Pad

English | 简体中文

Product Support Status

Product	Supported
Ascend 950PR&950DT products	√
Atlas A3 products	√
Atlas A2 products	√
Atlas 200I/500 A2 inference products	√
Atlas inference products	√
Atlas training products	√

Function Description

• Operator function: pads the beginning and end of each dimension of input x with a constant according to the configuration of the input paddings, and the padding value is 0. This operator is compatible with the Pad operator of TensorFlow.

• Example:

  Input x = [[0,1,2]], shape is [1,3]
  paddings = [[1,1],[2,2]]
  
  Output y = [[0,0,0,0,0,0,0],
            [0,0,0,1,2,0,0],
            [0,0,0,0,0,0,0]], shape is [3,7]
  

Parameter Description

Parameter Name	Input/Output/Attribute	Description	Data Type	Data Format
x	Input	Original tensor to be padded.	BFLOAT16, FLOAT16, FLOAT, DOUBLE, INT32, UINT8, INT16, INT8, COMPLEX64, INT64, QINT8, QUINT8, QINT32, QINT16, QUINT16, UINT16, COMPLEX128, UINT32, UINT64, BOOL, HIFLOAT8, FLOAT8_E5M2, FLOAT8_E4M3FN, FLOAT8_E8M0, FLOAT4_E2M1, FLOAT4_E1M2.	ND
paddings	Input	Configuration of padding. The shape is [rank(x), 2], and row i [a, b] indicates that a elements are padded at the beginning of dimension i of x and b elements are padded at the end.	INT32, INT64	ND
y	Output	Tensor after padding. Its data type is the same as that of x.	Same as the data type of x	ND

• Data types supported by the AICore path vary with the platform:
  • Atlas A2 products: FLOAT, FLOAT16, BFLOAT16, INT8, INT16, INT32, UINT16, INT64;
  • Some new-generation platforms additionally support the following types: UINT8, UINT32, UINT64, BOOL, HIFLOAT8, FLOAT8_E5M2, FLOAT8_E4M3FN, FLOAT8_E8M0, FLOAT4_E2M1, FLOAT4_E1M2;
  • Other platforms: FLOAT, FLOAT16, BFLOAT16, INT16, INT32, UINT16, INT64.
• Data types not in the preceding AICore support list (such as DOUBLE, COMPLEX64, COMPLEX128, and QINT8) are computed through the AICPU path.

Constraint Description

• The shape of paddings must be [rank(x), 2], where rank(x) is the number of dimensions of input x.
• If the data type of x is FLOAT4_E2M1 or FLOAT4_E1M2, the value of paddings should be an even number.
• If the data type of x is HIFLOAT8, FLOAT8_E5M2, FLOAT8_E4M3FN, or FLOAT8_E8M0, the value of paddings should be a non-negative integer.

Invocation Description

Invocation Method	Invocation Sample	Description
GE graph mode	test_geir_pad	Invokes the Pad operator by building a graph through the operator IR.