SignBitsPack

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: extracts the sign bits of floating-point numbers and packs them into uint8 bytes. The sign bits of every 8 floating-point numbers (0 indicates non-negative, and 1 indicates negative) are packed into 1 uint8 byte, where the sign bit of the 1st floating-point number corresponds to bit7 and that of the 8th corresponds to bit0. It is mainly used in the 1-bit Adam optimizer scenario to compress and store sign information.
• Computation formula:

$$b_i=\{\begin{array}{ll}0& x_i\ge 0\\ 1& x_i<0\end{array}$$

$$y_j=\sum _{k=0}^7{b}_{8j+k}\cdot 2^{7-k}$$

$$\text{out}=\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{h}\mathrm{a}\mathrm{p}\mathrm{e}\left(\{y_j\},(\text{size},L/\text{size})\right)$$

Here, $L=\lceil N/8\rceil$, and $N$ is the number of input elements. When $N$ is not a multiple of 8, the tail is padded with $-1.0$ (the sign bit is 1).

Parameter Description

Parameter Name	Input/Output/Attribute	Description	Data Type	Data Format
x	Input	Floating-point tensor to be packed, 1D. The sign bit of each element is extracted. Non-contiguous tensors are automatically converted to contiguous.	FLOAT16/FLOAT	ND
size	Required attribute	Size of the first dimension of the output. The value must be a positive integer (size ≥ 1). ceil(N/8) % size == 0 must be satisfied.	INT64	-
y	Output	Sign bit packing result, 2D. Its shape is [size, ceil(N/8)/size], and its dtype is always uint8.	UINT8	ND

Constraint Description

• Only the ND data format is supported.
• The input x must be 1D (rank=1). Scalars or multi-dimensional tensors are not supported.
• The type combination is fixed to FLOAT16→UINT8 or FLOAT→UINT8, and other dtype combinations are not supported.
• The attribute size must be a positive integer (size ≥ 1), and ceil(N/8) must be integer-divisible by size.
• Empty tensors (N=0) are supported, and an empty output with the shape [size, 0] is returned directly.
• Non-contiguous tensors are supported, and the framework automatically converts them to contiguous before execution.
• Both +0 and -0 are considered non-negative (the sign bit is 0). NaN is not supported, and its behavior is undefined.
• Deterministic implementation by default. The same input always produces the same output.

Invocation Description

Invocation Method	Invocation Sample	Description
Single-operator invocation	test_aclnn_sign_bits_pack	Invokes the SignBitsPack operator through the aclnn API.
Graph-mode invocation	test_geir_sign_bits_pack	Invokes the SignBitsPack operator by building a graph through GE-IR.