Classes
struct	BlockProperties
	BlockProperties is a template class that provides different characteristic of a block of each Tensor processed by each workgroup. More...

class	FullReductionKernelFunctor

struct	GeneralScalarContraction
	GeneralScalarContraction is a template class that provides the scalar value of Tensor -Tensor contraction operation, when all the dimensions are contracting dimensions. This Kernel reduces two tensors to an scalar. More...

struct	GeneralVectorTensor
	GeneralVectorTensor is a template class that provides Tensor -vector contraction operation, which is a special case of Tensor Tensor contraction. More...

class	GenericNondeterministicReducer

struct	OpDefiner

struct	OpDefiner< Eigen::internal::MeanReducer< CoeffReturnType >, CoeffReturnType, Index, false >

struct	OpDefiner< Eigen::internal::MeanReducer< CoeffReturnType >, CoeffReturnType, Index, true >

struct	PacketWrapper

struct	PartialReducerLauncher

struct	PartialReductionKernel

struct	ReductionPannel

struct	ScanAdjustmentKernelFunctor

struct	ScanInfo

struct	ScanKernelFunctor

struct	ScanLauncher_impl

struct	ScanParameters

struct	SecondStepFullReducer

struct	SecondStepPartialReduction

struct	SYCLAdjustBlockOffset

class	TensorContractionKernel
	TensorContractionKernel is a template class that provides Tensor -Tensor contraction operation. More...

struct	ThreadProperties
	ThreadProperties is a template class that provides each thread's properties within a workgroup. Please see the sycl-1.2.1 specification (https://www.khronos.org/registry/SYCL/specs/sycl-1.2.1.pdf) for the workgroup, work-items. More...

struct	TTPanelSize
	TTPanelSize, a template class used for setting the panel size required for launching General Tensor Tensor contraction kernel on various hardware devices. More...

struct	TVPanelSize
	TVPanelSize, a template class used for setting the panel size required for launching General TensorVector contraction kernel on various hardware devices. More...

Enumerations
enum	contraction_type { contraction_type::local, contraction_type::no_local }

enum	data_source { data_source::global_mem, data_source::local_mem, data_source::private_mem }

enum	reduction_dim { reduction_dim::inner_most, reduction_dim::outer_most }

enum	scan_step { scan_step::first, scan_step::second }

Functions
template<bool is_internal>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool	check_boundary (bool)
	check_boundary: is used to check the edge condition for non-internal blocks. More...

template<>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool	check_boundary< false > (bool cond)
	check_boundary: specialization of the check_boundary for non-internal blocks. More...

template<bool PacketLoad, bool is_coalesced_layout, bool , typename PacketType , typename TensorMapper , typename StorageIndex >
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ::Eigen::internal::enable_if< PacketLoad, PacketType >::type	read (const TensorMapper &tensorMapper, const StorageIndex &NCIndex, const StorageIndex &CIndex, const StorageIndex &ld)
	read, a template function used for loading the data from global memory. This function is used to guarantee coalesced and vectorized load whenever possible More...

template<bool PacketLoad, bool , bool IsRhs, typename PacketType , typename TensorMapper , typename StorageIndex >
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ::Eigen::internal::enable_if<!PacketLoad, PacketType >::type	read (const TensorMapper &tensorMapper, const StorageIndex &NCIndex, const StorageIndex &CIndex, const StorageIndex &)
	read, special overload of read function, when the read access is not vectorized More...

template<typename StorageIndex , StorageIndex ld, data_source dt, typename PacketType , typename DataScalar >
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename ::Eigen::internal::enable_if< dt !=data_source::global_mem, void >::type	write (PacketType &packet_data, DataScalar ptr)
	write, a template function used for storing the data to local memory. This function is used to guarantee coalesced and vectorized store whenever possible. More...

template<data_source dt, typename PacketType , typename DataScalar >
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ::Eigen::internal::enable_if< Eigen::internal::unpacket_traits< PacketType >::size !=1 &&dt==data_source::global_mem, void >::type	write (PacketType &packet_data, DataScalar *ptr)
	Overloading the write function for storing the data to global memory, when vectorization enabled This function is used to guarantee coalesced and vectorized store whenever possible. More...

template<data_source dt, typename PacketType , typename DataScalar >
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ::Eigen::internal::enable_if< Eigen::internal::unpacket_traits< PacketType >::size==1 &&dt==data_source::global_mem, void >::type	write (PacketType &packet_data, DataScalar *ptr)
	Overloading the write function for storing the data to global memory, when vectorization is disabled. More...

Enumeration Type Documentation

◆ contraction_type

enum Eigen::TensorSycl::internal::contraction_type

strong

Enumerator
local
no_local

Definition at line 126 of file TensorContractionSycl.h.

◆ data_source

enum Eigen::TensorSycl::internal::data_source

strong

Enumerator
global_mem
local_mem
private_mem

Definition at line 130 of file TensorContractionSycl.h.

◆ reduction_dim

enum Eigen::TensorSycl::internal::reduction_dim

strong

Enumerator
inner_most
outer_most

Definition at line 250 of file TensorReductionSycl.h.

◆ scan_step

enum Eigen::TensorSycl::internal::scan_step

strong

Enumerator
first
second

Definition at line 80 of file TensorScanSycl.h.

Function Documentation

◆ check_boundary()

template<bool is_internal>

EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool Eigen::TensorSycl::internal::check_boundary ( bool )

check_boundary: is used to check the edge condition for non-internal blocks.

Template Parameters

is_internal determines if the block is internal

Definition at line 277 of file TensorContractionSycl.h.

◆ check_boundary< false >()

template<>

EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool Eigen::TensorSycl::internal::check_boundary< false > ( bool cond )

check_boundary: specialization of the check_boundary for non-internal blocks.

