Hardware Acceleration Implementation Method of RNN Forward Propagation Model Based on Transversal Systolic Array

Abstract

The invention discloses a system and method for realizing hardware acceleration of an RNN forward propagation model based on a transverse pulsation array. First, the network parameters are configured, and after the data is initialized, based on the transverse pulsation array, the weight in the calculation adopts block design, and the hidden layer calculation The weight matrix is ​​divided into blocks by rows, and the hidden layer neurons are calculated through matrix multiplication vector and vector summation operation and activation function operation, and then according to the obtained hidden layer neurons, matrix multiplication vector, vector summation operation and activation function Operations to generate the results of the RNN output layer, and finally configure the information according to the length of the time series to generate the output results required by the RNN network. In this method, the hidden layer and the output layer are multi-dimensionally parallel, which improves the flow of calculation. At the same time, the weight matrix parameters in the RNN network The shared feature adopts block design to further improve the parallelism of computing, high flexibility, strong scalability, high utilization of storage resources, and high speed-up ratio, which greatly reduces computing.

Description

technical field [0001] The invention belongs to the technical field of artificial intelligence hardware acceleration, and in particular relates to a system and method for realizing hardware acceleration of an RNN forward propagation model based on a transverse pulsation array. Background technique [0002] In the era of artificial intelligence, neural network algorithms have been widely used in different fields such as image recognition, such as convolutional neural networks (Convolutional Neural Networks, CNN), deep neural networks (Deep Neural Networks, DNN), in such a wide and frequent use , the neural network algorithm also exposed some problems: the traditional neural network can only process one input alone, and the input at the previous and subsequent moments is completely irrelevant. For example, when we predict the next word of a sentence, the preceding and following words in the sentence are not independent, and CNN or DNN can't do anything about it, while Recurren...

AI Technical Summary

Problems solved by technology

[0002] In the era of artificial intelligence, neural network algorithms have been widely used in different fields such as image recognition, such as convolutional neural networks (Convolutional Neural Networks, CNN), deep neural networks (Deep Neural Networks, DNN), in such a wide and frequent use , the neural network algorithm also exposed some problems: the traditional neural network can only process one input alone, and the input at the previous and subsequent moments is completely irrelevant

[0003] The input layer, hidden layer, and output layer of RNN are all fully-connected computing modes, including a large number of multiply-accumulate operations, which are computationally intensive algorithms that usually need to be implemented using hardware acceleration methods, but traditional CPUs have fewer logical operation units. It cannot meet the increasingly large computing network. Although GPU has strong computing power, it cannot flexibly configure the hardware structure. As a dedicated processing chip, ASIC has low power consumption, small area, high performance and poor flexibility and cannot be configured for specific needs.

Images