Overview

The TensorTrait data structure is a basic template class that describes tensor-related information, including the tensor data type, logical position, and layout memory layout. With the template metaprogramming technology, this class completes computation and code generation during compilation, thereby reducing 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 Parameters

**Table 1** Template parameters
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. The LocalTensor/GlobalTensor defined by TensorTrait does not contain ShapeInfo. For example, the tensor corresponding to LocalTensor<float> that does not contain ShapeInfo is LocalTensor<TensorTrait<float>>.
pos	Logical position for storing data. 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<>>. This parameter must comply with 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 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 both the GlobalTensor/LocalTensor of the TensorTrait type and GlobalTensor/LocalTensor of the non-TensorTrait type.
Copy constructors and assignment operators are not supported between non-TensorTrait and TensorTrait GlobalTensor/LocalTensor 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 for pos and LayoutType.
The DataCopy slice data transfer API requires ShapeInfo and does not support the GlobalTensor/LocalTensor of the TensorTrait type.

**Table 2** APIs supported by TensorTrait
API Category	API
Basic APIs > Resource management > TQue/TQueBind	AllocTensor, FreeTensor, EnQue, DeQue
Basic APIs > 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 APIs > Vector computation > Logical computation	And, Or
Basic APIs > Vector computation > Compound computation	CastDeq, AddRelu, AddDeqRelu, SubRelu, MulAddDst, FusedMulAdd, FusedMulAddRelu, AddReluCast, SubReluCast, MulCast
Basic APIs > Data movement	DataCopy, Copy
Basic APIs > Matrix computation	InitConstValue, LoadData, LoadDataWithTranspose, SetAippFunctions, LoadImageToLocal, LoadUnzipIndex, LoadDataUnzip, LoadDataWithSparse, Mmad, MmadWithSparse, BroadCastVecToMM, Gemm, Fixpipe
Basic APIs > Vector computation > Comparison and selection	Compare, GetCmpMask, SetCmpMask, Select, GatherMask
Basic APIs > Vector computation > Type conversion	Cast
Basic APIs > Vector computation > Reduction computation	ReduceMax, BlockReduceMax, WholeReduceMax, ReduceMin, BlockReduceMin, WholeReduceMin, ReduceSum, BlockReduceSum, WholeReduceSum, RepeatReduceSum, PairReduceSum
Basic APIs > Vector computation > Data conversion	Transpose, TransDataTo5HD
Basic APIs > Vector computation > Data padding	Brcb
Basic APIs > Vector computation > Discretization and aggregation	Gather, Gatherb, Scatter
Basic APIs > Vector computation > Sorting and combination (ISASI)	ProposalConcat, ProposalExtract, RpSort16, MrgSort4, Sort32

Parent topic: TensorTrait