Input and Output

ELEWISE_CAST (Data Type Conversion)

Parameter

Dimension

Data Type

Format

Description

x

[dim_0, dim_1, ..., dim_n]

float16/float/int64/int32/bf16

ND

Input before conversion

out

[dim_0, dim_1, ..., dim_n]

float16/float/int64/int32/bf16

ND

Output after conversion

On Atlas 350 accelerator card, Atlas A3 inference products/Atlas A3 training products, and Atlas A2 training products/Atlas A2 inference products, data types can be converted between float16 and float, between int32 and int64, and between bf16 and float.
On Atlas inference products, data types can be converted between float16 and float.

ELEWISE_MULS (Element-wise Vector Multiplication)

Parameter	Dimension	Data Type	Format	Description
x	[dim_0, dim_1, ..., dim_n]	float16/float/bf16	ND	Input.
out	[dim_0, dim_1, ..., dim_n]	float/float16/bf16	ND	Output. The data type is the same as that of x.

ELEWISE_COS (Element-wise Cosine Computation)

Parameter	Dimension	Data Type	Format	Description
x	[dim_0, dim_1, ..., dim_n]	float/float16/bf16	ND	Input.
out	[dim_0, dim_1, ..., dim_n]	float/float16/bf16	ND	Output. The data type is the same as that of x.

ELEWISE_SIN (Element-wise Sine Computation)

Parameter	Dimension	Data Type	Format	Description
x	[dim_0, dim_1, ..., dim_n]	float/float16/bf16	ND	Input.
out	[dim_0, dim_1, ..., dim_n]	float/float16/bf16	ND	Output. The data type is the same as that of x.

ELEWISE_NEG (Opposite Number Computation Element-wise)

Parameter	Dimension	Data Type	Format	Description
x	[dim_0, dim_1, ..., dim_n]	float16/bf16	ND	Input.
out	[dim_0, dim_1, ..., dim_n]	float16/bf16	ND	Output. The data type is the same as that of x.

ELEWISE_QUANT (Quantization)

Parameter	Dimension	Data Type	Format	Description
x	[dim_0, dim_1, ..., dim_n]	float16	ND	Input. The size of the last dimension must be a multiple of 32.
out	[dim_0, dim_1, ..., dim_n]	int8	ND	Output.
x/out does not support the Atlas inference products.

ELEWISE_LOGICAL_NOT (Element-wise Logical NOT)

Parameter	Dimension	Data Type	Format	Description
x	[dim_0, dim_1, ..., dim_n]	int8	ND	Input.
out	[dim_0, dim_1, ..., dim_n]	int8	ND	Output.

ELEWISE_ADD (Element-wise Addition)

Parameter	Dimension	Data Type	Format	Description
x	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	float/float16/bf16/int32/int64	ND	Input 1.
y	[y_dim_0, ..., y_dim_n]	Same as x	ND	Input 2.
out	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	Same as x	ND	Output. The data types of output, input 1, and input 2 must be the same.

ELEWISE_MUL (Element-wise Vector-to-Vector Multiplication)

Parameter	Dimension	Data Type	Format	Description
x	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	float16/bf16	ND	Input 1.
y	[y_dim_0, ..., y_dim_n]	Same as x	ND	Input 2.
out	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	Same as x	ND	Output. The data types of output, input 1, and input 2 must be the same.

ELEWISE_REALDIV (Element-wise Vector-to-Vector Division)

Parameter	Dimension	Data Type	Format	Description
x	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	float/float16/bf16	ND	Input 1.
y	[y_dim_0, ..., y_dim_n]	Same as x	ND	Input 2.
out	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	Same as x	ND	Output. The data types of output, input 1, and input 2 must be the same.

ELEWISE_LOGICAL_AND (Element-wise Logical AND)

Parameter	Dimension	Data Type	Format	Description
x	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	int8	ND	Input 1.
y	[y_dim_0, ..., y_dim_n]	int8	ND	Input 2. The element values of input 1 and input 2 can only be 0 or 1.
out	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	int8	ND	Output.

