| With the development of multi-core processors, multi-thread is becoming the bottleneck of the performance of a program. Since a large amount of applications running on embedded multi-core processors are multimedia decoding programs, this thesis concentrates on the parallelization of these programs.The parallelization method consists of four stages:1)analyzing parallel area,2)making parallel plan,3)code generation, and4)runtime management. Among the four stages, we pay most attention to the stage of analyzing parallel area, which consists of program structure analysis and parallelization area detection.The proposed program structure analysis method combines static and dynamic analyses. First, it preprocesses the source code, then makes dynamic analysis and puts the cost into the source code as comments. Then it makes static analysis of the intermediate file and builds the program call graph. The nodes in the program call graph represents functions and loops, edges represents the call relations between the functions and loops, and the weight of the nodes represents the cost of functions and loops. The program call graph is used as a reference for the later parallel area detection.The proposed parallel area detection mechanism combines multi-grain parallel detection. For data parallellism, it detects the continuous write or read operations on the continuous memory addresses. For task parallellism, it detects the read and write dependences between functions. For pipeline parallellism, it detects the dependences between functions in a loop. Our method is based on dynamic analysis and can avoid the conservation in static analysis.We use APE and MP3decoding programs as case studies. We assess them on two and four-core platform, which get speedups of7.28and3.97, and power consumption ratios of0.29and0.47, respectively. The experienmental results show that our method can improve the perforemance of system and power consumption respectively and justify our method effectiveness. |
PDF Full Text Request