xgboost::linalg::TensorView< T, kDim > Class Template Reference

A tensor view with static type and dimension. More...

#include <linalg.h>

Public Types
using	ShapeT = size_t[kDim]
using	StrideT = ShapeT

Public Member Functions
template<typename I , int32_t D>
LINALG_HD	TensorView (common::Span< T > data, I const (&shape)[D], std::int32_t device)
	Create a tensor with data and shape.
template<typename I , int32_t D>
LINALG_HD	TensorView (common::Span< T > data, I const (&shape)[D], std::int32_t device, Order order)
template<typename I , std::int32_t D>
LINALG_HD	TensorView (common::Span< T > data, I const (&shape)[D], I const (&stride)[D], std::int32_t device)
	Create a tensor with data, shape and strides.
template<typename U , std::enable_if_t< common::detail::IsAllowedElementTypeConversion< U, T >::value > * = nullptr>
LINALG_HD	TensorView (TensorView< U, kDim > const &that)
template<typename... Index, detail::EnableIfIntegral< Index... > * = nullptr>
LINALG_HD T &	operator() (Index &&...index)
	Index the tensor to obtain a scalar value.
template<typename... Index, detail::EnableIfIntegral< Index... > * = nullptr>
LINALG_HD T const &	operator() (Index &&...index) const
	Index the tensor to obtain a scalar value.
template<typename... S>
LINALG_HD auto	Slice (S &&...slices) const
	Slice the tensor.
LINALG_HD auto	Shape () const
LINALG_HD auto	Shape (size_t i) const
	Get the shape for i^th dimension.
LINALG_HD auto	Stride () const
LINALG_HD auto	Stride (size_t i) const
	Get the stride for i^th dimension, stride is specified as number of items instead of bytes.
LINALG_HD std::size_t	Size () const
	Number of items in the tensor.
LINALG_HD bool	Contiguous () const
	Whether this is a contiguous array, both C and F contiguous returns true.
LINALG_HD bool	CContiguous () const
	Whether it's a c-contiguous array.
LINALG_HD bool	FContiguous () const
	Whether it's a f-contiguous array.
LINALG_HD auto	Values () const -> decltype(data_) const &
	Obtain a reference to the raw data.
LINALG_HD auto	DeviceIdx () const
	Obtain the CUDA device ordinal.

Static Public Attributes
size_t static constexpr	kValueSize = sizeof(T)
size_t static constexpr	kDimension = kDim

Detailed Description

template<typename T, int32_t kDim>
class xgboost::linalg::TensorView< T, kDim >

A tensor view with static type and dimension.

It implements indexing and slicing.

Most of the algorithms in XGBoost are implemented for both CPU and GPU without using much linear algebra routines, this class is a helper intended to ease some high level operations like indexing into prediction tensor or gradient matrix. It can be passed into CUDA kernel as normal argument for GPU algorithms.

Ideally we should add a template parameter bool on_host so that the compiler can prevent passing/accessing the wrong view, but inheritance is heavily used in XGBoost so some functions expect data types that can be used in everywhere (update prediction cache for example).

Constructor & Destructor Documentation

TensorView() [1/2]

template<typename T , int32_t kDim>

template<typename I , int32_t D>

LINALG_HD xgboost::linalg::TensorView< T, kDim >::TensorView ( common::Span< T >  data, I const (&)  shape[D], std::int32_t  device  )

inline

Create a tensor with data and shape.

Template Parameters

I	Type of the shape array element.
D	Size of the shape array, can be lesser than or equal to tensor dimension.

Parameters

data	Raw data input, can be const if this tensor has const type in its template parameter.
shape	shape of the tensor
device	Device ordinal

TensorView() [2/2]

template<typename T , int32_t kDim>

template<typename I , std::int32_t D>

LINALG_HD xgboost::linalg::TensorView< T, kDim >::TensorView ( common::Span< T >  data, I const (&)  shape[D], I const (&)  stride[D], std::int32_t  device  )

inline

Create a tensor with data, shape and strides.

Don't use this constructor if stride can be calculated from shape.

Member Function Documentation

operator()()

template<typename T , int32_t kDim>

template<typename... Index, detail::EnableIfIntegral< Index... > * = nullptr>

LINALG_HD T & xgboost::linalg::TensorView< T, kDim >::operator() ( Index &&...  index )

inline

Index the tensor to obtain a scalar value.

// Create a 3-dim tensor.
Tensor<float, 3> t {data, shape, 0};
float pi = 3.14159;
t(1, 2, 3) = pi;
ASSERT_EQ(t(1, 2, 3), pi);

Slice()

template<typename T , int32_t kDim>

template<typename... S>

LINALG_HD auto xgboost::linalg::TensorView< T, kDim >::Slice ( S &&...  slices ) const

inline

Slice the tensor.

The returned tensor has inferred dim and shape. Scalar result is not supported.

// Create a 3-dim tensor.
Tensor<float, 3> t {data, shape, 0};
// s has 2 dimensions (matrix)
auto s = t.Slice(1, All(), All());

---

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

• External/xgboost/include/xgboost/linalg.h