Research On Heterogeneous OpenCL Code Generation And Optimization Methods For Many-core Accelerators

In recent years,the processor performance has been limited,because of its increasing power consumption,interconnect delay and design complexity of the increasing complexity.The traditional single-core architecture has difficult to meet the market demand for performance.With the tremendous development of the integrated circuit technology,it is possible to integrate multiple processor cores on a single chip to accomplish more complex,large computational tasks,and the processor achieves have a transition from single-core to multi-core and many-core.However,there is also a bottleneck in improving performance by means of increasing the cores of same type processors.The number of cores in the CPU will no longer be able to increase performance by increasing the number of cores after reaching the limit.To further enhance the computing power,there has been a trend towards heterogeneous system architecture.Due to the underlying heterogeneous and multi-level storage level,heterogeneous systems programming difficulties is one of the bottlenecks restricting the development of heterogeneous systems,it has become the industry consensus.Therefore,OpenCL as the first heterogeneous parallel programming framework has win more and more people's favor,as an open standard for heterogeneous computing,OpenCL has attracted strong support of many manufacturers,it provides a free,open general standard for the heterogeneous system.To achieve the autonomy of the core processor,the Sunway TaihuLight supercomputer equipped with the domestic on-chip heterogeneous many-core processor SW26010,which was developed independenly by China.To reduce the programmers' difficulty of programming and improve the portability of the software,this paper designs and implements an OpenCL compilation system that supports the domestic many-core processor SW26010,and studies the OpenCL optimization method.The innovations of this paper are as follows:(1)Based on the OpenCL programming framework and the microstructure of the homegrown many-core processors,the paper proposes the mapping relationships between OpenCL platform model,memory model and execution model to SW26010 many-core processors.(2)In view of the hardware structure,this paper puts forward the OpenCL optimization methods for many-core processors,such as thread coarsening and data layout.