantAlgorithm< Gaudi::Algorithm >::isClonable

overridevirtualinherited

Specify if the algorithm is clonable.

Reentrant algorithms are clonable.

Definition at line 68 of file AthCommonReentrantAlgorithm.cxx.

 {
   // Reentrant algorithms are clonable.
   return true;
 }

◆ msg() [1/2]

MsgStream& AthCommonMsg< Gaudi::Algorithm >::msg ( ) const

inlineinherited

Definition at line 24 of file AthCommonMsg.h.

                                 {
     return this->msgStream();
   }

◆ msg() [2/2]

MsgStream& AthCommonMsg< Gaudi::Algorithm >::msg ( const MSG::Level lvl ) const

inlineinherited

Definition at line 27 of file AthCommonMsg.h.

                                                   {
     return this->msgStream(lvl);
   }

◆ msgLvl()

bool AthCommonMsg< Gaudi::Algorithm >::msgLvl ( const MSG::Level lvl ) const

inlineinherited

Definition at line 30 of file AthCommonMsg.h.

                                                {
     return this->msgLevel(lvl);
   }

◆ outputHandles()

virtual std::vector<Gaudi::DataHandle*> AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::outputHandles ( ) const

overridevirtualinherited

Return this algorithm's output handles.

We override this to include handle instances from key arrays if they have not yet been declared. See comments on updateVHKA.

◆ print_times()

void CaloGPUTimed::print_times	(	const std::string &	header,
		const size_t	time_size
	)		const

inlineprotectedinherited

Definition at line 143 of file CaloGPUTimed.h.

   {
     std::shared_lock<std::shared_mutex> lock(m_timeMutex);
     
     if (m_timeFileName.size() == 0)
       {
         return;
       }
     
     std::vector<size_t> indices(m_eventNumbers.size());
     
     std::iota(indices.begin(), indices.end(), 0);
     std::sort(indices.begin(), indices.end(), [&](size_t a, size_t b)
     {
       return m_eventNumbers[a] < m_eventNumbers[b];
     }
              );
     std::ofstream out(m_timeFileName);
  
     out << "Event_Number Total " << header << "\n";
  
     for (const size_t idx : indices)
       {
         out << m_eventNumbers[idx] << " ";
  
         size_t total = 0;
  
         for (size_t i = 0; i < time_size; ++i)
           {
             total += m_times[idx * time_size + i];
           }
  
         out << total << " ";
  
         for (size_t i = 0; i < time_size; ++i)
           {
             out << m_times[idx * time_size + i] << " ";
           }
         out << "\n";
       }
  
     out << std::endl;
  
     out.close();
   }

◆ record_times() [1/3]

template<class ... Args>

void CaloGPUTimed::record_times	(	const size_t	event_num,
		const size_t &	value
	)		const

inlineprotectedinherited

Definition at line 105 of file CaloGPUTimed.h.

   {
     const size_t time_size = 1;
  
     size_t old_size;
  
     {
       std::unique_lock<std::shared_mutex> lock(m_timeMutex);
       old_size = m_times.size();
       m_times.resize(old_size + time_size);
       m_eventNumbers.push_back(event_num);
     }
     {
       std::shared_lock<std::shared_mutex> lock(m_timeMutex);
       record_times_helper(old_size, value);
     }
   }

◆ record_times() [2/3]

template<class ... Args>

void CaloGPUTimed::record_times	(	const size_t	event_num,
		const size_t &	value,
		Args &&...	args
	)		const

inlineprotectedinherited

Definition at line 124 of file CaloGPUTimed.h.

   {
     const size_t time_size = sizeof...(args) + 1;
  
     size_t old_size;
  
     {
       std::unique_lock<std::shared_mutex> lock(m_timeMutex);
       old_size = m_times.size();
       m_times.resize(old_size + time_size);
       m_eventNumbers.push_back(event_num);
     }
     {
       std::shared_lock<std::shared_mutex> lock(m_timeMutex);
       record_times_helper(old_size, value, std::forward<Args>(args)...);
     }
  
   }

◆ record_times() [3/3]

void CaloGPUTimed::record_times	(	const size_t	event_num,
		const std::vector< size_t > &	times
	)		const

inlineprotectedinherited

Definition at line 86 of file CaloGPUTimed.h.

