functionality which is usable on the accelerator compute device from within a kernel. More...

Namespaces
namespace	atomic
namespace	concepts
namespace	cpu
namespace	layout
namespace	order
namespace	origin
	Origin of index domains.
namespace	range
namespace	scope
namespace	trait
namespace	traverse
namespace	unit
	Unit of index domains.
namespace	warp
namespace	worker

Classes
struct	Acc
class	AtomicAtomicRef
	The atomic ops based on atomic_ref for CPU accelerators. More...
struct	DomainSpec
class	FlatIdxContainer
struct	MultiDimensional
struct	SimdAlgo
	Creates a functor operate on contiguous data concurrently. More...
class	TiledIdxContainer
struct	WorkerGroup

Functions
template<typename T, typename T_Scope = scope::Device>
constexpr auto	atomicAdd (auto const &acc, T *const addr, T const &value, T_Scope const hier=T_Scope()) -> T
	Executes an atomic add operation.
template<typename T, typename T_Scope = scope::Device>
constexpr auto	atomicAnd (auto const &acc, T *const addr, T const &value, T_Scope const hier=T_Scope()) -> T
	Executes an atomic and operation.
template<typename T, typename T_Scope = scope::Device>
constexpr auto	atomicCas (auto const &acc, T *const addr, T const &compare, T const &value, T_Scope const hier=T_Scope()) -> T
	Executes an atomic compare-and-swap operation.
template<typename T, typename T_Scope = scope::Device>
constexpr auto	atomicDec (auto const &acc, T *const addr, T const &value, T_Scope const hier=T_Scope()) -> T
	Executes an atomic decrement operation.
template<typename T, typename T_Scope = scope::Device>
constexpr auto	atomicExch (auto const &acc, T *const addr, T const &value, T_Scope const hier=T_Scope()) -> T
	Executes an atomic exchange operation.
template<typename T, typename T_Scope = scope::Device>
constexpr auto	atomicInc (auto const &acc, T *const addr, T const &value, T_Scope const hier=T_Scope()) -> T
	Executes an atomic increment operation.
template<typename T, typename T_Scope = scope::Device>
constexpr auto	atomicMax (auto const &acc, T *const addr, T const &value, T_Scope const hier=T_Scope()) -> T
	Executes an atomic max operation.
template<typename T, typename T_Scope = scope::Device>
constexpr auto	atomicMin (auto const &acc, T *const addr, T const &value, T_Scope const hier=T_Scope()) -> T
	Executes an atomic min operation.
template<typename TOp, typename T, typename T_Scope = scope::Device>
constexpr auto	atomicOp (auto const &acc, T *const addr, T const &compare, T const &value, T_Scope const scope=T_Scope()) -> T
	Executes the given operation atomically.
template<typename TOp, typename T, typename T_Scope = scope::Device>
constexpr auto	atomicOp (auto const &acc, T *const addr, T const &value, T_Scope const scope=T_Scope()) -> T
	Executes the given operation atomically.
template<typename T, typename T_Scope = scope::Device>
constexpr auto	atomicOr (auto const &acc, T *const addr, T const &value, T_Scope const hier=T_Scope()) -> T
	Executes an atomic or operation.
template<typename T, typename T_Scope = scope::Device>
constexpr auto	atomicSub (auto const &acc, T *const addr, T const &value, T_Scope const hier=T_Scope()) -> T
	Executes an atomic sub operation.
template<typename T, typename T_Scope = scope::Device>
constexpr auto	atomicXor (auto const &acc, T *const addr, T const &value, T_Scope const hier=T_Scope()) -> T
	Executes an atomic xor operation.
template<typename T, size_t T_uniqueId>
constexpr decltype(auto)	declareSharedMdArray (concepts::Acc auto const &acc, alpaka::concepts::CVector auto const &extent)
	creates an M-dimensional array
template<typename T, size_t T_uniqueId>
constexpr decltype(auto)	declareSharedVar (concepts::Acc auto const &acc)
	Create a variable located in the thread blocks shared memory.
template<typename T>
constexpr auto	getDynSharedMem (concepts::Acc auto const &acc) -> T *
	Get block shared dynamic memory.
template<typename T>
void	isSupportedByAtomicAtomicRef ()
constexpr void	memFence (auto const &acc, concepts::Scope auto const scope, concepts::MemoryOrder auto const order)
template<alpaka::concepts::Layer T_Scope>
constexpr void	sync (concepts::Acc auto const &acc, T_Scope scope)
	Synchronize all threads within a given scope.
constexpr void	syncBlockThreads (concepts::Acc auto const &acc)
	Synchronize all threads within a thread block.
range-based loop indexable index container
template<concepts::IdxTraversing T_Traverse = traverse::Flat, concepts::IdxMapping T_IdxLayout = layout::Optimized>
ALPAKA_FN_HOST_ACC constexpr auto	makeIdxMap (auto const &acc, auto const workGroup, auto const range, T_Traverse traverse=T_Traverse{}, T_IdxLayout idxLayout=T_IdxLayout{})
	Creates an index container.
template<concepts::IdxTraversing T_Traverse = traverse::Flat, concepts::IdxMapping T_IdxLayout = layout::Optimized>
ALPAKA_FN_HOST_ACC constexpr auto	makeIdxMap (auto const &acc, auto const workGroup, auto const range, alpaka::concepts::BoundaryDirection auto const &bd, T_Traverse traverse=T_Traverse{}, T_IdxLayout idxLayout=T_IdxLayout{})
	Specialization for an index container with a given boundary direction of the volume described by the range.
template<typename T_ScalarIdxType, concepts::IdxTraversing T_Traverse = traverse::Flat, concepts::IdxMapping T_IdxLayout = layout::Optimized>
ALPAKA_FN_HOST_ACC constexpr auto	makeIdxMap (auto const &acc, auto const workGroup, auto const range, T_Traverse traverse=T_Traverse{}, T_IdxLayout idxLayout=T_IdxLayout{})
	Creates an index container.