Parameters

cond	true when the data is in range. Otherwise false

Definition at line 287 of file TensorContractionSycl.h.

◆ read() [1/2]

template<bool PacketLoad, bool is_coalesced_layout, bool , typename PacketType , typename TensorMapper , typename StorageIndex >

static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ::Eigen::internal::enable_if<PacketLoad, PacketType>::type Eigen::TensorSycl::internal::read	(	const TensorMapper &	tensorMapper,
		const StorageIndex &	NCIndex,
		const StorageIndex &	CIndex,
		const StorageIndex &	ld
	)

static

read, a template function used for loading the data from global memory. This function is used to guarantee coalesced and vectorized load whenever possible

Template Parameters

PacketLoad determines if the each element of this tensor block should be loaded in a packet mode

Parameters

is_coalesced_layout determines whether or not the Tensor data in a memory can be access coalesced and vectorized when possible. Coalesced memory access is a key factor in Kernel performance. When a tensor is 2d and the contracting dimension is 1, it is always possible to accessed tensor data coalesced and vectorized. This is the case when RHS(right hand side) Tensor is transposed or when LHS(left hand side) Tensor is not transposed.

Template Parameters

PacketType	determines the type of packet
TensorMapper	determines the input tensor mapper type
StorageIndex	determines the Index type

Parameters

tensorMapper	is the input tensor
NCIndex	is the non-contracting dim index
CIndex	is the contracting dim index
ld	is the leading dimension of the flattened tensor

Definition at line 159 of file TensorContractionSycl.h.

◆ read() [2/2]

template<bool PacketLoad, bool , bool IsRhs, typename PacketType , typename TensorMapper , typename StorageIndex >

static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ::Eigen::internal::enable_if<!PacketLoad, PacketType>::type Eigen::TensorSycl::internal::read	(	const TensorMapper &	tensorMapper,
		const StorageIndex &	NCIndex,
		const StorageIndex &	CIndex,
		const StorageIndex &
	)

static

read, special overload of read function, when the read access is not vectorized

Template Parameters

PacketLoad determines if the each element of this tensor block should be loaded in a packet mode

Parameters

is_coalesced_layout determines whether or not the Tensor data in a memory can be access coalesced and vectorized when possible. Coalesced memory access is a key factor in Kernel performance. When a tensor is 2d and the contracting dimension is 1, it is always possible to accessed tensor data coalesced and vectorized. This is the case when RHS(right hand side) Tensor is transposed or when LHS(left hand side) Tensor is not transposed.

Template Parameters

PacketType	determines the type of packet
TensorMapper	determines the input tensor mapper type
StorageIndex	determines the Index type

Parameters

tensorMapper	is the input tensor
NCIndex	is the non-contracting dim index
CIndex	is the contracting dim index

Definition at line 189 of file TensorContractionSycl.h.

◆ write() [1/3]

template<typename StorageIndex , StorageIndex ld, data_source dt, typename PacketType , typename DataScalar >

static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename ::Eigen::internal::enable_if<dt != data_source::global_mem, void>::type Eigen::TensorSycl::internal::write	(	PacketType &	packet_data,
		DataScalar	ptr
	)

static

write, a template function used for storing the data to local memory. This function is used to guarantee coalesced and vectorized store whenever possible.

Template Parameters

StorageIndex determines the Index type

Parameters

ld	is the leading dimension of the local memory. ld is a compile time value for the local memory

Template Parameters

data_source	an enum value representing if the location of the data in a memory hierarchy.
PacketType	determines the type of packet
DataScalar	determines the output data type

Parameters

packet_data	the data to be written in the local memory
ptr	a pointer to the local memory
CIndex	is the contracting dim index

Definition at line 220 of file TensorContractionSycl.h.

◆ write() [2/3]

template<data_source dt, typename PacketType , typename DataScalar >

static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ::Eigen::internal::enable_if< Eigen::internal::unpacket_traits<PacketType>::size != 1 && dt == data_source::global_mem, void>::type Eigen::TensorSycl::internal::write	(	PacketType &	packet_data,
		DataScalar *	ptr
	)

static

Overloading the write function for storing the data to global memory, when vectorization enabled This function is used to guarantee coalesced and vectorized store whenever possible.

Template Parameters

data_source	an enum value representing if the location of the data in a memory hierarchy.
PacketType	determines the type of packet
DataScalar	determines the output data type

Parameters

packet_data	the data to be written in the local memory
ptr	a pointer to the local memory

Definition at line 247 of file TensorContractionSycl.h.

◆ write() [3/3]

template<data_source dt, typename PacketType , typename DataScalar >

static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ::Eigen::internal::enable_if< Eigen::internal::unpacket_traits<PacketType>::size == 1 && dt == data_source::global_mem, void>::type Eigen::TensorSycl::internal::write	(	PacketType &	packet_data,
		DataScalar *	ptr
	)

static

Overloading the write function for storing the data to global memory, when vectorization is disabled.

Template Parameters

data_source	an enum value representing if the location of the data in a memory hierarchy.
PacketType	determines the type of packet
DataScalar	determines the output data type

Parameters

packet_data	the data to be written in the local memory
ptr	a pointer to the local memory

Definition at line 267 of file TensorContractionSycl.h.

Classes

Enumerations

Functions

Enumeration Type Documentation

◆ contraction_type

◆ data_source

◆ reduction_dim

◆ scan_step

Function Documentation

◆ check_boundary()

◆ check_boundary< false >()

◆ read() [1/2]

◆ read() [2/2]

◆ write() [1/3]

◆ write() [2/3]

◆ write() [3/3]