User-exposed API of Cooperative groups

Functions

User-exposed API of Cooperative groups

HIP API » Cooperative Groups

Collaboration diagram for User-exposed API of Cooperative groups:

Functions
template<class CGTy >
__hip_uint32_t	cooperative_groups::group_size (CGTy const &g)
	Returns the size of the group.
template<class CGTy >
__hip_uint32_t	cooperative_groups::thread_rank (CGTy const &g)
	Returns the rank of thread of the group.
template<class CGTy >
bool	cooperative_groups::is_valid (CGTy const &g)
	Returns true if the group has not violated any API constraints.
template<class CGTy >
void	cooperative_groups::sync (CGTy const &g)
	Synchronizes the threads in the group.
template<typename Group , typename Fn , typename... Args>
void	cooperative_groups::invoke_one (const Group &g, Fn &&fn, Args &&... args)
	Calls an invocable on a single thread of the group.
template<typename Fn , typename... Args>
auto	cooperative_groups::invoke_one_broadcast (const coalesced_group &g, Fn &&fn, Args &&... args) -> typename __hip_internal::remove_reference< decltype(impl::forward< Fn >(fn)(impl::forward< Args >(args)...))>::type
	Calls an invocable on a single thread of the group and broadcasts the result.
template<unsigned int Size, typename Parent , typename Fn , typename... Args>
auto	cooperative_groups::invoke_one_broadcast (const thread_block_tile< Size, Parent > &g, Fn &&fn, Args &&... args) -> typename __hip_internal::remove_reference< decltype(impl::forward< Fn >(fn)(impl::forward< Args >(args)...))>::type
	Calls an invocable on a single thread of the group and broadcasts the result.

Detailed Description

Function Documentation

◆ group_size()

template<class CGTy >

__hip_uint32_t cooperative_groups::group_size

(

CGTy const &

g

)

Returns the size of the group.

Total number of threads in the thread group, and this serves the purpose for all derived cooperative group types because their size is directly saved during the construction.

Template Parameters

CGTy	The cooperative group class template parameter.

Parameters

g	[in] The cooperative group for size returns.

Note

Implementation of publicly exposed wrapper API on top of basic cooperative group type APIs. This function is implemented on Linux and is under development on Microsoft Windows.

◆ invoke_one()

template<typename Group , typename Fn , typename... Args>

void cooperative_groups::invoke_one	(	const Group &	g,
		Fn &&	fn,
		Args &&...	args
	)

Calls an invocable on a single thread of the group.

Selects one thread of the group and uses it to call the supplied function with specified arguments. All group types are supported. The group is not synchronized, so communication or synchronization within the calling group is not allowed inside function.

Template Parameters

Group	The cooperative group class template parameter.
Fn	The type of the invocable.
Args	The types of the arguments of the invocable.

Parameters

g	[in] The group to select the thread from.
fn	[in] The invocable called by the selected thread.
args	[in] The arguments passed to the invocable.

◆ invoke_one_broadcast() [1/2]

template<typename Fn , typename... Args>

auto cooperative_groups::invoke_one_broadcast	(	const coalesced_group &	g,
		Fn &&	fn,
		Args &&...	args
	)		-> typename __hip_internal::remove_reference<decltype(impl::forward<Fn>(fn)( impl::forward<Args>(args)...))>::type

Calls an invocable on a single thread of the group and broadcasts the result.

Selects one thread of the group and uses it to call the supplied function with specified arguments and then, distributes the returned value to all threads of the group. The group is not synchronized, so communication or synchronization within the calling group is not allowed inside function. The return type of the function must be trivially copyable and at most 32 bytes in size.

Template Parameters

Fn	The type of the invocable.
Args	The types of the arguments of the invocable.

Parameters

g	[in] The group to select the thread from.
fn	[in] The invocable called by the selected thread.
args	[in] The arguments passed to the invocable.

◆ invoke_one_broadcast() [2/2]

template<unsigned int Size, typename Parent , typename Fn , typename... Args>

auto cooperative_groups::invoke_one_broadcast	(	const thread_block_tile< Size, Parent > &	g,
		Fn &&	fn,
		Args &&...	args
	)		-> typename __hip_internal::remove_reference<decltype(impl::forward<Fn>(fn)( impl::forward<Args>(args)...))>::type

Calls an invocable on a single thread of the group and broadcasts the result.

Selects one thread of the group and uses it to call the supplied function with specified arguments and then, distributes the returned value to all threads of the group. The group is not synchronized, so communication or synchronization within the calling group is not allowed inside function. The return type of the function must be trivially copyable and at most 32 bytes in size.

Template Parameters

Size	The size of the thread block tile group.
Parent	The cooperative group class template parameter of the input group.
Fn	The type of the invocable.
Args	The types of the arguments of the invocable.

Parameters

g	[in] The group to select the thread from.
fn	[in] The invocable called by the selected thread.
args	[in] The arguments passed to the invocable.

◆ is_valid()

template<class CGTy >

bool cooperative_groups::is_valid

(

CGTy const &

g

)

Returns true if the group has not violated any API constraints.

Template Parameters

CGTy	The cooperative group class template parameter.

Parameters

g	[in] The cooperative group for validity check.

Note

Implementation of publicly exposed wrapper API on top of basic cooperative group type APIs. This function is implemented on Linux and is under development on Microsoft Windows.

◆ sync()

template<class CGTy >

void cooperative_groups::sync

(

CGTy const &

g

)

Synchronizes the threads in the group.

Template Parameters

CGTy	The cooperative group class template parameter.

Parameters

g	[in] The cooperative group for synchronization.

Note

Implementation of publicly exposed wrapper API on top of basic cooperative group type APIs. This function is implemented on Linux and is under development on Microsoft Windows.

◆ thread_rank()

template<class CGTy >

__hip_uint32_t cooperative_groups::thread_rank

(

CGTy const &

g

)

Returns the rank of thread of the group.

Rank of the calling thread within [0, num_threads() ).

Template Parameters

CGTy	The cooperative group class template parameter.

Parameters

g	[in] The cooperative group for rank returns.

Note

Implementation of publicly exposed wrapper API on top of basic cooperative group type APIs. This function is implemented on Linux and is under development on Microsoft Windows.