CumsumOperation

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

Returns the cumulative sum along the given axes.

Figure 1 CumsumOperation

Input:

Array: The input of cumsum can be a multidimensional array.
Axis: You can select an axis (dimension) along which the cumulative sum operation is performed. By default, for a one-dimensional array, the cumulative sum is calculated backwards from the first element (including the element itself by default). For a multi-dimensional array, specify the axis along which the cumulative sum is performed.

Output:

Result of cumulative sum: The shape of the output is the same as that of the input. For each element, the sum includes the element and all previous elements.

Formula

If the input is a one-dimensional tensor of size N, the output is also a one-dimensional tensor of size N, where the values of the elements are:

Definition

      
           struct CumsumParam {
    SVector<int64_t> axes;
    bool exclusive = false;
    bool reverse =  false;
    uint8_t rsv[14] = {0};
};

Parameters

Member	Type	Default Value	Description
axes	SVector<int64_t>	-	Calculates the accumulated sum along a specified axis (dimension). Only one axis index can be included. The value of axes must be less than the number of dimensions of the input x.
exclusive	bool	false	Sequence number of the element from which the accumulation on an axis starts. The default value is false. true: The accumulation result does not contain the element itself (not supported currently). false: The accumulation result contains the element itself.
reverse	bool	false	Forward accumulation or reverse accumulation. The default value is false. true: performs summation in reverse direction (not supported currently). false: performs summation in forward direction.
rsv[14]	uint8_t	{0}	Reserved

Input

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

Output

Parameter	Dimension	Data Type	Format	Description
y	[dim_0, dim_1, ..., dim_n]	float16/bf16	ND	Output tensor.

Constraints

The size of the axes array must be 1.
The value of reverse must be false.
The value of exclusive must be false.

Parent topic: atb/infer_op_params.h