Abstract Task specialized for CUDA computation. More...

#include "abstract_cuda_task.h"

Inheritance diagram for hh::AbstractCUDATask< Separator, AllTypes >:

Collaboration diagram for hh::AbstractCUDATask< Separator, AllTypes >:

Public Member Functions
	AbstractCUDATask (std::string const &name, size_t numberThreads, bool enablePeerAccess, bool automaticStart=false)
	AbstractCUDATask full constructor.

	AbstractCUDATask (std::string const name="CudaTask", size_t numberThreads=1)
	Main constructor for a AbstractCUDATask.

	AbstractCUDATask (std::shared_ptr< hh::core::CoreTask< Separator, AllTypes... > > coreTask, bool enablePeerAccess)
	Custom core task constructor.

	~AbstractCUDATask () override
	Default destructor.

void	initialize () final
	Initialize an AbstractCUDATask to bind it to a CUDA device, and do the peer access if enabled. At the end will call AbstractCUDATask::initializeCuda.

void	shutdown () final
	Shutdown an AbstractCUDATask to destroy the task's CUDA stream created during AbstractCUDATask::initialize. First calls AbstractCUDATask::shutdownCuda.

virtual void	initializeCuda ()
	Virtual initialization step, where user defined data structure can be initialized.

virtual void	shutdownCuda ()
	Virtual shutdown step, where user defined data structure can be destroyed.

bool	enablePeerAccess () const
	Accessor for peer access choice.

cudaStream_t	stream () const
	Getter for CUDA task's stream.

bool	hasPeerAccess (int peerDeviceId)
	Accessor for peer access enabled for a specific device id.

Public Member Functions inherited from hh::AbstractTask< Separator, AllTypes... >
	AbstractTask ()
	Default constructor (Task with one thread named "Task")

	AbstractTask (std::string const &name, size_t const numberThreads=1, bool const automaticStart=false)
	AbstractTask main constructor.

	AbstractTask (std::shared_ptr< hh::core::CoreTask< Separator, AllTypes... > > coreTask)
	Construct a task from a user-defined core.

	~AbstractTask () override=default
	Default task destructor.

size_t	graphId () const
	Belonging graph id accessor.

bool	canTerminate () const override
	Default termination rule, it terminates if there is no predecessor connection and there is no input data.

Public Member Functions inherited from hh::behavior::TaskNode
	TaskNode (std::shared_ptr< hh::core::abstraction::TaskNodeAbstraction > core)
	Constructor using a core.

	~TaskNode () override=default
	Default destructor.

std::shared_ptr< AbstractMemoryManager > const &	memoryManager () const
	Memory manager accessor.

void	connectMemoryManager (std::shared_ptr< AbstractMemoryManager > mm)
	Connect a memory manager to a task.

std::shared_ptr< ManagedMemory >	getManagedMemory ()
	Get a managed memory for the memory manager attached to the task, can block if none are available at the time of the call.

virtual void	initialize ()
	initialize step for the task

virtual void	shutdown ()
	shutdown step for the task

virtual std::string	extraPrintingInformation () const
	Print extra information for the task.

Public Member Functions inherited from hh::behavior::Node
	Node (std::shared_ptr< hh::core::abstraction::NodeAbstraction > core)
	Constructor's node.

virtual	~Node ()=default
	Default destructor.

std::shared_ptr< hh::core::abstraction::NodeAbstraction > const &	core () const
	Core accessor.

std::string	name () const
	Node's name accessor.

Public Member Functions inherited from hh::behavior::CanTerminate
	CanTerminate ()=default
	Default constructor.

virtual	~CanTerminate ()=default
	Default destructor.

virtual bool	canTerminate () const =0
	Termination condition.

Public Member Functions inherited from hh::behavior::Cleanable
	Cleanable ()=default
	Default constructor.

	~Cleanable ()=default
	Default destructor.

virtual void	clean ()
	Clean a node.

