FillOperation

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	√
Atlas inference products	√
Atlas 200I/500 A2 inference products	√

Description

Sets a specified position to value or generates a tensor with a specified shape and fills it to value.

Figure 1 FillOperation

MaskedFill (withMask = true):
For an element in x, if the value of the element in the same location in mask is true or 1, the value of the element in the same location in output is set to value; the values of the elements in other locations in output are the same as those of x. The broadcast function is supported.

Figure 2 Implementation principle of the Fill operator

Figure 3 Implementation principle of the Fill operator (broadcast scenario)

Fill (withMask = false)
Generates a tensor with the shape specified by outDim. The values of all elements are value.

Definition

struct FillParam {
    bool withMask = true;
    SVector<float> value;
    SVector<int64_t> outDim;
    uint8_t rsv[8] = {0};
}

Parameters

Member	Type	Default Value	Description
withMask	bool	true	Masked filing.
value	SVector<float>	[]	Padded element. value is a SVector containing only one element.
outDim	SVector<float>	[]	Shape of the output tensor when withMask = false
rsv[8]	uint8_t	{0}	Reserved

withMask = true
Sets the element in the location specified by mask to value.

If the number of input tensors is 2, outDim does not need to be specified.

withMask = false
Generates a tensor with the shape specified by outDim and fills each element of the tensor with value.

If the number of input tensors is 0, outDim must be specified.

Input

Parameter	Dimension	Data Type	Format	Description
x	[dx_0, ..., dx_n]	float16/int32/bf16	ND	This parameter is valid only when withMask is set to true.
mask	[dmask_0, ..., dmask_m]	int8/bool	ND	This parameter is valid only when withMask is set to true. You can set the values in certain locations to true or 1 to set the value in the same location as x to value.

Output

Parameter	Dimension	Data Type	Format	Description
output	Same as x when withMask is true Specified by outDim when withMask is false	Same as x when withMask is true Must be of the float16 type when withMask is false	ND	Output tensor.

Restrictions

Restrictions on the parameters:
a. value.size() == 1.

Input and output constraints
- Input x cannot be broadcast.
- Each element of the input mask can only be 0 or 1.
- When withMask is true, the shape of mask must be the same as that of x or can be broadcast to the shape of x.

API Calling Example

Example 1: MaskedFill non-broadcast scenario

Input

param.withMask = true
param.value = [-1]
x = [[1, 2, 3],
    [4, 5, 6],
    [7, 8, 9]]
mask = [[ true, false,  true],
        [false,  true,  true],
        [ true,  true, false]]

Output

output = [[-1,  2, -1],
          [ 4, -1, -1],
          [-1, -1,  9]]

Example 2: MaskedFill broadcast scenario

Input

param.withMask = true
param.value = [-1]
x = [[1, 2, 3],
    [4, 5, 6],
    [7, 8, 9]]
mask = [true, false, true]

Output

output = [[-1,  2, -1],
          [-1,  5, -1],
          [-1,  8, -1]]

Example 3: Fill scenario

Input

param.withMask = false
param.value = [1]
param.outDim = [2, 3]

Output

output = [[1, 1, 1],
           [1, 1, 1]]

Parent topic: atb/infer_op_params.h