Convolutional hardware accelerator based on RS (Reed-Solomon) data stream and method thereof

Abstract

The invention discloses a convolution hardware accelerator based on RS (Reed-Solomon) data stream and a method thereof. The accelerator comprises an off-chip DDR (Double Data Rate) memory; an on-chip cache module is used for storing the original image data, the convolution kernel weight data and the convolution calculation intermediate result read from the off-chip DDR memory; a DDR controller is used for controlling data interaction between the off-chip DDR memory and the on-chip cache module; a data distribution module is used for transmitting the required weight and excitation data to the photoelectric calculation module according to the current calculation configuration information and the calculation rule of the photoelectric calculation module; the photoelectric calculation module is used for completing convolution operation of each layer of network; and the result collection module is used for receiving the convolution results of the photoelectric calculation module and splicing the results to obtain the input of the next layer, or storing the splicing results in an off-chip DDR memory. According to the method, convolution operation is carried out based on an RS data stream mode, data reuse is maximized, access to off-chip DDR is reduced, and energy efficiency is improved.

Description

technical field [0001] The invention relates to a row-fixed (RS) data flow-based convolution hardware accelerator architecture and a method thereof, and belongs to the technical field of hardware accelerated neural networks. Background technique [0002] In recent years, with the development of computer science and Internet technology, the scale of world data has exploded, and artificial intelligence has also transformed from early artificial feature engineering to learning from large amounts of data. Inspired by brain neuroscience, the neural network model formed after many years of evolution has achieved great results in the field of machine learning, and has been widely used in computer vision, speech recognition, and natural language processing. [0003] In order to deal with complex problems or improve the accuracy of the model, the scale of the neural network model is getting larger and larger, resulting in the need for a large amount of computing resources. Large-sca...

AI Technical Summary

Problems solved by technology

Although the convolutional neural network model can be deployed on a graphics processor GPU with a high degree of parallelism for training or reasoning, which can have a great speed advantage, due to factors such as its power consumption and cost, it is difficult to implement in the actual application of neural network algorithms. There are still many restrictions on deployment

Images