Variables
template<typename T>
constexpr bool	isOrigin_v = trait::IsOrigin<T>::value
template<typename T>
constexpr bool	isUnit_v = trait::IsUnit<T>::value
constexpr auto	linearized = MultiDimensional<false>{}

Detailed Description

functionality which is usable on the accelerator compute device from within a kernel.

Provides scopes for atomic and memory fence operations, analogous to NVIDIA CUDA's atomic and fence scopes.

This namespace defines the visibility scopes for atomic operations and memory fences, which control the visibility of memory operations across threads, blocks, and devices. The provided scopes are:

Block: Visibility within a thread block.
Device: Visibility across all thread blocks on the same device.
System: System-wide visibility, mapped to the strongest available atomic/fence by the backend.

See also: alpaka::onAcc::atomicAdd, alpaka::onAcc::memFence

Function Documentation

◆ atomicAdd()

template<typename T, typename T_Scope = scope::Device>

auto alpaka::onAcc::atomicAdd	(	auto const &	acc,
		T *const	addr,
		T const &	value,
		T_Scope const	hier = T_Scope() ) -> T

constexpr

Executes an atomic add operation.

Template Parameters

T	The value type.

Parameters

addr	The value to change atomically.
value	The value used in the atomic operation.

Definition at line 66 of file atomic.hpp.

◆ atomicAnd()

template<typename T, typename T_Scope = scope::Device>

auto alpaka::onAcc::atomicAnd	(	auto const &	acc,
		T *const	addr,
		T const &	value,
		T_Scope const	hier = T_Scope() ) -> T

constexpr

Executes an atomic and operation.

Template Parameters

T	The value type.

Parameters

addr	The value to change atomically.
value	The value used in the atomic operation.

Definition at line 143 of file atomic.hpp.

◆ atomicCas()

template<typename T, typename T_Scope = scope::Device>

auto alpaka::onAcc::atomicCas	(	auto const &	acc,
		T *const	addr,
		T const &	compare,
		T const &	value,
		T_Scope const	hier = T_Scope() ) -> T

constexpr

Executes an atomic compare-and-swap operation.

Template Parameters

T	The value type.

Parameters

addr	The value to change atomically.
compare	The comparison value used in the atomic operation.
value	The value used in the atomic operation.

Definition at line 177 of file atomic.hpp.

