Compression method and compression system

Abstract

The present invention provides a compression method, comprising the following steps of sequentially segmenting uncompressed data from the first bit to the last bit into m continuous uncompressed sub-data according to the segmentation byte threshold, so that each of the uncompressed sub-data It has an independent address, so that the uncompressed sub-data does not need to be arranged in a one-dimensional linear arrangement, thereby supporting the compression of data of any dimension. After removing the zero value, the compressed sub-data, flag bit data and number information are obtained. The flag Bit data can make compression less affected by zero value arrangement, improve the compression rate and the stability of compression rate, calculate address offset, calculate compressed sub-data storage address, store compressed sub-data, and calculate flag bit data storage address Storing the flag data, calculating the compression information storage address, storing the compression information, and storing the address offset without storing additional address information can effectively reduce the compression loss rate. The present invention also provides a compression system for realizing the compression method.

Description

technical field [0001] The invention relates to the technical field of convolutional neural networks, in particular to a compression method and a compression system. Background technique [0002] Deep neural network algorithms are widely used in artificial intelligence applications such as computer vision and speech recognition, and convolutional neural network is a very important algorithm in deep neural network algorithms, but due to the bandwidth requirements of convolutional neural network algorithms for accessing memory High, which limits the development of convolutional neural network accelerators. [0003] MIT (Massachusetts Institute of Technology, MIT) proposed a data compression method RLC (Run-LenthCompression, RLC) which is common to weight and intermediate feature data in the Eyriess convolutional neural network accelerator, but this compression method can only dynamically compress one Two-dimensional linear arrangement data cannot be applied to the compression...

AI Technical Summary

Problems solved by technology

[0002] Deep neural network algorithms are widely used in artificial intelligence applications such as computer vision and speech recognition, and convolutional neural network is a very important algorithm in deep neural network algorithms, but due to the bandwidth requirements of convolutional neural network algorithms for accessing memory High, limiting the development of convolutional neural network accelerators

[0003] MIT (Massachusetts Institute of Technology, MIT) proposed a data compression method RLC (Run-LenthCompression, RLC) which is common to weight and intermediate feature data in the Eyriess convolutional neural network accelerator, but this compression method can only dynamically compress one Two-dimensional linear arrangement data is not suitable for the compression of two-dimensional and three-dimensional arrangement feature data, and the compression rate is related to the arrangement of zero values, resulting in an unstable compression rate

[0004] In the Nvidia Deep Learning Accelerator (NVDLA) launched by Nvidia (NVIDIA), feature addresses can be arranged in two-dimensional and three-dimensional formats, but only statically compress weights, and cannot compress a large amount of intermediate feature data. Low, so that the bandwidth cannot be effectively saved, resulting in a high compression loss rate

Images