Arithmetic processor and arithmetic processing system for Point floating-point numbers

Abstract

The invention provides an arithmetic processor and an arithmetic processing system for Poit floating-point numbers, and relates to the technical field of computers. The operations of addition, subtraction, multiplication and FMA are directly carried out on intermediate data in a complement form acquired by a self-decoding circuit; and an operation result represented by the intermediate data is obtained in the complement form. The operation result represented by the intermediate data is directly inputted in the complement form into a coding circuit, so that the coding circuit directly convertsthe intermediate data in the complement form into Poit floating-point numbers. The mutual conversion of original codes and complement codes is prevented from being repeatedly carried out in the operational circuit, logic of the operational circuit is simplified, and therefore the area and power consumption of the operational circuit are reduced.

Description

technical field [0001] The present application relates to the field of computer technology, in particular to a Posit floating-point arithmetic processor and an arithmetic processing system. Background technique [0002] Floating-point numbers are a commonly used data representation method in the field of scientific computing and high-performance computing. In particular, floating-point numbers are widely used for data processing when the accuracy of calculation results is required to be high, such as in the fields of automatic driving, aerospace, and mechanical computing. . [0003] Posit standard floating-point numbers are more flexible than IEEE 754 standard floating-point numbers, and the sampling distribution of Posit floating-point numbers on real numbers is related to the sigmoid function, and the sigmoid function is a commonly used activation function in machine learning, so Using Posit floating-point numbers for machine learning, the efficiency of computer data proc...

AI Technical Summary

Problems solved by technology

[0004] The current method for processing Posit floating-point numbers is usually to use a decoding circuit to convert the Posit floating-point numbers participating in the calculation into the original code, and input the original code to the calculation circuit for calculation. Whether the number of points is a negative number, and then unify the original code of the negative Posit floating point number into the form of complement code, and then convert the operation result of the form of complement code into the original code after obtaining the operation result of the complement code form. The entire operation process needs to be repeated Convert the original code to the complement code, and the complement code to the original code, and it is necessary to judge the positive or negative of the Posit floating-point number at each operation node, which makes the entire operation logic of the Posit floating-point number very complicated and cumbersome, resulting in the circuit of the operation circuit The area is too large

Images