AsStridedOperation

Applicable Products

Hardware Model	Supported
Atlas 350 accelerator cards	x
Atlas A3 inference products / Atlas A3 training products	√
Atlas A2 training products / Atlas A2 inference products	√
Atlas training products	√
Atlas inference products	√
Atlas 200I/500 A2 inference products	√

Description

Creates a view of the data re-arranged based on an existing tensor. Elements can be rearranged by specifying the shape and strides.

Mapping (using the 3 x 3 input shape and 2 x 2 output shape as an example)

Input

inTensor = [[1, 2, 3], [4, 5, 6], [7, 8, 9]]
opParam.size = [2, 2]
opParam.stride = [2, 3]
opParam.offset = [0]

Output
```
outTensor = [[1, 4], [3, 6]]
```
Mapping: outTensor[i][j] -> v_inTensor[ i * stride[0] + j * stride[1] + offset[0]].

v_inTensor is the 1D vector of the input tensor.

Figure 1 Implementation principle of the AsStrided operator (offset=0)

Output when offset is not 0:
If offset is added, opParam.offset = [2].

Output: outTensor = [[3, 6], [5, 8]].

Figure 2 Implementation principle of the AsStrided operator (when offset is not 0)

Operator Context

Figure 3 AsStridedOperation

Definition

      
           struct AsStridedParam {
    SVector<int64_t> size;
    SVector<int64_t> stride;
    SVector<int64_t> offset;
    uint8_t rsv[8] = {0};
};

Parameters

Member	Type	Default Value	Description
size	SVector<int64_t>	-	Shape of the output tensor. The length of size must be less than or equal to 8, and each element must be greater than 0.
stride	SVector<int64_t>	-	Stride of each dimension of the output tensor derived based on the input tensor. The length of stride must be the same as that of size, and each element must be greater than or equal to 0.
offset	SVector<int64_t>	-	Offset of the output tensor memory relative to the input tensor memory, which is used as a constant. The length of offset must be 1, and the element must be greater than or equal to 0.
rsv[8]	uint8_t	{0}	Reserved

Input

Parameter	Dimension	Data Type	Format	Description
x	[din_0, ..., din_n]	float16/int64	ND	Input tensor

Output

Parameter	Dimension	Data Type	Format	Description
y	[dout_0, ..., dout_m]	float16/int64	ND	Output tensor

Constraints

Restrictions on the parameters:

offset.size() == 1, offset.at(i) ≥ 0.
size.size() ≤ 8, size.at(i) > 0.
stride.size() == size.size(), stride.at(i) ≥ 0.

Parent topic: atb/infer_op_params.h