Study for Effective Use of Vector Processors Based on Data-Parallel Model

我们从多个角度对数据并行计算模型及其在矢量处理器中的应用进行了基础性的研究。包括以下研究课题：1。基于数据并行模型的并行处理系统我们开发了一种能够高效执行数据并行程序的大规模并行计算机系统。该系统的硬件包括一台具有1024个处理器的SIMD机器SM-1，软件包括用于新语言的语言处理器，如扩展的C语言和扩展的Common Lisp.2。基于数据并行模型的应用对于一些高级应用程序，我们提出了新的数据并行算法，大大提高了执行性能。我们在我们的数据并行语言中实现了这些算法，并在SM-1上测试了它们的性能，证明了它们的有效性。基于数据并行模型的程序设计语言NCX为了使数据并行程序设计得到广泛应用，我们设计了一种独立于计算机体系结构的新型扩展C语言NCX。我们已经为不同架构的各种并行机器开发了NCX编译器。为了降低编译器开发的成本，我们开发了一种对NCX程序进行独立于体系结构分析的预处理器。到目前为止，所有的NCX编译器都使用了这个预处理器。矢量处理器数据并行语言的代码生成我们开发了一种在矢量处理器上高效执行数据并行程序的新方法。这种方法的基本思想是将数据并行程序中的一组数据对象作为一个可以在向量处理器上有效处理的单个向量来处理。我们在此基础上开发了一个实验性的NCX编译器。利用该编译器，我们在microvp上执行了一些应用程序，证明了该方法的有效性。

We have made fundamentl research on data-parallel computation model and its application to vector processors, from various points of view. The following research topics are included :1.Parallel processing system based on data-parallel modelWe have developed a massively parallel computer system which can execute data-parallel programs efficiently. The hardware of the system includes a SIMD machine SM-1 with 1024 processors, and the software includes language processors for new languages, such as an extended C language and an extended Common Lisp.2.Applications based on data-parallel modelFor some advanced application programs, we have proposed new data-parallel algorithms which greatly increase execution performance. We implement these algorithms in our data-parallel languages, measured their performance on SM-1, and proved the effectiveness.3.The programming language NCX based on data-parallel modelIn order to make data-parallel programming widely available, we have designed a new extended C language, called NCX,which is independent of computer architectures. We have developed NCX compilers for various parallel machines with difference architectures. In order to reduce the cost of compiler development, we have developed a preprocessor that performs architecture-independent analysis of NCX programs. This preprocessor is used in all NCX compilers so far developed.4.Code generation of a data-parallel language for vector processorswe have developed a new method to execute data-parallel programs efficiently on vector processors. The basic idea of this method is to handle a set of data objects in data-parallel programs as a single vector that can be processed efficiently on vector processors. We have developed an experimental NCX compiler based on this method. Using this compiler, we have executed some applications on the Micro VP and proved the effectiveness of the method.