Research On Parallel Acceleration Technology Of Heterogeneous Multi-core Platform Oriented Image Algorithm

With the rapid growth of computing power demand for computer applications such as artificial intelligence and big data,and the increasing diversification of application scenarios,the processor architecture is gradually diversified.The Central Processing Unit(CPU)architecture has evolved from a traditional single core to a multi-core parallel direction.The Graphics Processing Unit(GPU)has also evolved from a dedicated image rendering device to a general-purpose parallel computing device,mobile GPU and accelerators dedicated for artificial intelligence operations have also been developed and applied.Computing-intensive applications and diversified processor architectures have given birth to heterogeneous multi-core computer systems.The main control chip of embedded devices has begun to integrate multiple computing resources to improve computing capabilities,such as coprocessors,GPUs,and AI-specific accelerators.The collaborative parallel computing of heterogeneous multi-core platforms with multiple computing resources has gradually become a trend.Single heterogeneous acceleration methods such as GPU or FPGA are no longer suitable for embedded devices that integrate multiple computing resources.Therefore,it is urgent to study how to design parallel algorithms in the main control chip containing multiple computing resources in accordance with the hardware structure of the parallel processor and the difference in algorithm computing and memory access,and how to better coordinate multiple hardware parallel computing to more efficiently give full play to the computing power of the hardware.For example,in autonomous driving scenarios,target detection network carry functions such as extracting features and detecting pedestrians or vehicles to monitor road conditions,which are indispensable in the field of autonomous driving.However,autonomous driving scenes are mostly outdoor mobile scenes,and there are various environmental factors such as light intensity,rain and fog weather,and fuzzy occlusion,which affect the imaging quality of imaging equipment.As a result,the detection effect of target detection algorithms in automatic driving scenes is not good in practical applications ideally.In order to improve the accuracy of the algorithm,in addition to improving the accuracy of the model by improving the network structure and improving the network training strategy,it is necessary to perform preprocessing such as noise reduction and defogging on the input image of the detection and recognition algorithm.This paper is oriented to complex image processing and target detection tasks in autonomous driving scenarios.Aiming at computing resources such as ARM multi-core CPUs,Mali GPUs and neural network accelerators on heterogeneous multi-core chips on mobile equipment,this paper designs and implements an efficient heterogeneous collaborative parallelism Program.According to the characteristics of algorithm and computing architecture,the computing task is reasonably divided into multiple computing units,and data communication is realized among the computing units through circular queues,and then the overlapping execution of each computing module in time when multiple input images are realized in the form of pipeline.At the same time,for each computing unit,the parallel acceleration of dark channel defogging module is realized by using NEON vectorization instruction set and OpenMP multi-core technology in CPU,and Mali GPU is called to accelerate the parallel calculation of image non-local mean denoising module through OpenCL heterogeneous parallel framework in GPU,and the accelerated calculation of YOLOv3 target detection network is realized in neural network accelerator.Parallelization of computing tasks is realized at two different levels,and various computing resources on the mobile heterogeneous multi-core platform are fully utilized to accelerate the process of image processing and objection detection.