alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout > Struct Template Reference

Creates a functor operate on contiguous data concurrently. More...

#include <SimdAlgo.hpp>

Inheritance diagram for alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >:

Classes
struct	SimdPackConfig

Public Member Functions
constexpr	SimdAlgo (T_WorkGroup const workGroup, T_Traverse traverse=T_Traverse{}, T_IdxLayout idxLayout=T_IdxLayout{})
constexpr T_IdxLayout	getIdxLayoutPolicy () const
constexpr T_Traverse	getTraversePolicy () const
constexpr T_WorkGroup	getWorkGroup () const
template<uint32_t T_maxConcurrencyInByte, alpaka::concepts::Alignment T_MemAlignment = AutoAligned>
ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr auto	transformReduce (auto const &acc, alpaka::concepts::Vector auto extents, auto const &neutralElement, auto &&reduceFunc, auto &&transformFunc, alpaka::concepts::IDataSource auto &&data0, alpaka::concepts::IDataSource auto &&... dataN) const
	transform the input data and reduce it to a single value
ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr auto	transformReduce (auto const &acc, alpaka::concepts::Vector auto extents, auto const &neutralElement, auto &&reduceFunc, auto &&transformFunc, alpaka::concepts::IDataSource auto &&data0, alpaka::concepts::IDataSource auto &&... dataN) const
	transform the input data and reduce it to a single value
template<uint32_t T_maxConcurrencyInByte, alpaka::concepts::Alignment T_MemAlignment = AutoAligned>
ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr auto	transformReduce (auto const &acc, auto const &neutralElement, auto &&reduceFunc, auto &&transformFunc, alpaka::concepts::IDataSource auto &&data0, alpaka::concepts::IDataSource auto &&... dataN) const
	transform the input data and reduce it to a single value
ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr auto	transformReduce (auto const &acc, auto const &neutralElement, auto &&reduceFunc, auto &&transformFunc, alpaka::concepts::IDataSource auto &&data0, alpaka::concepts::IDataSource auto &&... dataN) const
	transform the input data and reduce it to a single value
ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr void	concurrent (auto const &acc, auto &&func, alpaka::concepts::IDataSource auto &&data0, alpaka::concepts::IDataSource auto &&... dataN) const
	execute the functor concurrently over the given data.
ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr void	concurrent (auto const &acc, alpaka::concepts::Vector auto extents, auto &&func, alpaka::concepts::IDataSource auto &&data0, alpaka::concepts::IDataSource auto &&... dataN) const
template<uint32_t T_maxConcurrencyInByte, alpaka::concepts::Alignment T_MemAlignment = AutoAligned>
ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr void	concurrent (auto const &acc, auto &&func, alpaka::concepts::IDataSource auto &&data0, alpaka::concepts::IDataSource auto &&... dataN) const
	execute the functor concurrently over the given data.
template<uint32_t T_maxConcurrencyInByte, alpaka::concepts::Alignment T_MemAlignment = AutoAligned>
ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr void	concurrent (auto const &acc, alpaka::concepts::Vector auto extents, auto &&func, alpaka::concepts::IDataSource auto &&data0, alpaka::concepts::IDataSource auto &&... dataN) const

Private Types
using	ConcurrentAlgo = internal::SimdConcurrent<SimdAlgo<T_WorkGroup, T_Traverse, T_IdxLayout>>
using	ReduceAlgo = internal::SimdTransformReduce<SimdAlgo<T_WorkGroup, T_Traverse, T_IdxLayout>>

Static Private Member Functions
template<typename T_ValueType>
static consteval SimdPackConfig< T_ValueType >	calcSimdPackConfig (alpaka::concepts::Api auto api, alpaka::concepts::DeviceKind auto deviceKind, uint32_t maxConcurrencyInByte)
	Generate a SIMD config for the API and device kind.
template<typename T_Type, uint32_t T_maxConcurrencyInByte, uint32_t T_cacheLineInByte>
static constexpr auto	calcSimdWidth ()

Private Attributes
friend	ConcurrentAlgo
T_WorkGroup	m_workGroup
friend	ReduceAlgo

Additional Inherited Members
Protected Member Functions inherited from alpaka::onAcc::internal::SimdConcurrent< SimdAlgo< T_WorkGroup, traverse::Flat, layout::Optimized > >
ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr void	concurrent (auto const &acc, alpaka::concepts::Vector auto extents, auto &&func, alpaka::concepts::IDataSource auto &&data0, alpaka::concepts::IDataSource auto &&... dataN) const
constexpr	SimdConcurrent ()=default
Protected Member Functions inherited from alpaka::onAcc::internal::SimdTransformReduce< SimdAlgo< T_WorkGroup, traverse::Flat, layout::Optimized > >
ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr auto	transformReduce (concepts::Acc auto const &acc, alpaka::concepts::Vector auto extents, auto const &neutralElement, auto &&reduceFunc, auto &&func, alpaka::concepts::IDataSource auto &&data0, alpaka::concepts::IDataSource auto &&... dataN) const
constexpr	SimdTransformReduce ()=default