◆ atomicDec()

template<typename T, typename T_Scope = scope::Device>

auto alpaka::onAcc::atomicDec	(	auto const &	acc,
		T *const	addr,
		T const &	value,
		T_Scope const	hier = T_Scope() ) -> T

constexpr

Executes an atomic decrement operation.

Template Parameters

T	The value type.

Parameters

addr	The value to change atomically.
value	The value used in the atomic operation.

Definition at line 132 of file atomic.hpp.

◆ atomicExch()

template<typename T, typename T_Scope = scope::Device>

auto alpaka::onAcc::atomicExch	(	auto const &	acc,
		T *const	addr,
		T const &	value,
		T_Scope const	hier = T_Scope() ) -> T

constexpr

Executes an atomic exchange operation.

Template Parameters

T	The value type.

Parameters

addr	The value to change atomically.
value	The value used in the atomic operation.

Definition at line 110 of file atomic.hpp.

◆ atomicInc()

template<typename T, typename T_Scope = scope::Device>

auto alpaka::onAcc::atomicInc	(	auto const &	acc,
		T *const	addr,
		T const &	value,
		T_Scope const	hier = T_Scope() ) -> T

constexpr

Executes an atomic increment operation.

Template Parameters

T	The value type.

Parameters

addr	The value to change atomically.
value	The value used in the atomic operation.

Definition at line 121 of file atomic.hpp.

◆ atomicMax()

template<typename T, typename T_Scope = scope::Device>

auto alpaka::onAcc::atomicMax	(	auto const &	acc,
		T *const	addr,
		T const &	value,
		T_Scope const	hier = T_Scope() ) -> T

constexpr

Executes an atomic max operation.

Template Parameters

T	The value type.

Parameters

addr	The value to change atomically.
value	The value used in the atomic operation.

Definition at line 99 of file atomic.hpp.

◆ atomicMin()

template<typename T, typename T_Scope = scope::Device>

auto alpaka::onAcc::atomicMin	(	auto const &	acc,
		T *const	addr,
		T const &	value,
		T_Scope const	hier = T_Scope() ) -> T

constexpr

Executes an atomic min operation.

Template Parameters

T	The value type.

Parameters

addr	The value to change atomically.
value	The value used in the atomic operation.

Definition at line 88 of file atomic.hpp.

◆ atomicOp() [1/2]

template<typename TOp, typename T, typename T_Scope = scope::Device>

auto alpaka::onAcc::atomicOp	(	auto const &	acc,
		T *const	addr,
		T const &	compare,
		T const &	value,
		T_Scope const	scope = T_Scope() ) -> T

constexpr

Executes the given operation atomically.

Template Parameters

TOp	The operation type.
T	The value type.

Parameters

addr	The value to change atomically.
compare	The comparison value used in the atomic operation.
value	The value used in the atomic operation.

Definition at line 44 of file atomic.hpp.

◆ atomicOp() [2/2]

template<typename TOp, typename T, typename T_Scope = scope::Device>

auto alpaka::onAcc::atomicOp	(	auto const &	acc,
		T *const	addr,
		T const &	value,
		T_Scope const	scope = T_Scope() ) -> T

constexpr

Executes the given operation atomically.

Template Parameters

TOp	The operation type.
T	The value type.

Parameters

addr	The value to change atomically.
value	The value used in the atomic operation.

Definition at line 26 of file atomic.hpp.

◆ atomicOr()

template<typename T, typename T_Scope = scope::Device>

auto alpaka::onAcc::atomicOr	(	auto const &	acc,
		T *const	addr,
		T const &	value,
		T_Scope const	hier = T_Scope() ) -> T

constexpr

Executes an atomic or operation.

Template Parameters

T	The value type.

Parameters

addr	The value to change atomically.
value	The value used in the atomic operation.

Definition at line 154 of file atomic.hpp.

◆ atomicSub()

template<typename T, typename T_Scope = scope::Device>

auto alpaka::onAcc::atomicSub	(	auto const &	acc,
		T *const	addr,
		T const &	value,
		T_Scope const	hier = T_Scope() ) -> T

constexpr

Executes an atomic sub operation.

Template Parameters

T	The value type.

Parameters

addr	The value to change atomically.
value	The value used in the atomic operation.

Definition at line 77 of file atomic.hpp.

