SliceOperation

Applicable Products

Hardware Model	Supported	Remarks
Atlas 350 accelerator cards	√	-
Atlas A3 inference products/Atlas A3 training products	√	-
Atlas A2 training products/Atlas A2 inference products	√	-
Atlas training products	√	-
Atlas inference products	√	The bf16 data type is not supported.
Atlas 200I/500 A2 inference products	√	Only the float16 data type is supported.

Description

Extracts a slice with a specified size from a start location of the input tensor.

Operator Context

Figure 1 Context of the SliceOperation operator

Definition

struct SliceParam {
    SVector<int64_t> offsets;
    SVector<int64_t> size;
    uint8_t rsv[8] = {0};
};

Parameters

Member	Type	Default Value	Description
offsets	SVector<int64_t>	-	Start point of each dimension slice. offsets[i] can be positive or negative. offsets[i] ≥ 0: The actual start point is offsets[i]. offsets[i] < 0: The actual start point is dimNum + offsets[i]. The actual start position must be greater than or equal to 0 and less than dimNum.
size	SVector<int64_t>	-	Size of each dimension slice. The value of size[i] must be greater than 0 or equal to -1. The value 0 is not allowed. If size[i] is -1, the slice extends to the end of the i^th dimension. If size[i] is greater than 0, the slice length is size[i]. For the i^th dimension, the dimension size is dimNum. The actual start point and length are defined as follows: Actual start point: If offsets[i] ≥ 0: actual_offset = offsets[i] If offsets[i] < 0: actual_offset = dimNum + offsets[i] Actual length: If size[i] > 0: actual_size = size[i] If size[i] = -1: actual_size = dimNum - actual_offset Validity check: actual_offset >= 0 actual_offset < dimNum actual_offset + actual_size <= dimNum actual_size > 0
rsv[8]	uint8_t	{0}	Reserved

Member

Type

Default Value

Description

offsets

SVector<int64_t>

Start point of each dimension slice.

offsets[i] can be positive or negative.

offsets[i] ≥ 0: The actual start point is offsets[i].
offsets[i] < 0: The actual start point is dimNum + offsets[i].
The actual start position must be greater than or equal to 0 and less than dimNum.

size

SVector<int64_t>

Size of each dimension slice.

The value of size[i] must be greater than 0 or equal to -1. The value 0 is not allowed.
- If size[i] is -1, the slice extends to the end of the i^th dimension.
- If size[i] is greater than 0, the slice length is size[i].

For the i^th dimension, the dimension size is dimNum. The actual start point and length are defined as follows:
- Actual start point:
  If offsets[i] ≥ 0: actual_offset = offsets[i]
  
  If offsets[i] < 0: actual_offset = dimNum + offsets[i]
- Actual length:
  If size[i] > 0: actual_size = size[i]
  
  If size[i] = -1: actual_size = dimNum - actual_offset

Validity check:
- actual_offset >= 0
- actual_offset < dimNum
- actual_offset + actual_size <= dimNum
- actual_size > 0

rsv[8]

uint8_t

{0}

Reserved

Input

Parameter	Dimension	Data Type	Format	Description
x	[dim_0, dim_1, ..., dim_n]	float16/float/int8/bool/int32/uint32/bf16	ND	Input tensor.

Output

Parameter	Dimension	Data Type	Format	Description
output	Same as the shape indicated by size	float16/float/int8/bool/int32/uint32/bf16	ND	Output tensor. The dimension size is the value specified by size. Same as the data type of x.

Example API Call

Input

x = [[ -78.9,   78.8,  32.22,   49.7,  82.8, 24.77],
     [  72.2,  20.89,  55.16,   85.5,  72.8,  94.8],
    [ 11.36, -83.56,  65.94, -46.94, 72.44,  68.7]]
offsets = [1, 2]

size = [2, 4]

In dimension 0, the slice length is 2 from index 1.

In the first dimension, the slice length is 4 (2 + 4 < = 6) from index 2.

Output

y.dim = [2, 4]
y = [[ 55.1562,  85.5000,  72.8125,  94.8125],
     [ 65.9375, -46.9375,  72.4375,  68.6875]]

If the dimension of size is -1, the output shape is automatically derived.

Parent topic: atb/infer_op_params.h