A method and computing system based on gpgpu reconfigurable architecture

Abstract

The invention discloses a method and a computing system based on a GPGPU reconfigurable architecture. The method includes selecting an architecture mode according to the participation degree of matrix multiplication or convolution operation in the calculation, wherein the architecture mode includes a general-purpose graphics processor mode and Hybrid systolic array mode; when the general-purpose graphics processor mode is selected, the registers in the computing system are configured to perform data access and operation in the general-purpose matrix multiplication mode of the general-purpose graphics processor; when the hybrid systolic array mode is selected, the computing system The registers are configured to perform data access and operations in a direct convolution manner with a hybrid systolic array. The computing system includes registers, and the architectural pattern of the registers is configured using the method described above. The method and computing system based on the GPGPU reconfigurable architecture disclosed by the present invention can be flexibly configured according to the needs of different scenarios, which can not only meet the general requirements, but also make up for the shortcomings of the lack of flexibility of special accelerators.

Description

technical field [0001] The present invention relates to the computing field, and more specifically to a method and a computing system based on a GPGPU reconfigurable architecture. Background technique [0002] Deep neural network (DNN) is the basis of current artificial intelligence (AI) applications, such as: speech recognition, image recognition, video processing, automatic driving, cancer detection, game confrontation, etc., and in some fields, DNN has surpassed human accuracy sex. [0003] figure 1 shows a relatively simple neural network calculation process: the neurons in the input layer pass the received information to the middle layer of the neural network (also known as the hidden layer), and the "weight sum" of the hidden layer is finally transmitted to the output layer, and finally presented to the user. The input or output of a neuron is usually called "activation", and the synapse is called "weight". In the present invention, including the specification and ...

AI Technical Summary

Problems solved by technology

[0064] 1. Reshape and Duplicate are performed on Tensor during the calculation of convolution to matrix, which will increase bandwidth consumption;

[0065] 2. The execution in the time domain brings repeated data loading, which is not as high as the energy efficiency ratio of the air domain execution of the systolic array

[0067] 1. Loss of flexibility. When the neural network model is continuously updated and iterated, it cannot flexibly support all GEMM acceleration algorithms, cannot support general-purpose operations other than matrix operations, and narrow application scenarios will lead to reduced market demand.

[0068] 2. There are still certain limitations here, that is, they only use the accumulation characteristics of the airspace, but completely abandon the time-domain accumulation characteristics of GPGPU, so that the convolution still cannot bypass the matrix operation

Images