基于特征向量匹配的CUDA程序PTX代码提取

将CUDA程序移植到其它异构众核处理器平台,对扩展CUDA程序应用范围并发挥目标平台的众核优势具有现实意义。采用二进制翻译技术移植CUDA程序时,PTX代码的识别和提取是关键。从软件逆向角度,通过静态分析方法评价了PTX代码提取的可行性,抽象出PTX提取模型,设计了基于特征向量匹配PTX代码提取算法。实验结果表明,这种提取方法成功率接近100%。

PTX Code Extraction in CUDA Program Based on Characteristic Vector Matching

Porting CUDA program to other heterogeneous and many core target platforms is significant for enriching CUDA application and taking advantage of many core property. When porting the CUDA program with the technology of binary translation, the identification and extraction of PTX code is the key step. On the perspective of software converse, the feasibility of the PTX code extraction is evaluated based on the static analysis of CUDA binary code. A model is proposed for PTX code extraction and the corresponding algorithm is designed based on characteristic vector matching. The experimental result validates the correctness of the algorithm.