Detailed Description

template<typename T_WorkGroup, concepts::IdxTraversing T_Traverse = traverse::Flat, concepts::IdxMapping T_IdxLayout = layout::Optimized>
struct alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >

Creates a functor operate on contiguous data concurrently.

The class is automatically configured to use the best fitting SIMD width for the given data type and is able to expose instruction level parallelism.

Parameters

T_WorkGroup	participating thread description. More than one thread can have the same index within the group. All worker with the same id will get the same index as result.
T_Traverse	Policy to configure the method used to find the next valid index for a worker.

See also

namespace traverse

Parameters

T_IdxLayout

Policy to define how index data will be mapped to worker threads.

See also

namespace layout

Definition at line 33 of file SimdAlgo.hpp.

Member Typedef Documentation

ConcurrentAlgo

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

using alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::ConcurrentAlgo = internal::SimdConcurrent<SimdAlgo<T_WorkGroup, T_Traverse, T_IdxLayout>>

private

Definition at line 292 of file SimdAlgo.hpp.

ReduceAlgo

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

using alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::ReduceAlgo = internal::SimdTransformReduce<SimdAlgo<T_WorkGroup, T_Traverse, T_IdxLayout>>

private

Definition at line 293 of file SimdAlgo.hpp.

Constructor & Destructor Documentation

SimdAlgo()

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

alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::SimdAlgo ( T_WorkGroup const workGroup, T_Traverse traverse = T_Traverse{}, T_IdxLayout idxLayout = T_IdxLayout{} )

inlineconstexpr

Definition at line 37 of file SimdAlgo.hpp.

Member Function Documentation

calcSimdPackConfig()

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

template<typename T_ValueType>

consteval SimdPackConfig< T_ValueType > alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::calcSimdPackConfig ( alpaka::concepts::Api auto api, alpaka::concepts::DeviceKind auto deviceKind, uint32_t maxConcurrencyInByte )

inlinestaticnodiscardconstevalprivate

Generate a SIMD config for the API and device kind.

Produces an optimized SIMD configuration based on technical constraints. The SIMD is set to a power of two. If possible, the SIMD configuration is aligned to the cacheline size for the given device kind.

Parameters

maxConcurrencyInByte

The upper limit in bytes a SIMD configuration must not exceed, except a single value is larger. This parameter is used to control the register pressure.

Returns

a configuration with the number of SIMD pack which should be used in parallel for a single invocation. And the width of a single SIMD pack.

Definition at line 326 of file SimdAlgo.hpp.

calcSimdWidth()

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

template<typename T_Type, uint32_t T_maxConcurrencyInByte, uint32_t T_cacheLineInByte>

constexpr auto alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::calcSimdWidth ( )

inlinestaticconstexprprivate

Definition at line 299 of file SimdAlgo.hpp.

concurrent() [1/4]

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

template<uint32_t T_maxConcurrencyInByte, alpaka::concepts::Alignment T_MemAlignment = AutoAligned>

ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr void alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::concurrent ( auto const & acc, alpaka::concepts::Vector auto extents, auto && func, alpaka::concepts::IDataSource auto && data0, alpaka::concepts::IDataSource auto &&... dataN ) const

inlineconstexpr

Parameters

extents

number of elements to process in each dimension

Definition at line 144 of file SimdAlgo.hpp.

concurrent() [2/4]

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

ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr void alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::concurrent ( auto const & acc, alpaka::concepts::Vector auto extents, auto && func, alpaka::concepts::IDataSource auto && data0, alpaka::concepts::IDataSource auto &&... dataN ) const

inlineconstexpr

Parameters

extents

number of elements to process in each dimension

Definition at line 85 of file SimdAlgo.hpp.

concurrent() [3/4]

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

template<uint32_t T_maxConcurrencyInByte, alpaka::concepts::Alignment T_MemAlignment = AutoAligned>

ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr void alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::concurrent ( auto const & acc, auto && func, alpaka::concepts::IDataSource auto && data0, alpaka::concepts::IDataSource auto &&... dataN ) const

inlineconstexpr

execute the functor concurrently over the given data.

