uni20::async::TbbCudaScheduler Class Reference

Unified oneTBB scheduler with one host arena and one arena per CUDA device. More...

#include <tbb_cuda_scheduler.hpp>

Inheritance diagram for uni20::async::TbbCudaScheduler:

Collaboration diagram for uni20::async::TbbCudaScheduler:

Public Member Functions
bool	has_device (int device) const noexcept
	Report whether a CUDA device has an execution arena in this scheduler.
TbbCudaScheduler &	operator= (TbbCudaScheduler const &)=delete
void	schedule (AsyncTask &&t) override
	Schedule a coroutine for initial execution.
void	schedule (CudaTask &&task) override
	Submit a CUDA task without establishing device affinity.
void	schedule (CudaTask &&task, int device) override
	Bind and submit a CUDA task to an enrolled device arena.
	TbbCudaScheduler (std::vector< cuda::Device > devices, TbbCudaSchedulerOptions options={})
	Construct a scheduler for an explicit set of CUDA devices.
	TbbCudaScheduler (TbbCudaScheduler const &)=delete
	TbbCudaScheduler (TbbCudaSchedulerOptions options={})
	Construct a scheduler enrolling every visible CUDA device.
	~TbbCudaScheduler () noexcept override
Public Member Functions inherited from uni20::async::TbbScheduler
void	help_while_waiting (const WaitPredicate &is_ready) override
	Allow a scheduler to advance queued work while a thread is blocking.
void	pause () override
	Pause the scheduler. Don't execute scheduled tasks, but instead add them to a queue.
void	resume () override
	Unpause the scheduler, and execute any tasks that have been queued.
void	run_all ()
	Block until all tasks scheduled on this scheduler are complete.
void	schedule (AsyncTask &&t) override
	Schedule a coroutine for initial execution.
	TbbScheduler (int max_concurrency=oneapi::tbb::task_arena::automatic, TbbSchedulerBatchOptions batch_options={})
	Construct a TBB scheduler with a given arena concurrency.
	TbbScheduler (oneapi::tbb::task_arena::constraints constraints, TbbSchedulerBatchOptions batch_options={})
	Construct a TBB scheduler constrained to a specific NUMA node.
void	wait_for (const WaitPredicate &is_ready) override
	Accept an already-satisfied predicate; pending TBB waits require a WaitRequest notification.
void	wait_for (WaitRequest const &request) override
	Wait for targeted readiness while allowing the arena to execute other work.
	~TbbScheduler () noexcept override
Public Member Functions inherited from uni20::async::IAsyncScheduler
template<class Function > requires std::invocable<Function&, std::size_t>
void	execute_batch (std::size_t size, Function &&function)
	Execute a synchronous batch of independent lightweight work items.
	~IAsyncScheduler () override=default
	Virtual destructor.
Public Member Functions inherited from uni20::async::IScheduler
virtual	~IScheduler ()=default
	Virtual destructor.
Public Member Functions inherited from uni20::async::ICudaScheduler
	~ICudaScheduler () override=default

Additional Inherited Members
Public Types inherited from uni20::async::IAsyncScheduler
using	WaitPredicate = std::function< bool()>
using	WaitWakeup = std::function< void()>
using	WaitWakeupRegistration = std::function< void(WaitWakeup)>
Protected Member Functions inherited from uni20::async::TbbScheduler
template<typename Prepare >
void	dispatch_handle_in_arena (oneapi::tbb::task_arena &arena, TaskHandle handle, detail::TbbTaskAdmissionContext context, Prepare &&prepare)
	Submit one activation to an arena through the shared task group.
void	enqueue_task (BasicTask &&t)
	Enqueue one task through the scheduler's domain-routing hook.
void	wait_for_submitted_tasks ()
	Wait for every activation registered with the shared task group.

Detailed Description

Unified oneTBB scheduler with one host arena and one arena per CUDA device.

Ordinary and CUDA tasks share scheduler ownership, task-group accounting, pause state, waits, and rescheduling. CUDA activations are routed by the promise's optional affinity or the scheduler's default CUDA device.

Constructor & Destructor Documentation

TbbCudaScheduler() [1/3]

uni20::async::TbbCudaScheduler::TbbCudaScheduler ( TbbCudaSchedulerOptions  options = {} )

inlineexplicit

Construct a scheduler enrolling every visible CUDA device.

Parameters

options

Host and per-device arena configuration.

TbbCudaScheduler() [2/3]

uni20::async::TbbCudaScheduler::TbbCudaScheduler ( std::vector< cuda::Device >  devices, TbbCudaSchedulerOptions  options = {}  )

inlineexplicit

Construct a scheduler for an explicit set of CUDA devices.

Parameters

devices	Validated devices to enroll.
options	Host and per-device arena configuration.

TbbCudaScheduler() [3/3]

uni20::async::TbbCudaScheduler::TbbCudaScheduler ( TbbCudaScheduler const &  )

delete

~TbbCudaScheduler()

uni20::async::TbbCudaScheduler::~TbbCudaScheduler ( )

inlineoverridenoexcept

Member Function Documentation

has_device()

bool uni20::async::TbbCudaScheduler::has_device ( int  device ) const

inlinenoexcept

Report whether a CUDA device has an execution arena in this scheduler.

operator=()

TbbCudaScheduler & uni20::async::TbbCudaScheduler::operator= ( TbbCudaScheduler const &  )

delete

schedule() [1/3]

void uni20::async::TbbScheduler::schedule ( AsyncTask &&  t )

inlineoverridevirtual

Schedule a coroutine for initial execution.

Implements uni20::async::IAsyncScheduler.

schedule() [2/3]

void uni20::async::TbbCudaScheduler::schedule ( CudaTask &&  task )

inlineoverridevirtual

Submit a CUDA task without establishing device affinity.

Parameters

task	CUDA task to admit through the default device arena.

Implements uni20::async::ICudaScheduler.

schedule() [3/3]

void uni20::async::TbbCudaScheduler::schedule ( CudaTask &&  task, int  device  )

inlineoverridevirtual

Bind and submit a CUDA task to an enrolled device arena.

Parameters

task	CUDA task to admit.
device	CUDA runtime device ordinal selected for the task.

Implements uni20::async::ICudaScheduler.

---

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

• tbb_cuda_scheduler.hpp