Research On Heterogeneous System Accelerated Computing Based On OpenCL

Due to factors such as bandwidth,heat dissipation,power consumption,and cost,the performance growth curve of general-purpose processors gradually slows down.With the complexity of computing tasks and scale upgrades,the performance of generalpurpose processors has reached a bottleneck.Facing the performance bottleneck of the CPU,customizing a dedicated processor for a specific field has become an efficient solution.Computing platforms built with processors of different system architectures are called heterogeneous systems.In the face of performance issues and power consumption issues,the heterogeneous system breaks through the limitation of a single architecture,providing users with better scalability and applicability.The heterogeneous parallel computing model is the most critical basic software on a heterogeneous system.It connects applications and hardware resources and directly affects performance.OpenCL is an open source,free,unified heterogeneous parallel computing framework,Because the abstract model of OpenCL is applicable to different platforms,it is supported by many hardware manufacturers.In the OpenCL development environment,users can uniformly control various processors under different platforms,and perform parallel computing at various levels.Therefore,based on the OpenCL framework,it has application value to study the performance of heterogeneous systems.Based on the OpenCL parallel framework,this thesis aims at the high performance of heterogeneous systems to carry out research on computing acceleration of heterogeneous systems.The main work is divided into two parts:(1)Single-kernel multi-devices computing mode based on code analysis and conversion.This mode analyzes and converts the source code of the OpenCL kernel,and distributes the workload of the kernel to multiple devices to realize the calculation acceleration of a single kernel in a heterogeneous system.Pattern composition includes:kernel decomposition,kernel global memory access pattern analysis,kernel performance prediction.The kernel decomposition module converts the kernel source code to realize the division of computing load and the merging of output data;The kernel global memory access mode analysis module analyzes the memory access instructions of the source code,and optimizes the transmission volume of the global memory between devices;The Kernel Performance Prediction module analyzes different ways of predicting kernel execution times.(2)Multi-kernels heterogeneous scheduling algorithm under the OpenCL parallel framework.Taking the heterogeneous system of the HunXin Digtal Signal Processor as a case,explore the execution mode of OpenCL multi-kernels on a multi-devices system.The content includes:analyzing the parallelism between OpenCL kernels,and proposing two ways to obtain OpenCL task graphs:the task graph extraction method based on command queue,and the task graph extraction method based on abstract syntax tree;Based on the traditional heterogeneous task scheduling algorithm HEFT,combined with heterogeneous system symmetry and kernel parallel computing characteristics,the HDGEFT scheduling algorithm is proposed.Starting from the above work content,this thesis designs experiments to verify the behavior of the acceleration scheme on heterogeneous systems.Through the comparison of experimental data,the efficiency of computing acceleration of heterogeneous systems under the framework of OpenCL is verified.