Compute Unified Device Architecture Based On Binary Translation Technology

In recent years, The serial computing architecture of CPU has gradually become the bottleneck for promoting high-performance computing performance. The new heterogeneous architecture combined CPU with GPU has greatly improved the performance of some application program, but software resources which be able to take full advantage of this new architecture currently is less. Therefore, it is a new subject to be studied in depth how to effectively use the new heterogeneous architecture of CPU and GPU to increase the executive performance of traditional serial program and reduce cost and risk of software development.Binary translation technology provides the ability that the executable file of the old architecture can be automatically converted to the executable file of the new architecture without the help of compiler. Generally, binary translation technique is used to solve different architecture software transplant problem. Therefore, this paper uses binary translation technology and the new heterogeneous architecture of CPU and GPU to improve the performance of many traditional serial programs running on CPU. The main contributions of this paper include:Firstly, according to the similarity of anti-compiler technology, compiler technology and binary translation technology, this paper presents a approach using the anti-compiler front-end as the binary translator front-end and and using the compiler back-end as the binary translator back-end a binary translator. This method avoids to rewrite a static binary system, and greatly reducing development time.Secondly, to solve the contradiction between resource rich of the traditional serial program and resource scarcity of new architecture software, this paper proposes a solution compute unified device architecture based on binary translation technology. A traditional serial program running on the CPU is converted to a new binary file running in the CPU and GPU Heterogeneous System by the bCUDA solution.Finally, according to the proposed solutions bCUDA, we use the front-end of a decompiler, the back-end of compiler, PTX back-end and advanced language back-end to achieve a static binary translation system SBTG. Through the matrix multiplication results show that the parallel program which is generated by the static binary translation system SBTG not only properly perform the function of traditional serial program, but also its execution speed faster than the implementation of traditional serial procedures.