TensorTrait Overview

The TensorTrait data structure is a basic template class that describes tensor information, including the data type, logical location, and layout memory layout of the tensor. With the help of template metaprogramming technology, this class completes computation and code generation at compile time, reducing the runtime overhead.

Header Files to Be Included

#include "kernel_operator_tensor_trait.h"

Prototype

template <typename T, TPosition pos = TPosition::GM, typename LayoutType = Layout<Shape<>, Stride<>>>
struct TensorTrait {
    using LiteType = T;
    using LiteLayoutType = LayoutType;
    static constexpr const TPosition tPos = pos; // This constant member is reserved for future function extension.
public:
    __aicore__ inline TensorTrait(const LayoutType& t = {});

    __aicore__ inline LayoutType& GetLayout();
    __aicore__ inline const LayoutType& GetLayout() const;

    __aicore__ inline void SetLayout(const LayoutType& t);

};

Template parameter

**Table 1** Parameters in the template
Parameter	Description
T	Only the following basic data types are supported: int4b_t, uint8_t, int8_t, int16_t, uint16_t, bfloat16_t, int32_t, uint32_t, int64_t, uint64_t, float, and half. In the TensorTrait structure, the using keyword is used to define a type alias LiteType, which is the same as the template parameter T type. The LocalTensor/GlobalTensor defined by TensorTrait does not contain ShapeInfo. For example, the tensor that does not contain the ShapeInfo information corresponding to LocalTensor<float> is LocalTensor<TensorTrait<float>>.
pos	Logical position where data is stored. The value is of the Tposition type. The default value is TPosition::GM.
LayoutType	Layout data type. The default value is empty, that is, Layout<Shape<>, Stride<>>. The input data type LayoutType must meet the Restrictions.

Member Function

__aicore__ inline TensorTrait(const LayoutType& t = {})
__aicore__ inline LayoutType& GetLayout()
__aicore__ inline const LayoutType& GetLayout() const
__aicore__ inline void SetLayout(const LayoutType& t)

Related APIs

// TensorTrait structure construction method
template <typename T, TPosition pos, typename LayoutType>
__aicore__ inline constexpr auto MakeTensorTrait(const LayoutType& t)

// is_tensorTrait prototype definition
template <typename T> struct is_tensorTrait

Restrictions

The same API does not support the input of GlobalTensor and LocalTensor of the TensorTrait and non-TensorTrait types at the same time.
Copy constructors and assignment operators are not supported between GlobalTensor and LocalTensor of the non-TensorTrait and TensorTrait types.

Currently, the TensorTrait feature supports only the following APIs.

When used with APIs, the TensorTrait structure does not support the configuration of the pos and LayoutType template parameters. You need to use the constructor to construct the TensorTrait, and retain the default values of pos and LayoutType.
The DataCopy API for data movement requires the ShapeInfo information and does not support the input of GlobalTensor or LocalTensor of the TensorTrait type.

**Table 2** APIs supported by TensorTrait
API Category	API
Basic API> resource management > TQue/TQueBind	AllocTensor, FreeTensor, EnQue, DeQue
Basic API> vector computation > Basic arithmetic	Exp, Ln, Abs, Reciprocal, Sqrt, Rsqrt, Relu, Add, Sub, Mul, Div, Max, Min, Adds, Muls, Maxs, Mins, VectorPadding, BilinearInterpolation, LeakyRelu
Basic API> vector computation > Logical computation	And, Or
Basic API> vector computation > Compound computation	CastDeq, AddRelu, AddDeqRelu, SubRelu, MulAddDst, FusedMulAdd, FusedMulAddRelu, AddReluCast, SubReluCast, MulCast
Basic APIs > Data movement	DataCopy, Copy
Basic API> matrix computation	InitConstValue, LoadData, LoadDataWithTranspose, SetAippFunctions, LoadImageToLocal, LoadUnzipIndex, LoadDataUnzip, LoadDataWithSparse, Mmad, MmadWithSparse, BroadCastVecToMM, Gemm, Fixpipe
Basic API> vector computation > Comparison and selection	Compare, GetCmpMask, SetCmpMask, Select, GatherMask
Basic API> vector computation > Type conversion	Cast
Basic API> vector computation > Reduction computation	ReduceMax, BlockReduceMax, WholeReduceMax, ReduceMin, BlockReduceMin, WholeReduceMin, ReduceSum, BlockReduceSum, WholeReduceSum, RepeatReduceSum, PairReduceSum
Basic API> vector computation > Data conversion	Transpose, TransDataTo5HD
Basic API> vector computation > Data filling	Brcb
Basic API> vector computation > Discretization and aggregation	Gather, Gatherb, Scatter
Basic API> vector computation > Sorting and combination (ISASI)	ProposalConcat, ProposalExtract, RpSort16, MrgSort4, Sort32

Parent topic: TensorTrait