Attention

The number of elements to process is derived from the first MdSpan object. All other MdSpan objects must have at least the same number of elements.

Parameters

T_maxConcurrencyInByte	Maximum number of bytes to be used for concurrency. Concurrency bytes describe a virtual simd pack size which is not exceeded. Internally a best fitting SIMD width is calculated and instruction parallelism is exposed based on T_maxConcurrencyInByte.
T_MemAlignment	alignment of the memory, if no alignments is given the alignment will be derived from the MdSpan data descriptions
func	the functor to be executed
data0	the first data to be processed
dataN	the remaining data to be processed

Definition at line 126 of file SimdAlgo.hpp.

concurrent() [4/4]

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

ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr void alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::concurrent ( auto const & acc, auto && func, alpaka::concepts::IDataSource auto && data0, alpaka::concepts::IDataSource auto &&... dataN ) const

inlineconstexpr

execute the functor concurrently over the given data.

Attention

The number of elements to process is derived from the first MdSpan object. All other MdSpan objects must have at least the same number of elements. The optimal concurrency is also derived from the first MdSpan.

Parameters

func	the functor to be executed
data0	the first data to be processed
dataN	the remaining data to be processed

Definition at line 73 of file SimdAlgo.hpp.

getIdxLayoutPolicy()

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

T_IdxLayout alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::getIdxLayoutPolicy ( ) const

inlineconstexpr

Definition at line 56 of file SimdAlgo.hpp.

getTraversePolicy()

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

T_Traverse alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::getTraversePolicy ( ) const

inlineconstexpr

Definition at line 51 of file SimdAlgo.hpp.

getWorkGroup()

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

T_WorkGroup alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::getWorkGroup ( ) const

inlineconstexpr

Definition at line 46 of file SimdAlgo.hpp.

transformReduce() [1/4]

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

template<uint32_t T_maxConcurrencyInByte, alpaka::concepts::Alignment T_MemAlignment = AutoAligned>

ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr auto alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::transformReduce ( auto const & acc, alpaka::concepts::Vector auto extents, auto const & neutralElement, auto && reduceFunc, auto && transformFunc, alpaka::concepts::IDataSource auto && data0, alpaka::concepts::IDataSource auto &&... dataN ) const

inlineconstexpr

transform the input data and reduce it to a single value

Attention

If no extent is given the number of elements to process is derived from the first MdSpan object. All other MdSpan objects must have at least the same number of elements.

Parameters

neutralElement	the neutral element for the reduction operation
reduceFunc	The binary reduction operation to be executed, e.g. std::plus. The functor should support Simd packages.
transformFunc	N-nary functor to be executed, values of all containers will be passed to the functor as arguments. The functor should support Simd packages. If not you can enforce the element wise execution by wrapping into ScalarFunc. If you would like to support stencil executions wrapp fn into StencilFunc. StencilFunc is getting all arguments as SimdPtr. If StencilFunc is used you should take care to not read outside of valid memory ranges by using sub-views to your input and output data. Optionally a transformFn can have an accelerator as first argument. If the result of this functor is a structured value providing an overload to simdize the type can improve the performance see alpaka::makeSimdized.
data0	the first data to be processed
dataN	the remaining data to be processed

Returns

A single reduced value.

Parameters

extents

number of elements to process in each dimension

Template Parameters

T_maxConcurrencyInByte	Maximum number of bytes to be used for concurrency. Concurrency bytes describe a virtual simd pack size which is not exceeded. Internally a best fitting SIMD width is calculated and instruction parallelism is exposed based on T_maxConcurrencyInByte.
T_MemAlignment	alignment of the memory, if no alignments is given the alignment will be derived from the MdSpan data descriptions

Definition at line 272 of file SimdAlgo.hpp.

transformReduce() [2/4]

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

ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr auto alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::transformReduce ( auto const & acc, alpaka::concepts::Vector auto extents, auto const & neutralElement, auto && reduceFunc, auto && transformFunc, alpaka::concepts::IDataSource auto && data0, alpaka::concepts::IDataSource auto &&... dataN ) const

inlineconstexpr

transform the input data and reduce it to a single value

Attention

If no extent is given the number of elements to process is derived from the first MdSpan object. All other MdSpan objects must have at least the same number of elements.

