Fixed-point method and system for floating-point operation

A technology of floating-point calculation and floating-point data, which is applied in neural learning methods, calculations using number system representations, and biological neural network models. It can solve the problems that neural networks cannot meet computing performance requirements, and achieve the effect of improving performance.

Active Publication Date: 2017-12-08
HANGZHOU FEISHU TECH CO LTD
View PDF6 Cites 72 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0007] The present invention aims at the disadvantage that the neural network using standard floating-point data operations in the p

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Fixed-point method and system for floating-point operation
  • Fixed-point method and system for floating-point operation
  • Fixed-point method and system for floating-point operation

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0056] A fixed-point method for floating-point operations, comprising the following steps:

[0057] S1. Adjust the exponent and significant digit of the adjustable precision floating point data according to the preset numerical range and calculation accuracy requirements, and convert the parameter data in single precision floating point format into adjustable precision floating point data in advance;

[0058] S2. During calculation, convert the adjustable-precision floating-point data into fixed-point data and use it for calculation, and convert the intermediate results generated by the calculation into adjustable-precision floating-point data;

[0059] S3. After all calculations are completed, a final result is generated, and the final result is converted into data in a single-precision floating-point format.

[0060] In a general neural network system, data in single-precision floating-point format is used for calculations. The network model parameter data involved in the ca...

specific example

[0092] In the process of converting single-precision floating-point numbers into adjustable-precision floating-point data, the operation of multiplying the coefficient k is used, so the actual fixed-point operation is performed with k times the original value. When two adjustable-precision floating-point data After a multiplication operation, the actual result is the square of k of the original result. In the FPGA internal data operation process, all values ​​​​need to be calculated according to a unified multiple, so the result after the multiplication operation needs to be divided by k before subsequent operations. .

[0093] Figure 4 It is a schematic diagram of the fixed-point operation process. The most commonly used operations in the neural network are multiplication and addition. Therefore, taking Y=A×B+C as an example, the data processing and operation process described above are explained. The data are expressed in decimal. It can be seen that after one multiplicati...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

PUM

No PUM Login to view more

Abstract

The invention discloses a fixed-point method for a floating-point operation. The method comprises the steps that according to a preset numerical range and the calculation precision requirement, the exponents and significant bits of adjustable-precision floating-point data are adjusted, and the parameter data in a single-precision floating-point format are converted into adjustable-precision floating-point data; when calculating is carried out, the adjustable-precision floating-point data are converted into fixed-point data and used for calculation, and an intermediate result generated by calculation is converted into adjustable-precision floating-point data; a final result is generated after all calculation; and the final result is converted to data in the single-precision floating-point format. According to the invention, standard floating-point data can be represented with a 16-bit bit width or a smaller bit width to realize fixed-point conversion of the floating-point operation; the storage space is saved; the operation performance is improved; and the calculation precision is not influenced.

Description

technical field [0001] The invention relates to a programmable processor, in particular to a fixed-point floating-point calculation method and system. Background technique [0002] At present, due to the continuous advancement of deep learning algorithms, artificial intelligence technology has made great breakthroughs. In deep learning, a machine can learn a task from large amounts of data in a supervised or unsupervised manner. Large-scale supervised learning has achieved success in applications such as machine vision, speech recognition, and natural language processing. The artificial neural network (Artificial Neural Network) is currently the most widely used in the field of deep learning. Using a neural network to identify and classify requires a large number of floating-point operations. Therefore, computing hardware has become the most important bottleneck in the current application of neural networks. Moreover, with the substantial increase in the amount of computin...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

Application Information

Patent Timeline
no application Login to view more
IPC IPC(8): G06N3/063G06N3/08G06F7/38
CPCG06F7/38G06N3/063G06N3/084
Inventor 丁昊杰王文华
Owner HANGZHOU FEISHU TECH CO LTD
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic.

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap