UnPad

Applicability

Product	Supported
Atlas 350 Accelerator Card	√
Atlas A3 training product/Atlas A3 inference product	√
Atlas A2 training product/Atlas A2 inference product	√
Atlas 200I/500 A2 inference product	x
Atlas inference product AI Core	√
Atlas inference product Vector Core	x
Atlas training product	x

Function Usage

Unpads a two-dimensional tensor (height x width) in the width direction. If the width of the tensor is not 32-byte aligned, this API should not be called for unpadding. This API is used in the following scenario: The width of the tensor is already 32-byte aligned. Take half as an example, for example, 16 x 16, and unpad it to 16 x 15.

Prototype

Due to the complex computation involved in the internal implementation of this API, additional temporary space is required to store intermediate variables generated during computation. The method of obtaining the temporary space size (BufferSize) is as follows: Obtain the required maximum and minimum temporary space sizes using the GetUnPadMaxMinTmpSize API provided in UnPad Tiling. The minimum space can ensure correct functionality, while the maximum space is used to improve performance.

The temporary space can be allocated through the API framework or passed by developers through the sharedTmpBuffer input parameter. Therefore, there are two types of function prototypes for the UnPad API.

Pass the temporary space through the sharedTmpBuffer input parameter.

template <typename T>
__aicore__ inline void UnPad(const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, UnPadParams& unPadParams, LocalTensor<uint8_t>& sharedTmpBuffer, UnPadTiling& tiling)

This method requires you to allocate and manage the temporary buffer space on your own, and reuse the buffer after calling the API, so that the buffer is not repeatedly allocated or deallocated, improving the flexibility and buffer utilization.

Allocate the temporary space through the API framework.

template <typename T>
__aicore__ inline void UnPad(const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, UnPadParams& unPadParams, UnPadTiling& tiling)

When using this method, you do not need to allocate the space, but must reserve the required temporary space.

Parameters

**Table 1** Template parameters
Parameter	Description
T	Operand data type. For the Atlas 350 Accelerator Card, the supported data types are int16_t, uint16_t, half, int32_t, uint32_t, and float. For the Atlas A3 training product/Atlas A3 inference product, the supported data types are int16_t, uint16_t, half, int32_t, uint32_t, and float. For the Atlas A2 training product/Atlas A2 inference product, the supported data types are int16_t, uint16_t, half, int32_t, uint32_t, and float. For the Atlas inference product AI Core, the supported data types are int16_t, uint16_t, half, int32_t, uint32_t, and float.

Table 2 API parameters

Parameter

Input/Output

Description

dstTensor

Output

Destination operand, with a two-dimensional shape. For details about the definition of the LocalTensor data structure, see LocalTensor.

The type is LocalTensor, and TPosition can be VECIN, VECCALC, or VECOUT.

srcTensor

Input

Source operand, with a two-dimensional shape. For details about the definition of the LocalTensor data structure, see LocalTensor.

The type is LocalTensor, and TPosition can be VECIN, VECCALC, or VECOUT.

UnPadParams

Input

Detailed UnPad parameters of the UnPadParams data type. The specific parameters of the struct are as follows:

leftPad: size of data for left unpadding. The value of leftPad must be less than 32 bytes. The unit is column. This parameter does not take effect currently.
rightPad: size of data for right unpadding. The value of rightPad must be greater than 0 and less than 32 bytes. The unit is column. Currently, the unpadding operation can be performed only on the right.

The definition of the UnPadParams struct is as follows:

struct UnPadParams {
    uint16_t leftPad = 0;
    uint16_t rightPad = 0;
};

sharedTmpBuffer

Input

Shared buffer, which is used to store temporary data generated during internal API computation. This enables you to manage the sharedTmpBuffer space and reuse the buffer after calling the API, so that the buffer is not repeatedly allocated and deallocated, improving the flexibility and buffer utilization. For details about how to obtain the size of the shared buffer, see UnPad Tiling.

The type is LocalTensor, and TPosition can be VECIN, VECCALC, or VECOUT.

tiling

Input

Tiling information required for computation. For details about how to obtain the tiling information, see UnPad Tiling.

Returns

None

Constraints

For details about the operand address alignment requirements, see General Address Alignment Restrictions.

Examples

In this sample, the width of the tensor is already 32-byte aligned. Take half as an example: For instance, 16 x 16 becomes 16 x 15 after unpadding. The input data type is half.

For details about the complete example, see UnPad operator sample.

// dstLocal: output tensor
// srcLocal: input tensor
// unPadParams: unpadding parameters
AscendC::UnPadParams unPadParams{0, 1}; // Remove 0 columns from the left and 1 column from the right. Currently, unpadding on the left is not supported.
AscendC::UnPad(srcOutLocal, dstLocal, unPadParams, tilingData.unpadTilingData);

Parent topic: Tensor Transformation