Analysis Of Essential Parts Of Source Codes For Parallel Computing

The development of semiconductor technology offers faster transistors speed, smallerchip area and lower power consumption. Multicore parallel computing platform has takenplace serial computing platform with saturated performance and clock rate. Heterogeneousmulticore processor parallel platforms can manage the specific tasks on heterogeneousapplication specific processors to improve the performance.Application codes running on embedded heterogeneous multicore parallel computingplatforms are implemented as both software running on processors and application specifichardware. In system design, hardware software co-design methodology is thus used to meetthe need of performance and flexibility. At the initial architecture design phase, the impactof hardware and software partition must be considered. For example, function partition,memory subsystem structure, precision trade off on software and hardware. However, thesetasks heavily rely on the experience of designer. This thesis proposes methods and toolstowards quantitative design and has obtained the following achievements:(1) Analysis and identification of program hot spotsDesign a program hotspot analysis tool is based on compiler technology. Weimplemented a hot spot identification tool for quantitative approach of hardware softwarepartitioning, instruction-set acceleration, and stream-kernel code partition. Theimplementation of the hot spots analysis framework is based on the compiler and programintermediate presentation.(2) Memory size estimation and trade offBased on the dynamic profiler techniques, we achieved the information of programmemory access, for example data access frequency and data live-time interval. Wethereafter designed a model and a tool for memory subsystem power consumption. The toolcan provide program feature for memory size estimation and data distribution.(3) Optimum word length analysis for fixed-point processingUsing simulation-based search methods, we implemented a design an algorithm foroptimizing the data precisions, i.e. the word length. The proposed method can dynamically convert data from float-point to fixed-point by optimizing parameters such as fractionalword length, allowed saturation points, and fixed-point word-length through iterations. Thetool can also provide fixed-point processor word-length suggestions.