ScatterElementsV2Operation

Applicable Products

Hardware Model	Supported or Not
Atlas 350 accelerator card	x
Atlas A3 inference products/Atlas A3 training products	√
Atlas A2 training products/Atlas A2 inference products	√
Atlas training products	x
Atlas inference products	x
Atlas 200I/500 A2 inference products	x

Description

Writes all values in the tensor update to the tensor input based on the indices position.

Definition

struct ScatterElementsV2Param {
    enum ReductionType {
        NONE = 0,
        ADD, 
    };
    int32_t axis = -1;
    ReductionType reduction = NONE;
    uint8_t rsv[24] = {0};
};

Parameters

Member	Type	Default Value	Value Range	Mandatory or Not	Description
axis	int32_t	-1	-1	No	Used in performing an update or accumulation along the specified axis. Currently, only the value -1 is supported.
reduction	int32_t	0	0,1	No	Whether to perform in-place update or in-place accumulation. 0: in-place update. 1: in-place accumulation.
rsv[24]	[uint8_t]	{0}	[0]	No	Reserved

Member

Type

Default Value

Value Range

Mandatory or Not

Description

axis

int32_t

-1

Used in performing an update or accumulation along the specified axis. Currently, only the value -1 is supported.

reduction

int32_t

0,1

Whether to perform in-place update or in-place accumulation.

0: in-place update.
1: in-place accumulation.

rsv[24]

[uint8_t]

{0}

[0]

Reserved

Input

Parameter	Dimension	Data Type	Format	Description
input	[x_0, x_1, ..., x_n]	float16/float/uint8/int8/int32/bfloat16	ND	Original input tensor, to which the final result is written in place.
indice	[y_0, y_1, ..., y_n]	int64/int32	ND	Index. The shape of the non-tail axis and non-axis 0 must be the same as that of input. Axis 0 and the tail axis cannot be greater than those of input. The value range of the corresponding dimension dim in indice and input must be [0, corresponding dimension dim in input - 1]. For example, if the shape of input is [3, 3, 3], the value range of indice is [0, 2].
update	[y_0, y_1, ..., y_n]	Same as input	ND	Tensor to be updated or accumulated to the original input tensor.

Parameter

Dimension

Data Type

Format

Description

input

[x_0, x_1, ..., x_n]

float16/float/uint8/int8/int32/bfloat16

Original input tensor, to which the final result is written in place.

indice

[y_0, y_1, ..., y_n]

int64/int32

Index. The shape of the non-tail axis and non-axis 0 must be the same as that of input.

Axis 0 and the tail axis cannot be greater than those of input.

The value range of the corresponding dimension dim in indice and input must be [0, corresponding dimension dim in input - 1]. For example, if the shape of input is [3, 3, 3], the value range of indice is [0, 2].

update

[y_0, y_1, ..., y_n]

Same as input

Tensor to be updated or accumulated to the original input tensor.

Output

Parameter	Dimension	Data Type	Format	Description
input	[dim_0, dim_1,...,dim_n]	float16/float/uint8/int8/int32/bfloat16	ND	Original input tensor, to which the final result is written in place.

Parent topic: atb/infer_op_params.h