SignBitsUnpack

English | 简体中文

Product Support Status

Product	Supported
Ascend 950PR&950DT products	√
Atlas A3 products	√
Atlas A2 products	√
Atlas 200I/500 A2 inference products	√
Atlas inference products	√
Atlas training products	√

Function Description

• Operator function: unpacks the 1-bit Adam sign bits of the UINT8 type into a tensor of the FLOAT32 or FLOAT16 type.

• Computation formula:

Assume that the input self is a 1D UINT8 tensor with the length of N. Each byte contains 8 sign bits in LSB-first order. The unpacking rule for the $i$-th bit (from the least significant to the most significant) in the $j$-th byte $b_j$ ($i=0,1,\dots ,7$) is as follows:

$$ou{t}_{j\cdot 8+i}=\{\begin{array}{ll}+1.0,& {\text{bit}}_i(b_j)=1\\ -1.0,& {\text{bit}}_i(b_j)=0\end{array}$$

The output tensor has $N\times 8$ elements in total and is reshaped to 2D. Its shape is:

$$out.shape=(size,\frac{N\times 8}{size})$$

Here, size is the size of the first dimension of the output specified by the input parameter.

Parameter Description

Parameter Name	Input/Output/Attribute	Description	Data Type	Data Format
self	Input	Input parameter for the SignBitsUnpack computation. It is a 1D tensor, and each UINT8 element contains 8 sign bits. Empty tensors are supported.	UINT8	ND
size	Input attribute	Indicates dimension processing, which is the first dimension of the output tensor during reshape.	INT64	-
dtype	Input attribute	Indicates the data type of the quantized output tensor.	FLOAT16, FLOAT	-
out	Output	Output parameter for the SignBitsUnpack computation. It is a 2D tensor, and its data type is determined by the dtype parameter.	FLOAT16, FLOAT	ND

Constraint Description

• self must be a 1-dimensional tensor, and out must be a 2-dimensional tensor.
• size must be greater than 0, and (the number of elements in self × 8) must be integer-divisible by size.
• The first dimension of out must be equal to size.
• The dtype parameter supports only FLOAT16 and FLOAT, and the data type of out must be the same as that specified by the dtype parameter. UINT8 is the data type of self, and INT64 is the API type of size. Neither of them is an optional value of the dtype parameter.
• Only the ND data format is supported.
• The empty tensor scenario is supported.
• Deterministic computation: aclnnSignBitsUnpack defaults to a deterministic implementation.

Invocation Description

Invocation Method	Invocation Sample	Description
aclnn invocation	test_aclnn_sign_bits_unpack	Invokes the SignBitsUnpack operator through the aclnnSignBitsUnpack API.
Graph-mode invocation	test_geir_sign_bits_unpack	Invokes the SignBitsUnpack operator by building a graph through the operator IR.