Parameters

neutralElement	the neutral element for the reduction operation
reduceFunc	The binary reduction operation to be executed, e.g. std::plus. The functor should support Simd packages.
transformFunc	N-nary functor to be executed, values of all containers will be passed to the functor as arguments. The functor should support Simd packages. If not you can enforce the element wise execution by wrapping into ScalarFunc. If you would like to support stencil executions wrapp fn into StencilFunc. StencilFunc is getting all arguments as SimdPtr. If StencilFunc is used you should take care to not read outside of valid memory ranges by using sub-views to your input and output data. Optionally a transformFn can have an accelerator as first argument. If the result of this functor is a structured value providing an overload to simdize the type can improve the performance see alpaka::makeSimdized.
data0	the first data to be processed
dataN	the remaining data to be processed

Returns

A single reduced value.

Parameters

extents

number of elements to process in each dimension

Definition at line 205 of file SimdAlgo.hpp.

transformReduce() [3/4]

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

template<uint32_t T_maxConcurrencyInByte, alpaka::concepts::Alignment T_MemAlignment = AutoAligned>

ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr auto alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::transformReduce ( auto const & acc, auto const & neutralElement, auto && reduceFunc, auto && transformFunc, alpaka::concepts::IDataSource auto && data0, alpaka::concepts::IDataSource auto &&... dataN ) const

inlineconstexpr

transform the input data and reduce it to a single value

Attention

If no extent is given the number of elements to process is derived from the first MdSpan object. All other MdSpan objects must have at least the same number of elements.

Parameters

neutralElement	the neutral element for the reduction operation
reduceFunc	The binary reduction operation to be executed, e.g. std::plus. The functor should support Simd packages.
transformFunc	N-nary functor to be executed, values of all containers will be passed to the functor as arguments. The functor should support Simd packages. If not you can enforce the element wise execution by wrapping into ScalarFunc. If you would like to support stencil executions wrapp fn into StencilFunc. StencilFunc is getting all arguments as SimdPtr. If StencilFunc is used you should take care to not read outside of valid memory ranges by using sub-views to your input and output data. Optionally a transformFn can have an accelerator as first argument. If the result of this functor is a structured value providing an overload to simdize the type can improve the performance see alpaka::makeSimdized.
data0	the first data to be processed
dataN	the remaining data to be processed

Returns

A single reduced value.

Template Parameters

T_maxConcurrencyInByte	Maximum number of bytes to be used for concurrency. Concurrency bytes describe a virtual simd pack size which is not exceeded. Internally a best fitting SIMD width is calculated and instruction parallelism is exposed based on T_maxConcurrencyInByte.
T_MemAlignment	alignment of the memory, if no alignments is given the alignment will be derived from the MdSpan data descriptions

Definition at line 241 of file SimdAlgo.hpp.

transformReduce() [4/4]

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

ALPAKA_FN_INLINE ALPAKA_FN_ACC constexpr auto alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::transformReduce ( auto const & acc, auto const & neutralElement, auto && reduceFunc, auto && transformFunc, alpaka::concepts::IDataSource auto && data0, alpaka::concepts::IDataSource auto &&... dataN ) const

inlineconstexpr

transform the input data and reduce it to a single value

Attention

If no extent is given the number of elements to process is derived from the first MdSpan object. All other MdSpan objects must have at least the same number of elements.

Parameters

neutralElement	the neutral element for the reduction operation
reduceFunc	The binary reduction operation to be executed, e.g. std::plus. The functor should support Simd packages.
transformFunc	N-nary functor to be executed, values of all containers will be passed to the functor as arguments. The functor should support Simd packages. If not you can enforce the element wise execution by wrapping into ScalarFunc. If you would like to support stencil executions wrapp fn into StencilFunc. StencilFunc is getting all arguments as SimdPtr. If StencilFunc is used you should take care to not read outside of valid memory ranges by using sub-views to your input and output data. Optionally a transformFn can have an accelerator as first argument. If the result of this functor is a structured value providing an overload to simdize the type can improve the performance see alpaka::makeSimdized.
data0	the first data to be processed
dataN	the remaining data to be processed

Returns

A single reduced value.

Definition at line 183 of file SimdAlgo.hpp.

Member Data Documentation

ConcurrentAlgo

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

friend alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::ConcurrentAlgo

private

Definition at line 295 of file SimdAlgo.hpp.

m_workGroup

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

T_WorkGroup alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::m_workGroup

private

Definition at line 364 of file SimdAlgo.hpp.

ReduceAlgo

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

friend alpaka::onAcc::SimdAlgo< T_WorkGroup, T_Traverse, T_IdxLayout >::ReduceAlgo

private

Definition at line 296 of file SimdAlgo.hpp.

---

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

• /home/docs/checkouts/readthedocs.org/user_builds/alpaka3/checkouts/latest/include/alpaka/onAcc/SimdAlgo.hpp