Constructor

Function Usage

Constructs a specified tensor object. The TensorV2 class is used to describe the information and behavior of a tensor object, including the shape, data format, data type, step, offset, and tensor data.

Prototype

TensorV2()   //Default constructor
TensorV2 (const StorageShape &storage_shape, const StorageFormat &storage_format, const TensorPlacement placement, const ge::DataType data_type, TensorAddress addr)
TensorV2(const StorageShape &storage_shape, const StorageFormat &storage_format, ge::DataType data_type)
TensorV2(const StorageShape &storage_shape, const StorageFormat &storage_format, const TensorPlacement placement, const ge::DataType data_type, TensorAddress addr, TensorAddrManager manager)
TensorV2(const StorageShape &storage_shape, const StorageFormat &storage_format, const TensorPlacement placement, const ge::DataType data_type, TensorAddress addr, TensorAddrManager manager, const Stride &stride, const int64_t offset)

Parameters

Parameter

Input/Output

Description

storage_shape

Input

Shape of a specified tensor. It is of the StorageShape type.

storage_format

Input

Data format of a specified tensor. It is of the StorageFormat type.

placement

Input

Location of the device where the actual tensor data is stored. It is of the TensorPlacement type.

data_type

Input

Data type of a specified tensor. It is of the ge::DataType.

addr

Input

Memory address that stores the actual tensor data. It is of the TensorAddress type, which is defined as follows:

using TensorAddress = void *;

manager

Input

Management function of the tensor data. If the value of manager is empty, addr is considered as the tensor data address and the data does not need to be deallocated.

using TensorAddrManager = ge::graphStatus (*)(TensorAddress addr, TensorOperateType operate_type, void **out);
enum TensorOperateType {
  kGetTensorAddress,  ///< Obtain the tensor address.
  kFreeTensor,        ///< Free the tensor.
  kPlusShareCount,    ///< Share the tensor.
  kTensorOperateType
};

stride

Input

Step of a specified tensor. It is of the Stride type.

offset

Input

Offset of a specified tensor. It is of the int64_t type.

Returns

An initialized tensor object

Constraints

The values of the members (such as format, shape, stride, and offset) in the tensor must be configured explicitly. Otherwise, the values are undefined.

The TensorV2 class cannot be integrated into a graph.

After a TensorV2 pointer is converted into a Tensor pointer, the tensor member function can be called.

Before converting a Tensor pointer to a TensorV2 pointer, you need to call GetVersion() to obtain the version. Only kTensorV2 supports the conversion.

Examples

TensorV2 tensor{{{8, 3, 224, 224}, {16, 3, 224, 224}},       // shape                
                {ge::FORMAT_ND, ge::FORMAT_FRACTAL_NZ, {}},  // format                              
                kFollowing,                                  // placement                
                ge::DT_FLOAT16,                              //dt                
                nullptr};

Parent topic: TensorV2