◆ atomicXor()

template<typename T, typename T_Scope = scope::Device>

auto alpaka::onAcc::atomicXor	(	auto const &	acc,
		T *const	addr,
		T const &	value,
		T_Scope const	hier = T_Scope() ) -> T

constexpr

Executes an atomic xor operation.

Template Parameters

T	The value type.

Parameters

addr	The value to change atomically.
value	The value used in the atomic operation.

Definition at line 165 of file atomic.hpp.

◆ declareSharedMdArray()

template<typename T, size_t T_uniqueId>

decltype(auto) alpaka::onAcc::declareSharedMdArray	(	concepts::Acc auto const &	acc,
		alpaka::concepts::CVector auto const &	extent )

constexpr

creates an M-dimensional array

// creates a MdSpan view to a float value, do NOT use a reference here

auto fooArrayMd = declareSharedVar<float, uniqueId()>(acc, CVec<uint32_t, 5, 8>{});

alpaka::onAcc::declareSharedVar

constexpr decltype(auto) declareSharedVar(concepts::Acc auto const &acc)

Create a variable located in the thread blocks shared memory.

Definition Acc.hpp:152

alpaka::CVec

Vec< T, sizeof...(T_values), detail::CVec< T, T_values... > > CVec

A vector with compile-time known values.

Definition CVec.hpp:31

Attention: The data is not initialized it can contains garbage.

Template Parameters

T	type which should be created, the constructor is not called
T_uniqueId	id those is unique inside a kernel. Reusing the id will return the same memory declared before with the same id.

Parameters

extent M-dimensional extent in elements for each dimension, 1 - M dimensions are supported

Returns: MdSpan non owning view to the corresponding data, you should NOT store a reference to the handle

Definition at line 173 of file Acc.hpp.

◆ declareSharedVar()

template<typename T, size_t T_uniqueId>

decltype(auto) alpaka::onAcc::declareSharedVar ( concepts::Acc auto const & acc )

constexpr

Create a variable located in the thread blocks shared memory.

// creates a reference to a float value

auto& foo = declareSharedVar<float, uniqueId()>(acc);

Attention: The data is not initialized; it can contain garbage.

Template Parameters

T	The type that should be created; the constructor is not called
T_uniqueId	ID that is unique inside a kernel. Reusing the id will return the same memory declared before with the same id.

Returns: Result should be taken by reference

Definition at line 152 of file Acc.hpp.

◆ getDynSharedMem()

template<typename T>

auto alpaka::onAcc::getDynSharedMem ( concepts::Acc auto const & acc ) -> T *

constexpr

Get block shared dynamic memory.

The available size of the memory can be defined by specializing 'onHost::trait::GetDynSharedMemBytes' or adding a public member variable 'uint32_t dynSharedMemBytes' for a kernel. The Memory can be accessed by all threads within a block. Access to the memory is not thread safe.

Template Parameters

T	The element type.

Returns: Pointer to pre-allocated contiguous memory.

Definition at line 197 of file Acc.hpp.

◆ isSupportedByAtomicAtomicRef()

template<typename T>

void alpaka::onAcc::isSupportedByAtomicAtomicRef ( )

Definition at line 47 of file atomic.hpp.

◆ makeIdxMap() [1/3]

template<concepts::IdxTraversing T_Traverse = traverse::Flat, concepts::IdxMapping T_IdxLayout = layout::Optimized>

ALPAKA_FN_HOST_ACC constexpr auto alpaka::onAcc::makeIdxMap	(	auto const &	acc,
		auto const	workGroup,
		auto const	range,
		alpaka::concepts::BoundaryDirection auto const &	bd,
		T_Traverse	traverse = T_Traverse{},
		T_IdxLayout	idxLayout = T_IdxLayout{} )

constexpr

Specialization for an index container with a given boundary direction of the volume described by the range.

Definition at line 75 of file interface.hpp.

◆ makeIdxMap() [2/3]

template<typename T_ScalarIdxType, concepts::IdxTraversing T_Traverse = traverse::Flat, concepts::IdxMapping T_IdxLayout = layout::Optimized>

ALPAKA_FN_HOST_ACC constexpr auto alpaka::onAcc::makeIdxMap	(	auto const &	acc,
		auto const	workGroup,
		auto const	range,
		T_Traverse	traverse = T_Traverse{},
		T_IdxLayout	idxLayout = T_IdxLayout{} )

constexpr

Creates an index container.

