PadV2

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 input tensor with a constant of a specified value.

• Example:

  Input tensor([[0,1,2]])
  paddings([[0,0],[2,2]])
  constant_values(0)
  
  The output is ([[0,0,0,1,2,0,0]])
  

Parameter Description

Parameter Name	Input/Output/Attribute	Description	Data Type	Data Format
x	Input	Original tensor to be padded.	FLOAT, FLOAT16, INT32, INT64, INT16, INT8, UINT8, UINT16, UINT32, UINT64, BOOL, DOUBLE, BFLOAT16, HIFLOAT8, FLOAT8_E5M2, FLOAT8_E4M3FN, FLOAT8_E8M0, FLOAT4_E2M1, FLOAT4_E1M2.	ND
paddings	Input	Padding configuration. shape=[N, 2], where N is the number of dimensions of x.	INT32, INT64	ND
constant_values	Input	Padding constant value, which is a scalar tensor.	FLOAT, FLOAT16, INT32, INT64, INT16, INT8, UINT8, UINT16, UINT32, UINT64, BOOL, DOUBLE, BFLOAT16, HIFLOAT8, FLOAT8_E5M2, FLOAT8_E4M3FN, FLOAT8_E8M0, FLOAT4_E2M1, FLOAT4_E1M2.	ND
y	Output	Tensor after padding.	FLOAT, FLOAT16, INT32, INT64, INT16, INT8, UINT8, UINT16, UINT32, UINT64, BOOL, DOUBLE, BFLOAT16, HIFLOAT8, FLOAT8_E5M2, FLOAT8_E4M3FN, FLOAT8_E8M0, FLOAT4_E2M1, FLOAT4_E1M2.	ND

• Atlas training products, Atlas inference products, and Atlas 200I/500 A2 inference products: the data types BFLOAT16, HIFLOAT8, FLOAT8_E5M2, FLOAT8_E4M3FN, FLOAT8_E8M0, FLOAT4_E2M1, and FLOAT4_E1M2 are not supported.
• Atlas A3 products and Atlas A2 products: the data types HIFLOAT8, FLOAT8_E5M2, FLOAT8_E4M3FN, FLOAT8_E8M0, FLOAT4_E2M1, and FLOAT4_E1M2 are not supported.

Output Shape Computation

y.shape[d] = x.shape[d] + paddings[d][0] + paddings[d][1]

Constraint Description

1. Dimension constraints:

   • The number of dimensions of x must be in the range [1, 8]
   • The first dimension of paddings must be equal to the number of dimensions of x
   • The second dimension of paddings must be equal to 2

2. Data type constraints:

   • x, constant_values, and y must use the same data type
   • paddings must use INT32 or INT64

3. fp8/fp4 type constraints:

   • When x is HIFLOAT8, FLOAT8_E5M2, FLOAT8_E4M3FN, FLOAT8_E8M0, FLOAT4_E2M1, or FLOAT4_E1M2, each dimension of paddings must be a non-negative value
   • When x is FLOAT4_E2M1 or FLOAT4_E1M2, the shape of the trailing axis of the input must be an even number, and the paddings of the trailing axis must be an even number

4. Padding constraints:

   • For negative padding (slice), the output shape = x.shape[d] + left + right >= 0

5. paddings parameter constraints

   • The shape of paddings must be [rank, 2], where rank is the number of dimensions of input x (1 to 8)
   • Each row [left, right] indicates the number of paddings in the corresponding dimension
     • left: the number of elements padded at the beginning of the dimension
     • right: the number of elements padded at the end of the dimension
   • The value of paddings can be:
     • A positive number: indicates padding
     • A negative number: indicates slice (cropping)
     • Zero: indicates no padding

Invocation Description

Invocation Method	Sample Code	Description
GE graph mode	test_geir_pad_v2.cpp	Invokes the PadV2 operator through the operator IR API