SetValueOperation

Applicable Products

Hardware Model	Supported	Remarks
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	√	The bf16 data type is not supported.
Atlas 200I/500 A2 inference products	√	-

Description

Copies the content of the input source tensor to the specified location of the input target tensor.

The copy is an in-place copy, and the final result is modified in the input target tensor.

Example:

The input target tensor is dst: (a, b, c), and the input source tensor is src: (d, e, f).

dst[starts[0]: ends[0], starts[1]: ends[1], starts[2]: ends[2]] = src

The length of starts[0]:ends[0] is ends[0] - starts[0], the length of starts[1]:ends[1] is ends[1] - starts[1], and the length of starts[2]:ends[2] is ends[2] - starts[2].

Operator Context

Definition

struct SetValueParam {
     SVector<int64_t> starts;
     SVector<int64_t> ends;
     SVector<int64_t> strides;
     uint8_t rsv[8] = {0};
};

Members

Member	Type	Default Value	Description
starts	SVector<int64_t>	-	Start position for copying each dimension.
ends	SVector<int64_t>	-	One position after the end position of each dimension copy. The copy ends at the previous position of this position.
strides	SVector<int64_t>	-	Copy step of each dimension. Currently, strides can only be all 1s.
rsv[8]	uint8_t	{0}	Reserved

Input

Parameter	Dimension	Data Type	Format	Description
dst	[dim_0, dim_1, ..., dim_n]	float16/float/int32/int64/bf16	ND	Input tensor 1
src	All dimensions of src are less than or equal to those of dst.	Same as dst	ND	Input tensor 2

Output

Parameter	Dimension	Data Type	Format	Description
dst	[dim_0, dim_1, ..., dim_n]	Same as src	ND	Output tensor. Note: The output dst points to the same address as the input with the same name, that is, in-place modification is performed.

Parameter

Dimension

Data Type

Format

Description

dst

[dim_0, dim_1, ..., dim_n]

Same as src

Output tensor.

Note: The output dst points to the same address as the input with the same name, that is, in-place modification is performed.

Constraints

The number of dimensions of input src must be the same as that of input dst.
The size of each dimension of input src must be less than or equal to that of the corresponding dimension of input dst.
One or two dimensions of input src and input dst must be different and meet the following requirements:
- If one dimension is different, this dimension cannot be the highest dimension (dimension 0).
- If two dimensions are different, one of the different dimensions must be the highest dimension (dimension 0).

API Calling Example

Input:
starts = {1,2}
ends = {4,4}
strides = {1,1}
dst=[[ 14,  -2,   2, -10, -12],
     [-11,   3,   2, -18,   6],
     [  1, -18, -12,   9,  -6],
     [ 16, -11,   6,  -6,   3],
     [ 12,  18,  14,  10,   9]]
src=[[  1,   0],
     [ -4,   7],
     [  9, -18]]
Output:
dst=[[ 14,  -2,   2, -10, -12],
     [-11,   3,   1,   0,   6],
     [  1, -18,  -4,   7,  -6],
     [ 16, -11,   9, -18,   3],
     [ 12,  18,  14,  10,   9]]

Parent topic: atb/infer_op_params.h