Public Member Functions inherited from hh::behavior::Copyable< AbstractTask< Separator, AllTypes... > >
	Copyable (size_t const numberThreads)
	Copyable constructor, set the number of threads for a node.

virtual	~Copyable ()=default
	Default destructor.

size_t	numberThreads () const
	Number of threads accessor.

std::vector< AbstractTask< Separator, AllTypes... > const * >	group () const
	Get the group of nodes that hold the current nodes.

virtual std::shared_ptr< AbstractTask< Separator, AllTypes... > >	copy ()
	Copy method called to either create a group of node or duplicate a node when an execution pipeline is created.

Private Attributes
bool	enablePeerAccess_ = false
	Enable CUDA Peer Access through all CUDA devices available.

std::unordered_set< int >	peerDeviceIds_ = {}
	Sparse matrix of linked CUDA devices.

cudaStream_t	stream_ = {}
	CUDA stream linked to the task.

Additional Inherited Members
Protected Member Functions inherited from hh::AbstractTask< Separator, AllTypes... >
std::shared_ptr< hh::core::CoreTask< Separator, AllTypes... > > const &	coreTask () const
	Accessor to the core task.

int	deviceId ()
	Accessor to device id linked to the task (default 0 for CPU task)

Detailed Description

template<size_t Separator, class ... AllTypes>
class hh::AbstractCUDATask< Separator, AllTypes >

Abstract Task specialized for CUDA computation.

At initialization, the device is set to the task (cudaSetDevice), and a stream is created and bound to the task (cudaStreamCreate). During shutdown, the stream is destroyed (cudaStreamDestroy).

Virtual functions: Execute::execute (one for each of TaskInputs)
AbstractCUDATask::copy (only used if number of threads is greater than 1 or used in an ExecutionPipeline)
AbstractCUDATask::initializeCuda
AbstractCUDATask::shutdownCuda
Node::canTerminate
Node::extraPrintingInformation

Template Parameters

Separator	Separator position between input types and output types
AllTypes	List of input and output types

Definition at line 46 of file abstract_cuda_task.h.

Constructor & Destructor Documentation

◆ AbstractCUDATask() [1/3]

template<size_t Separator, class ... AllTypes>

hh::AbstractCUDATask< Separator, AllTypes >::AbstractCUDATask	(	std::string const &	name,
		size_t	numberThreads,
		bool	enablePeerAccess,
		bool	automaticStart = `false`
	)

inline

AbstractCUDATask full constructor.

Parameters

name	Task name
numberThreads	Number of threads for the task
enablePeerAccess	Enable peer access for NVIDIA GPUs
automaticStart	Flag for automatic start (Cf. AbstractTask)

Definition at line 58 of file abstract_cuda_task.h.

Here is the call graph for this function:

◆ AbstractCUDATask() [2/3]

template<size_t Separator, class ... AllTypes>

hh::AbstractCUDATask< Separator, AllTypes >::AbstractCUDATask	(	std::string const	name = `"CudaTask"`,
		size_t	numberThreads = `1`
	)

inlineexplicit

Main constructor for a AbstractCUDATask.

Parameters

name	Name of the AbstractCUDATask
numberThreads	Number of thread for this task (default 1)

Definition at line 68 of file abstract_cuda_task.h.

◆ AbstractCUDATask() [3/3]

template<size_t Separator, class ... AllTypes>

hh::AbstractCUDATask< Separator, AllTypes >::AbstractCUDATask	(	std::shared_ptr< hh::core::CoreTask< Separator, AllTypes... > >	coreTask,
		bool	enablePeerAccess
	)

inline

Custom core task constructor.

Parameters

coreTask	Custom core to use
enablePeerAccess	Enable per access for NVIDIA GPUs

Definition at line 74 of file abstract_cuda_task.h.

Here is the call graph for this function:

◆ ~AbstractCUDATask()

template<size_t Separator, class ... AllTypes>

hh::AbstractCUDATask< Separator, AllTypes >::~AbstractCUDATask ( )

inlineoverride

Default destructor.

Definition at line 82 of file abstract_cuda_task.h.

Here is the call graph for this function:

Member Function Documentation

◆ enablePeerAccess()

template<size_t Separator, class ... AllTypes>

bool hh::AbstractCUDATask< Separator, AllTypes >::enablePeerAccess ( ) const

inline

Accessor for peer access choice.