ELEWISE_LOGICAL_OR (Element-wise Logical OR)

Parameter	Dimension	Data Type	Format	Description
x	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	int8	ND	Input 1.
y	[y_dim_0, ..., y_dim_n]	int8	ND	Input 2. The element values of input 1 and input 2 can only be 0 or 1.
out	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	int8	ND	Output.

ELEWISE_LESS (Element-wise Less Than Comparison)

Parameter	Dimension	Data Type	Format	Description
x	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	int64/float/float16/bf16	ND	Input 1.
y	[y_dim_0, ..., y_dim_n]	Same as x	ND	Input 2. The data types of input 1 and input 2 must be the same.
out	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	int8	ND	Output.

ELEWISE_GREATER (Element-wise Greater Than Comparison)

Parameter	Dimension	Data Type	Format	Description
x	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	int64/float/float16/bf16	ND	Input 1.
y	[y_dim_0, ..., y_dim_n]	Same as x	ND	Input 2. The data types of input 1 and input 2 must be the same.
out	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	int8	ND	Output.

ELEWISE_SUB (Element-wise Subtraction)

Parameter	Dimension	Data Type	Format	Description
x	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	float16/int64/bf16	ND	Input 1.
y	[y_dim_0, ..., y_dim_n]	Same as x	ND	Input 2. The data types of input 1 and input 2 must be the same.
out	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	Same as x	ND	Output.

ELEWISE_EQUAL (Element-wise Equal Comparison)

Parameter	Dimension	Data Type	Format	Description
x	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	float/float16/bf16	ND	Input 1.
y	[y_dim_0, ..., y_dim_n]	Same as x	ND	Input 2. The data types of input 1 and input 2 must be the same.
out	[x_dim_0, x_dim_1, ..., x_dim_n, y_dim_0, ..., y_dim_n]	int8	ND	Output.

ELEWISE_QUANT_PER_CHANNEL (Quantization per Channel)

Parameter	Dimension	Data Type	Format	Description
x	[x_dim_0, x_dim_1, ..., x_dim_n, dim_0, ..., dim_n]	float16/bf16	ND	Input 1.
scale	[dim_0, ..., dim_n]	float16/bf16	ND	Input 2. The element in scale cannot be 0. The data types of input 1 and input 2 must be the same.
offset	[dim_0, ..., dim_n]	int8	ND	Input 3. offset can be an empty tensor.
y	[x_dim_0, x_dim_1, ..., x_dim_n, dim_0, ..., dim_n]	int8	ND	Quantization output result

ELEWISE_DEQUANT_PER_CHANNEL (Dequantization per Channel)

Parameter	Dimension	Data Type	Format	Description
y	[x_dim_0, x_dim_1, ..., x_dim_n, dim_0, ..., dim_n]	int8	ND	Input vector 1.
scale	[dim_0, ..., dim_n]	float16	ND	Input vector 2.
offset	[dim_0, ..., dim_n]	int8	ND	Input vector 3. offset can be an empty tensor. The data types of input 1 and input 3 must be the same.
x	[x_dim_0, x_dim_1, ..., x_dim_n, dim_0, ..., dim_n]	float16	ND	Dequantization output result.

ELEWISE_DYNAMIC_QUANT (Element-wise Dynamic Quantization)

Parameter	Dimension	Data Type	Format	Description
x	[dim_0, dim_1, ..., dim_n]	float16/bf16	ND	Input vector, where n is greater than 1.
z	[dim_0, dim_1, ..., dim_n]	int8	ND	Quantization output result
scale	[dim_0, ..., dim_n-1]	float	ND	Output, scale required for dequantization.
offset	[dim_0, ..., dim_n-1]	float	ND	Output, offset required for dequantization. This parameter is returned only when asymmetric is set to true. Currently, asymmetric cannot set to true.

ELEWISE_TANH (Element-wise Hyperbolic Tangent Computation)

Parameter	Dimension	Data Type	Format	Description
x	[dim_0, ..., dim_n]	float16	ND	Input.
out	[dim_0, ..., dim_n]	float16	ND	Output.

Parent topic: ElewiseOperation