A data processing apparatus and method based on an NPU processor

By introducing preprocessing and Taylor expansion calculations into the NPU processor, the problems of low efficiency of the CORDIC algorithm and insufficient accuracy of the Taylor expansion are solved, and efficient and high-precision calculations of trigonometric functions are achieved.

CN117688282BActive Publication Date: 2026-07-07STREAM COMPUTING INC
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
STREAM COMPUTING INC
Filing Date
2022-09-09
Publication Date
2026-07-07

AI Technical Summary

Technical Problem

When processing trigonometric functions, existing NPU processors suffer from low computational efficiency due to the CORDIC algorithm, while the Taylor expansion approximation algorithm has a limited accuracy range, making it difficult to meet the requirements for efficient and high-precision computation.

Method used

A data processing device and method based on an NPU processor is adopted. The input value is transformed into a target input value and an auxiliary value through a preprocessing unit, and the first and second calculated values ​​are calculated using Taylor expansion. The output unit determines the target trigonometric function value according to the magnitude of the input value and improves the calculation accuracy by combining the auxiliary value.

Benefits of technology

While meeting the preset accuracy requirements, it improves the processing speed and efficiency of trigonometric function operations, and has higher data processing efficiency compared to the CORDIC algorithm.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117688282B_ABST
    Figure CN117688282B_ABST
Patent Text Reader

Abstract

Embodiments of the present application disclose a data processing device and method based on an NPU processor. After determining a target input value, the device and method introduce an auxiliary value, obtain a first calculation value corresponding to the target input value and a second calculation value corresponding to the auxiliary value based on a Taylor expansion, and finally determine a target trigonometric function value from the first calculation value and the second calculation value according to the size of the target input value. Thus, when running the Taylor expansion approximation algorithm in the NPU processor, the accuracy of the value range of the part of the trigonometric function that is difficult to ensure is ensured by processing the auxiliary value corresponding to the target input value.
Need to check novelty before this filing date? Find Prior Art