   {
     size_t old_size;
     {
       std::unique_lock<std::shared_mutex> lock(m_timeMutex);
       old_size = m_times.size();
       m_times.resize(old_size + times.size());
       m_eventNumbers.push_back(event_num);
     }
     {
       std::shared_lock<std::shared_mutex> lock(m_timeMutex);
       for (size_t i = 0; i < times.size(); ++i)
         {
           m_times[old_size + i] = times[i];
         }
     }
   }

◆ record_times_helper() [1/3]

template<class Arg >

void CaloGPUTimed::record_times_helper	(	const size_t	index,
		Arg &&	arg
	)		const

inlineprivateinherited

Definition at line 70 of file CaloGPUTimed.h.

   {
     // coverity[missing_lock]
     m_times[index] = std::forward<Arg>(arg);
     
     //This is called within a function that holds the lock itself.
   }

◆ record_times_helper() [2/3]

void CaloGPUTimed::record_times_helper ( const size_t ) const

inlineprivateinherited

Definition at line 64 of file CaloGPUTimed.h.

   {
     //Do nothing
   }

◆ record_times_helper() [3/3]

template<class ... Args>

void CaloGPUTimed::record_times_helper	(	size_t	index,
		Args &&...	args
	)		const

inlineprivateinherited

Definition at line 79 of file CaloGPUTimed.h.

   {
     (record_times_helper(index++, std::forward<Args>(args)), ...);
   }

◆ renounce()

std::enable_if_t<std::is_void_v<std::result_of_t<decltype(&T::renounce)(T)> > && !std::is_base_of_v<SG::VarHandleKeyArray, T> && std::is_base_of_v<Gaudi::DataHandle, T>, void> AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::renounce ( T & h )

inlineprotectedinherited

Definition at line 380 of file AthCommonDataStore.h.

   {
     h.renounce();
     PBASE::renounce (h);
   }

◆ renounceArray()

void AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::renounceArray ( SG::VarHandleKeyArray & handlesArray )

inlineprotectedinherited

remove all handles from I/O resolution

Definition at line 364 of file AthCommonDataStore.h.

                                                           {
     handlesArray.renounce();
   }

◆ setFilterPassed()

virtual void AthCommonReentrantAlgorithm< Gaudi::Algorithm >::setFilterPassed	(	bool	state,
		const EventContext &	ctx
	)		const

inlinevirtualinherited

Definition at line 100 of file AthCommonReentrantAlgorithm.h.

                                                                             {
     execState( ctx ).setFilterPassed( state );
   }

◆ sysExecute()

StatusCode AthCommonReentrantAlgorithm< Gaudi::Algorithm >::sysExecute ( const EventContext & ctx )

overridevirtualinherited

Execute an algorithm.

We override this in order to work around an issue with the Algorithm base class storing the event context in a member variable that can cause crashes in MT jobs.

Definition at line 85 of file AthCommonReentrantAlgorithm.cxx.

 {
   return BaseAlg::sysExecute (ctx);
 }

◆ sysInitialize()

StatusCode AthCommonReentrantAlgorithm< Gaudi::Algorithm >::sysInitialize

overridevirtualinherited

Override sysInitialize.

Override sysInitialize from the base class.

Loop through all output handles, and if they're WriteCondHandles, automatically register them and this Algorithm with the CondSvc

Scan through all outputHandles, and if they're WriteCondHandles, register them with the CondSvc

Reimplemented from AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >.

Reimplemented in InputMakerBase, and HypoBase.

Definition at line 61 of file AthCommonReentrantAlgorithm.cxx.

                                                                {
   StatusCode sc=AthCommonDataStore<AthCommonMsg<BaseAlg>>::sysInitialize();
  
   if (sc.isFailure()) {
     return sc;
   }
   
   ServiceHandle<ICondSvc> cs("CondSvc",name());
   for (auto h : outputHandles()) {
     if (h->isCondition() && h->mode() == Gaudi::DataHandle::Writer) {
       // do this inside the loop so we don't create the CondSvc until needed
       if ( cs.retrieve().isFailure() ) {
         ATH_MSG_WARNING("no CondSvc found: won't autoreg WriteCondHandles");
         return StatusCode::SUCCESS;
       }      
       if (cs->regHandle(this,*h).isFailure()) {
         sc = StatusCode::FAILURE;
         ATH_MSG_ERROR("unable to register WriteCondHandle " << h->fullKey()
                       << " with CondSvc");
       }
     }
   }
   return sc;  
 }

◆ sysStart()

virtual StatusCode AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::sysStart ( )

overridevirtualinherited

Handle START transition.

We override this in order to make sure that conditions handle keys can cache a pointer to the conditions container.

◆ updateVHKA()

void AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::updateVHKA ( Gaudi::Details::PropertyBase & )

inlineinherited

Definition at line 308 of file AthCommonDataStore.h.