Returns: True if peer access is enabled, else False

Definition at line 136 of file abstract_cuda_task.h.

◆ hasPeerAccess()

template<size_t Separator, class ... AllTypes>

bool hh::AbstractCUDATask< Separator, AllTypes >::hasPeerAccess ( int peerDeviceId )

inline

Accessor for peer access enabled for a specific device id.

Parameters

peerDeviceId Device id to test

Returns: True if peer access enable for device id peerDeviceId, else False

Definition at line 145 of file abstract_cuda_task.h.

◆ initialize()

template<size_t Separator, class ... AllTypes>

void hh::AbstractCUDATask< Separator, AllTypes >::initialize ( )

inlinefinalvirtual

Initialize an AbstractCUDATask to bind it to a CUDA device, and do the peer access if enabled. At the end will call AbstractCUDATask::initializeCuda.

Reimplemented from hh::behavior::TaskNode.

Definition at line 90 of file abstract_cuda_task.h.

Here is the call graph for this function:

◆ initializeCuda()

template<size_t Separator, class ... AllTypes>

virtual void hh::AbstractCUDATask< Separator, AllTypes >::initializeCuda ( )

inlinevirtual

Virtual initialization step, where user defined data structure can be initialized.

Definition at line 129 of file abstract_cuda_task.h.

Here is the caller graph for this function:

◆ shutdown()

template<size_t Separator, class ... AllTypes>

void hh::AbstractCUDATask< Separator, AllTypes >::shutdown ( )

inlinefinalvirtual

Shutdown an AbstractCUDATask to destroy the task's CUDA stream created during AbstractCUDATask::initialize. First calls AbstractCUDATask::shutdownCuda.

Reimplemented from hh::behavior::TaskNode.

Definition at line 123 of file abstract_cuda_task.h.

Here is the call graph for this function:

◆ shutdownCuda()

template<size_t Separator, class ... AllTypes>

virtual void hh::AbstractCUDATask< Separator, AllTypes >::shutdownCuda ( )

inlinevirtual

Virtual shutdown step, where user defined data structure can be destroyed.

Definition at line 132 of file abstract_cuda_task.h.

Here is the caller graph for this function:

◆ stream()

template<size_t Separator, class ... AllTypes>

cudaStream_t hh::AbstractCUDATask< Separator, AllTypes >::stream ( ) const

inline

Getter for CUDA task's stream.

Returns: CUDA stream

Definition at line 140 of file abstract_cuda_task.h.

Member Data Documentation

◆ enablePeerAccess_

template<size_t Separator, class ... AllTypes>

bool hh::AbstractCUDATask< Separator, AllTypes >::enablePeerAccess_ = false

private

Enable CUDA Peer Access through all CUDA devices available.

Definition at line 48 of file abstract_cuda_task.h.

◆ peerDeviceIds_

template<size_t Separator, class ... AllTypes>

std::unordered_set<int> hh::AbstractCUDATask< Separator, AllTypes >::peerDeviceIds_ = {}

private

Sparse matrix of linked CUDA devices.

Definition at line 49 of file abstract_cuda_task.h.

◆ stream_

template<size_t Separator, class ... AllTypes>

cudaStream_t hh::AbstractCUDATask< Separator, AllTypes >::stream_ = {}

private

CUDA stream linked to the task.

Definition at line 50 of file abstract_cuda_task.h.

Public Member Functions

Private Attributes

Additional Inherited Members

Detailed Description

Constructor & Destructor Documentation

◆ AbstractCUDATask() [1/3]

◆ AbstractCUDATask() [2/3]

◆ AbstractCUDATask() [3/3]

◆ ~AbstractCUDATask()

Member Function Documentation

◆ enablePeerAccess()

◆ hasPeerAccess()

◆ initialize()

◆ initializeCuda()

◆ shutdown()

◆ shutdownCuda()

◆ stream()

Member Data Documentation

◆ enablePeerAccess_

◆ peerDeviceIds_

◆ stream_