Convolutional neural network multinuclear parallel computing method facing GPDSP

Abstract

The invention discloses a convolutional neural network multinuclear parallel computing method facing GPDSP. The method comprises the following steps that S1, two data cache regions and a weight data cache region are constructed by a CPU core in an off-chip memory; S2, convolution kernel data of the designated number is combined by the CPU core for processing and stored in the weight data cache region; S3, the CPU core accesses to-be-calculated image data of the designated amplitude for combination processing, and the data is transmitted to a free data cache region; S4, if DSP cores are leisure, and data ready of the data cache regions is achieved, an address is transmitted to the DSP cores; S5, all the DSP cores conduct convolution neural network calculation; S6, a calculation result of the current time is output; S7, the steps of S3-S6 are repeated until all the calculations are completed. The performance and multi-level parallelism of the CPU core and the DSP cores in the GPDSP can be fully exerted, and efficient convolutional neural network calculation is achieved.

Description

technical field [0001] The present invention relates to the technical field of deep learning, in particular to a convolutional neural network multi-core parallel computing method oriented to GPDSP (General-Purpose Digital Signal Processor). Background technique [0002] At present, the deep learning model based on Convolutional Neural Networks (CNN) has made remarkable achievements in various aspects such as image recognition and classification, machine translation, automatic text processing, speech recognition, automatic driving, and video analysis. become a research hotspot in various fields. Convolutional neural network is a deep feed-forward neural network, which is usually composed of several convolutional layers, activation layers and pooling layers alternately. The convolutional layer performs feature extraction through the convolution operation of the convolution kernel and the input features, so that Learn the characteristics of each level. In the calculation of c...

AI Technical Summary

Problems solved by technology

The powerful computing power of GPDSP makes it possible to become a very good platform for accelerating convolutional neural network calculations. However, GPDSP is a heterogeneous multi-core processor that includes CPU cores and DSP cores, including register files, scalar memories, and on-chip vector arrays. Memory, on-chip shared memory array, off-chip DDR memory and other multi-level storage architectures, the existing convolutional neural network computing methods cannot be directly applied to GPDSP.

To implement convolutional neural network calculations by GPDSP, there are still problems such as how to map convolutional neural network calculations to GPDSP's CPU core and multiple DSP cores with 64-bit vector processing arrays, and how to utilize the multi-level parallelism of GPDSP. There is no effective solution for convolutional neural network computing based on GPDSP, and it is urgent to provide a GPDSP-oriented convolutional neural network multi-core parallel computing method to take advantage of the structural features and multi-level parallelism of GPDSP to improve the performance of convolutional neural networks. Network Computational Efficiency

Images