A Study. on A Super-Pipelined Ultra-Parallel Processing Mechanism and Its Unified System Description

为了在单个硅片上实现复杂系统的最终集成，该项目的主要目标是建立一种跨越硬件和软件的新的系统开发范例。结果表明，数据驱动原理能够为片上系统的整体设计提供公共基础，该设计涵盖了给定规范的硬件和软件实现。结果表明，软件部分在完全由自定时流水线方案组成的分布在单个芯片上的超流水线动态数据驱动处理器上执行效率最高。本课题取得的成果概括如下：1.提出了超流水线超并行处理机制。该机制利用流水线之间的灵活交互，实现遗传算法中的遗传函数等算法功能和非算法功能，并以数据驱动的多媒体处理器DDMP为核心处理器进行开发。当前版本的DDMP每秒可以执行数百亿次信号处理操作，功耗低至2 W。建立了系统功能、行为和数据结构的自然表示的多边图形化规范方案。在该环境中，表示功能块之间的数据依赖关系的图形化描述的框图被直接转换为保持目标系统的处理结构的可执行数据驱动程序。包含转换的附加基本信息的行为和数据结构也用图表形式表示。例如，增强的序列图、数据结构图、状态转换图和决策表。并将该变换方案应用于复杂视频信号处理等多维流处理中，验证了该变换方案的实用性。

Towards ultimate integration of complex systems on a single silicon wafer, the main objective of this project is to establish a novel system development paradigm spanning both hardware and software. It is demonstrated that data-driven principle is able to provide common basis for integral design of systems on a chip covering both hardware and software implementations of a given specification. Then it is shown that the software part is most effectively executed on super-pipelined dynamic data-driven processors entirely composed of a self-timed pipeline scheme and distributed over a single chip. The results obtained in the project are summarized as following :1. Super-pipelined ultra-parallel processing mechanism was proposed. In this mechanism, flexible interaction among pipelines can be utilized to implement both algorithmic and non-algorithmic functions such as the genetic functions in GA.Furthermore, during the project, the data-driven multimedia processor DDMP was developed as a core processor in the mechanism. The current version of the DDMP can execute tens billions of signal processing operations per second with power consumption as low as 2 W.2. Multilateral diagrammatic specification scheme was established for natural representation of system function, behavior and data-structures. In the environment, a block diagram like graphical descriptions representing data dependencies among functional blocks are directly transformed into executable data-driven programs preserving the processing structures of target systems. Behavior and data structures which contain additional essential information for the transformation are also represented by diagrammatic forms. e.g. augmented sequence charts, data-structure diagrams, state transition diagrams and decision tables. Further, the transformation scheme was applied to multi-dimensional stream processing such as complex video signal processing and its practicability was demonstrated.