                                                {
     // debug() << "updateVHKA for property " << p.name() << " " << p.toString() 
     //         << "  size: " << m_vhka.size() << endmsg;
     for (auto &a : m_vhka) {
       std::vector<SG::VarHandleKey*> keys = a->keys();
       for (auto k : keys) {
         k->setOwner(this);
       }
     }
   }

Member Data Documentation

◆ ATLAS_THREAD_SAFE [1/3]

CaloRecGPU::Helpers::separate_thread_holder<CaloRecGPU::EventDataHolder> m_eventDataThreadedHolder CaloGPUHybridClusterProcessor::ATLAS_THREAD_SAFE

mutableprivate

A way to reduce allocations over multiple threads by keeping a cache of previously allocated objects that get assigned to the threads as they need them.

It's all thread-safe due to an internal mutex ensuring no objects get assigned to different threads.

Definition at line 159 of file CaloGPUHybridClusterProcessor.h.

◆ ATLAS_THREAD_SAFE [2/3]

CaloRecGPU::Helpers::separate_thread_holder<simple_GPU_pointer_holder> m_temporariesThreadedHolder CaloGPUHybridClusterProcessor::ATLAS_THREAD_SAFE

mutableprivate

A way to reduce allocations over multiple threads by keeping a cache of previously allocated objects that get assigned to the threads as they need them.

It's all thread-safe due to an internal mutex ensuring no objects get assigned to different threads.

Definition at line 224 of file CaloGPUHybridClusterProcessor.h.

◆ ATLAS_THREAD_SAFE [3/3]

CaloRecGPU::ConstantDataHolder m_constantData CaloGPUHybridClusterProcessor::ATLAS_THREAD_SAFE

mutableprivate

Constant data, common for all events and persisted throughout the run.

Is mutable to deal with the cases where the data preparation is deferred to the first event.

Definition at line 239 of file CaloGPUHybridClusterProcessor.h.

◆ m_clusterCellLinkOutput

SG::WriteHandleKey<CaloClusterCellLinkContainer> CaloGPUHybridClusterProcessor::m_clusterCellLinkOutput {this, "ClusterCellLinksOutputName", "", "The name of the key in StoreGate for the output CaloClusterCellLinkContainer"}

private

The name of the key in StoreGate for the output CaloClusterCellLinkContainer.

Definition at line 142 of file CaloGPUHybridClusterProcessor.h.

◆ m_clusterOutput

SG::WriteHandleKey<xAOD::CaloClusterContainer> CaloGPUHybridClusterProcessor::m_clusterOutput {this, "ClustersOutputName", "", "The name of the key in StoreGate for the output CaloClusterContainer"}

private

The name of the key in StoreGate for the output CaloClusterContainer.

Definition at line 138 of file CaloGPUHybridClusterProcessor.h.

◆ m_constantDataSent

std::atomic<bool> CaloGPUHybridClusterProcessor::m_constantDataSent

mutableprivate

A flag to signal that the constant data has been adequately sent to the GPU.

This is required for everything to work properly in a multi-threaded context...

Definition at line 245 of file CaloGPUHybridClusterProcessor.h.

◆ m_deferConstantDataToFirstEvent

Gaudi::Property<bool> CaloGPUHybridClusterProcessor::m_deferConstantDataToFirstEvent {this, "DeferConstantDataPreparationToFirstEvent", true, "Convert and send event data on first event instead of during initialize (needed for exporting geometry and noise properly)?"}

private

If true, the constant data is only converted and sent to the GPU on the first event, in case not all the necessary information is available during the initialize phase.

Definition at line 148 of file CaloGPUHybridClusterProcessor.h.

◆ m_detStore

StoreGateSvc_t AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::m_detStore

privateinherited

Pointer to StoreGate (detector store by default)

Definition at line 393 of file AthCommonDataStore.h.

◆ m_doPlots

Gaudi::Property<bool> CaloGPUHybridClusterProcessor::m_doPlots {this, "DoPlots", false, "Do plots based on the plotter tool optionally provided."}

private

If true, calls the plotter specified by m_plotterTool at every tool execution.

It should be the plotter's responsibility to only take data from the tools it wants to.

Definition at line 113 of file CaloGPUHybridClusterProcessor.h.

◆ m_evtStore

StoreGateSvc_t AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::m_evtStore

privateinherited

Pointer to StoreGate (event store by default)

Definition at line 390 of file AthCommonDataStore.h.

