High-Speed Transcendental Function Operation Unit Design
Tian Ze, Fan Feihu, Zhang Jun, Xianglong Ren, Wei Yang · 2022
With the rapid popularization and rapid development of embedded devices such as mobile terminals, people have become increasingly demanding for high-resolution picture rendering effects and fast computing capabilities of various processors. The enhancement of these capabilities of the processor depends on the computer’s floating-point computing capabilities, which makes people to the GPU’s rendering quality and speed, mobile devices, digital signal processing (DSP), and other computing units’ floating-point computing capabilities Higher requirements have been raised [1]. The transcendental function arithmetic unit based on single-precision floating-point numbers is a very important component in these devices. For example, the perspective division, vector normalization, and coordinate transformation required in the GPU all require transcendental function operations [2] (For example, find the inverse, square root, trigonometric function, exponent, logarithm, etc.). In addition to the algebraic operations implemented in the field of mathematics, it is difficult to directly use a computer to calculate some complex functions. It can only complete complex functions through a limited number of simple operations, but for transcendental functions, in other words, the relationship between their input variables and output variables cannot be expressed by a finite number of basic operations such as addition, subtraction, multiplication, division, and power. Therefore, in a computer, the operation of a transcendental function cannot be directly implemented using a limited number of basic function operation units. A special operation unit is required to complete the operation of the transcendental function.