The traversal policy (T_Traverse) defines how the next valid index is found for a worker and defaults to traverse::Flat. The mapping policy (T_IdxLayout) defines how the index is mapped to worker threads and defaults to layout::Optimized.

Template Parameters

T_ScalarIdxType scalar index type sed for the indices inside the iterator

Parameters

workGroup	Description of the participating thread group. More than one thread can have the same index within the group; all workers with the same id obtain the same index as result.
range	Index range description.
traverse	Policy describing how the next value can be found.
idxLayout	Policy describing how real worker threads will be mapped to the range.

Returns: A index container that can be used in a range-based for loop.

Definition at line 113 of file interface.hpp.

◆ makeIdxMap() [3/3]

template<concepts::IdxTraversing T_Traverse = traverse::Flat, concepts::IdxMapping T_IdxLayout = layout::Optimized>

ALPAKA_FN_HOST_ACC constexpr auto alpaka::onAcc::makeIdxMap	(	auto const &	acc,
		auto const	workGroup,
		auto const	range,
		T_Traverse	traverse = T_Traverse{},
		T_IdxLayout	idxLayout = T_IdxLayout{} )

constexpr

Creates an index container.

The index data type is deduced from the supplied range. The traversal policy (T_Traverse) defines how the next valid index is found for a worker and defaults to traverse::Flat. The mapping policy (T_IdxLayout) defines how the index is mapped to worker threads and defaults to layout::Optimized.

Parameters

workGroup	Description of the participating thread group. More than one thread can have the same index within the group; all workers with the same id obtain the same index as result.
range	Index range description.
traverse	Policy describing how the next value can be found.
idxLayout	Policy describing how real worker threads will be mapped to the range.

Returns: A index container that can be used in a range-based for loop.

Definition at line 57 of file interface.hpp.

◆ memFence()

void alpaka::onAcc::memFence	(	auto const &	acc,
		concepts::Scope auto const	scope,
		concepts::MemoryOrder auto const	order )

constexpr

Definition at line 18 of file memFence.hpp.

◆ sync()

template<alpaka::concepts::Layer T_Scope>

void alpaka::onAcc::sync	(	concepts::Acc auto const &	acc,
		T_Scope	scope )

constexpr

Synchronize all threads within a given scope.

Definition at line 126 of file Acc.hpp.

◆ syncBlockThreads()

void alpaka::onAcc::syncBlockThreads ( concepts::Acc auto const & acc )

constexpr

Synchronize all threads within a thread block.

Definition at line 132 of file Acc.hpp.

Variable Documentation

◆ isOrigin_v

template<typename T>

bool alpaka::onAcc::isOrigin_v = trait::IsOrigin<T>::value

constexpr

Definition at line 88 of file tag.hpp.

◆ isUnit_v

template<typename T>

bool alpaka::onAcc::isUnit_v = trait::IsUnit<T>::value

constexpr

Definition at line 91 of file tag.hpp.

◆ linearized

auto alpaka::onAcc::linearized = MultiDimensional<false>{}

constexpr

Definition at line 21 of file WorkerGroup.hpp.

Namespaces

Classes

Functions

Variables

Detailed Description

Function Documentation

◆ atomicAdd()

◆ atomicAnd()

◆ atomicCas()

◆ atomicDec()

◆ atomicExch()

◆ atomicInc()

◆ atomicMax()

◆ atomicMin()

◆ atomicOp() [1/2]

◆ atomicOp() [2/2]

◆ atomicOr()

◆ atomicSub()

◆ atomicXor()

◆ declareSharedMdArray()

◆ declareSharedVar()

◆ getDynSharedMem()

◆ isSupportedByAtomicAtomicRef()

◆ makeIdxMap() [1/3]

◆ makeIdxMap() [2/3]

◆ makeIdxMap() [3/3]

◆ memFence()

◆ sync()

◆ syncBlockThreads()

Variable Documentation

◆ isOrigin_v

◆ isUnit_v

◆ linearized