◆ m_extendedExtraObjects

DataObjIDColl AthCommonReentrantAlgorithm< Gaudi::Algorithm >::m_extendedExtraObjects

privateinherited

Extra output dependency collection, extended by AthAlgorithmDHUpdate to add symlinks.

Empty if no symlinks were found.

Definition at line 114 of file AthCommonReentrantAlgorithm.h.

◆ m_GPUoperations

ToolHandleArray<CaloClusterGPUProcessor> CaloGPUHybridClusterProcessor::m_GPUoperations {this, "GPUTools", {}, "Tools to be applied to the clusters on the GPU"}

private

Tools to be applied to the clusters on the GPU.

Definition at line 95 of file CaloGPUHybridClusterProcessor.h.

◆ m_mDecor_ncells

SG::WriteDecorHandleKey<xAOD::CaloClusterContainer> CaloGPUHybridClusterProcessor::m_mDecor_ncells {this, "Decor_ncells", "nCells", "Decorator containing the number of cells associated to a cluster"}

private

Key to the handle for writing the number of cells as a decoration.

Definition at line 127 of file CaloGPUHybridClusterProcessor.h.

◆ m_measureTimes

Gaudi::Property<bool> CaloGPUTimed::m_measureTimes

protectedinherited

If true, times are recorded to the file given by m_timeFileName.

Defaults to false.

Definition at line 46 of file CaloGPUTimed.h.

◆ m_mutex

std::mutex CaloGPUHybridClusterProcessor::m_mutex

mutableprivate

This mutex is locked when sending the constant data on the first event to ensure thread safety.

Otherwise, it's unused.

Definition at line 251 of file CaloGPUHybridClusterProcessor.h.

◆ m_numPreAllocatedGPUData

Gaudi::Property<size_t> CaloGPUHybridClusterProcessor::m_numPreAllocatedGPUData {this, "NumPreAllocatedDataHolders", 0, "Number of event data holders to pre-allocate on GPU memory"}

private

Number of events for which to pre-allocate space on GPU memory (should ideally be set to the expected number of threads to be run with).

Definition at line 134 of file CaloGPUHybridClusterProcessor.h.

◆ m_plotterTool

ToolHandle<ICaloClusterGPUPlotter> CaloGPUHybridClusterProcessor::m_plotterTool {this, "PlotterTool", "", "An optional plotter, for testing and/or debugging purposes"}

private

An optional plotter, for testing and/or debugging purposes.

Definition at line 119 of file CaloGPUHybridClusterProcessor.h.

◆ m_postGPUoperations

ToolHandleArray<CaloClusterCollectionProcessor> CaloGPUHybridClusterProcessor::m_postGPUoperations {this, "AfterGPUTools", {}, "Tools to be applied to the clusters on the CPU after returning from the GPU"}

private

Tools to be applied to the clusters after returning from the GPU.

Definition at line 107 of file CaloGPUHybridClusterProcessor.h.

◆ m_preGPUoperations

ToolHandleArray<CaloClusterCollectionProcessor> CaloGPUHybridClusterProcessor::m_preGPUoperations {this, "BeforeGPUTools", {}, "Tools to be applied to the clusters on the CPU before processing them on the GPU"}

private

Tools to be applied to the clusters before being sent to the GPU for processing.

Definition at line 82 of file CaloGPUHybridClusterProcessor.h.

◆ m_skipConversions

Gaudi::Property<bool> CaloGPUHybridClusterProcessor::m_skipConversions {this, "SkipConversions", false, "If true, skip converting CPU to GPU data (useful if only instanting CPU tools)"}

private

If true, both constant and event data conversion is skipped.

Definition at line 153 of file CaloGPUHybridClusterProcessor.h.

◆ m_temporariesSize

size_t CaloGPUHybridClusterProcessor::m_temporariesSize

private

The size of the temporary buffer to allocate for the GPU tools that will be called.

Will correspond to the maximum of all the necessary sizes for all the GPU tools.

Warning: Every tool should consider the buffer as filled with uninitialized memory at their start.

Definition at line 232 of file CaloGPUHybridClusterProcessor.h.

◆ m_timeFileName

Gaudi::Property<std::string> CaloGPUTimed::m_timeFileName

protectedinherited

File to which times should be saved.

Definition at line 50 of file CaloGPUTimed.h.

◆ m_timeMutex

std::shared_mutex CaloGPUTimed::m_timeMutex

mutableprotectedinherited

Mutex that is locked when recording times.

Definition at line 32 of file CaloGPUTimed.h.

