Research And Application Of Parallel Optimization Methods For Multi-core Vector Processors

It has long been a trend that multi-core and vectorization are widely equipped in modern processors.To fully utilize these features,the methodology of parallel programming is indispensable.Thus,this research would mainly focus on exploring the parallelization and optimization methods for multi-core vector processor architectures,considering the similar multi core and vector characteristics dwelled in CPUs and GPUs.My thesis starts by investigating the similarity among multi-core vector processors.Then this research deduces the parallel programming patterns for multi-core vector processors from the general one,which includes the programming patterns for Problem Space,Algorithm Structure,Program Structure and Implementation Mechanism.After that,a recipe for parallelizing codes is presented,consisting of Problem Analyzing,Algorithm Structure Design,Program Structure Design and Implementation.In the Program Design step,I propose setting loop as a bridge between multi-core vector processors and code,provide the methods to transform other program structure to loop,and discuss the issues of converting loop into multicore-vectorized code.Besides,this thesis also proposed two optimization methods,i.e.storage pattern optimization and loop regularizing optimization.The methodology studied is then put into operation in two representative application fields,including Digital Image Correlation(DIC)and Genomic Alignment.The algorithm of Pathindependent Digital Image Correlation(PiDIC)proposed by Zhenyu Jiang and BWA-MEM are studied,and the algorithm of cuDIC and vecMEM are designed and implemented.The finegrained regular data parallel schema in PiDIC and the nearly regular process pattern in Smith Waterman operation repeatedly called in BWA-MEM implies the feasibility of applying multicore vector parallelization and optimization.In the Algorithm Structure Design step,I harness a na?ve while efficient structure for DIC optimization,and devise an algorithm structure maintaining the original pipeline pattern in MEM for the development of vecMEM.In the Program Structure Design step,changes are made center on loop.Then,I implement the cuDIC in GPU while vecMEM in CPU.After the implementation,the storage pattern optimization is applied,and the loop regularizing optimization is further leveraged in vecMEM for it's sophisticated while predictable processing pattern.Experimental study shows that the parallelization and optimization methodology proposed could improve the performance of the program to some extend.To be specific,cu DIC achieves performance improvement over it's sequential counterpart of PiDIC as much as 19.62 times,and vecMEM's performance is 30.6% times higher than multi-thread optimized BWA-MEM in 100 bp genomic alignment task.The result of the experiments proves the effectiveness of the parallelization and optimization methods proposed.