Research On Some Key Parallel Optimization Technologies For Computer Vision Applications Aiming At The Multicore Platforms

The risen of Multicore and Manycore architecture is a big shift in the history of computer development.When the traditional performance improving method which mainly relies on increasing the frequency of the processor faced with the enormous difficulties, the researchers and industry turned to a different way, viz. increasing the density of the IC unit to improve the performance of the processors continuously to fill the increasing computation power needs required by the latest emerged Tera-Scale RMS typed applications, including computer vision applications. The rapid development and application of Multicore and Manycore architecture in recent years, on one hand, provides the probability and the hardware basics for implementing these performance hungry applications in real time, whereas, on the other hand, since the potential compute power of the Multicore and Manycore hardware can be fully utilized only by the means of parallel, it also brings out more challenges on parallel optimization for these kind of applications, especially for computer vision applications. And the key challenges for the programmer to achieving the performance indices, is how to best exploit the optimal mapping from potential algorithm parallelism to the available hardware parallelism, considering the complex input dependent computation characters of computer vision applications.This thesis commences from the study of the model of the computer vision applications and the workload characters. After the evaluations of the characters and the available parallel optimization methodology, strategy and technologies, the three key capability requirements and6key data parallelism patterns for best optimize the computer vision applications are identified and summarized; based on this acknowledgement, the thesis proposes that under the guidance of domain knowledge, a dynamic-static integrated parallel optimization scheme which combined both pre-knowledge and the real timed workload feature is an ideal solution to the problem. This thesis gives a thoroughly discussion aiming at the key optimize technologies related to the above solution.Firstly, for the parallel abstract and representation topic, this thesis introduces an extended concept of composition relationship and weighted character, brings out an improved TStreams parallel computation model to express the various parallelism lies in the algorithm, and at the same time, act as the basics for the further stage parallel optimization.Secondly, for the parallel exploit topic, a DSL oriented to the computer vision applications is employed to describe the available parallel optimization rules. The optimization rules is an abstraction from the domain knowledge, gives the various transform of specified functions. The optimization rule description for the whole application consists of an overall solution space for the parallel optimization. With the compiler-driven autotuner which is introduced by the thesis, the solution space for the parallel optimization can be constructed, and the candidate solutions can be searched in a semi-automatic way. In such a way, the complicated and difficult parallel task can be converted into the description task for the algorithm and hardware’s inherent parallelism characters. It largely reduced the difficulty of the parallel optimization and improved the efficency of parallelization greately.Thirdly, for dynamic task scheduling topic, a concept of "character weight aware Runtime" is proposed by this thesis. It enables the runtime system to perform real timed low level optimization depending on the runtime information, so as to get a better localized, balanced dynamic scheduling with large degree of parallelism.At last, by synthesized the above research fruits, this thesis bring out a semi-automatic parallel optimization framework aiming for the computer vision application on Manycore and Multicore platforms. And the application it into the optimization for vehicle recognition algorithm (adopted in Drive Assistant System) has achieved good result in parallel optimization.