◆ m_transformBackToCPU

ToolHandle<ICaloClusterGPUOutputTransformer> CaloGPUHybridClusterProcessor::m_transformBackToCPU {this, "GPUToEventDataTool", {}, "Tool for getting the data from the GPU back to the CPU Athena data structures"}

private

The tool that will convert the data from the GPU back to the CPU.

Definition at line 101 of file CaloGPUHybridClusterProcessor.h.

◆ m_transformConstantData

ToolHandle<ICaloClusterGPUConstantTransformer> CaloGPUHybridClusterProcessor::m_transformConstantData {this, "ConstantDataToGPUTool", "", "Tool for transforming the constant data and sending it to the GPU"}

private

The tool that will convert the constant data from the CPU to the GPU.

Definition at line 76 of file CaloGPUHybridClusterProcessor.h.

◆ m_transformForGPU

ToolHandle<ICaloClusterGPUInputTransformer> CaloGPUHybridClusterProcessor::m_transformForGPU {this, "EventDataToGPUTool", "", "Tool for transforming the event data and sending it to the GPU"}

private

The tool that will actually convert the data from the CPU to the GPU.

Definition at line 89 of file CaloGPUHybridClusterProcessor.h.

◆ m_varHandleArraysDeclared

bool AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::m_varHandleArraysDeclared

privateinherited

Definition at line 399 of file AthCommonDataStore.h.

◆ m_vhka

std::vector<SG::VarHandleKeyArray*> AthCommonDataStore< AthCommonMsg< Gaudi::Algorithm > >::m_vhka

privateinherited

Definition at line 398 of file AthCommonDataStore.h.

◆ m_writeTriggerSpecificInfo

Gaudi::Property<bool> CaloGPUHybridClusterProcessor::m_writeTriggerSpecificInfo {this, "WriteTriggerSpecificInfo", false, "Write some trigger-specific decorations and use the trigger auxiliary container."}

private

If true, writes some trigger-specific decorations.

Definition at line 123 of file CaloGPUHybridClusterProcessor.h.

The documentation for this class was generated from the following files:

Classes

Public Member Functions

Protected Member Functions

Protected Attributes

Private Types

Private Member Functions

Private Attributes

Detailed Description

Member Typedef Documentation

◆ StoreGateSvc_t

Constructor & Destructor Documentation

◆ CaloGPUHybridClusterProcessor()

◆ ~CaloGPUHybridClusterProcessor()

Member Function Documentation

◆ cardinality()

◆ declareGaudiProperty() [1/4]

◆ declareGaudiProperty() [2/4]

◆ declareGaudiProperty() [3/4]

◆ declareGaudiProperty() [4/4]

◆ declareProperty() [1/6]

◆ declareProperty() [2/6]

◆ declareProperty() [3/6]

◆ declareProperty() [4/6]

◆ declareProperty() [5/6]

◆ declareProperty() [6/6]

◆ detStore()

◆ evtStore() [1/2]

◆ evtStore() [2/2]

◆ execute()

◆ extraDeps_update_handler()

◆ extraOutputDeps()

◆ filterPassed()

◆ finalize()

◆ handle()

◆ initialize()

◆ initialize_CUDA()

◆ initialize_non_CUDA()

◆ inputHandles()

◆ isClonable()

◆ msg() [1/2]

◆ msg() [2/2]

◆ msgLvl()

◆ outputHandles()

◆ print_times()

◆ record_times() [1/3]

◆ record_times() [2/3]

◆ record_times() [3/3]

◆ record_times_helper() [1/3]

◆ record_times_helper() [2/3]

◆ record_times_helper() [3/3]

◆ renounce()

◆ renounceArray()

◆ setFilterPassed()

◆ sysExecute()

◆ sysInitialize()

◆ sysStart()

◆ updateVHKA()

Member Data Documentation

◆ ATLAS_THREAD_SAFE [1/3]

◆ ATLAS_THREAD_SAFE [2/3]

◆ ATLAS_THREAD_SAFE [3/3]

◆ m_clusterCellLinkOutput

◆ m_clusterOutput

◆ m_constantDataSent

◆ m_deferConstantDataToFirstEvent

◆ m_detStore

◆ m_doPlots

◆ m_evtStore

◆ m_extendedExtraObjects

◆ m_GPUoperations

◆ m_mDecor_ncells

◆ m_measureTimes

◆ m_mutex

◆ m_numPreAllocatedGPUData

◆ m_plotterTool

◆ m_postGPUoperations

◆ m_preGPUoperations

◆ m_skipConversions

◆ m_temporariesSize

◆ m_timeFileName