Research and Development of an Automatic Vectorizing/Parallelizing Compiler

本研究从1990年开始，历时3年。在此基础上，设计并实现了自动向量化/并行化编译器和虚拟并行向量处理器模拟器。作为本工作的平台，采用了前人开发的自动向量化编译器V-Pascal。研究活动可分为：-1。自动向量化和并行化函数的新设计和实现；2 .虚拟并行矢量处理器模拟器高效堆叠/同步机制的设计与实现；3 .修改目标代码生成模块；4 .虚拟并行矢量处理器模拟器的实现；5 . BV-Pascal的并行化修订；全系统性能评估。新编译器的主要特点是：它可以以比任何其他商业编译器更精确的方式执行依赖性分析；建立了整个目标程序的依赖关系，实现了不同粒度代码片段的并行化；允许潜在并行任务之间的分层表示；通过估算每个任务的执行时间成本，可以轻松地拆分或合并这些并行任务，从而自动生成最优并行任务。这些方案在为本项目购买的Titan计算机上进行了测试和验证。

This research was performed for 3 years starting from 1990. As planned in the proposal, an automatic vectorizing/parallelizing compiler, as well as a simulator of a virtual parallel vector processor, was designed and implemented. As the platform this work, the previously developed automatic vectorizing compiler V-Pascal was used. The research activities can be itemized as : -1. New degign and implementation of automated vectorization and parallelization functions ;2. Design and implementation of an efficient stacking / synchronizing mechanism for the virtual parallel vector processor simulator ;3. Revision of object-code generation modules ;4. Implementation of the virtual parallel vector processor simulator ;5. Revision of BV-Pascal for parallelization ;6. System-wide performance evaluation.The main features of the new compiler are : it can perform dependence analyses in a way much more precise than any other commercial compilers ; dependence relations are obtained over the whole objective program to parallelize code fragments of various grain sizes ; allowing for hierarchical representation between potential parallel tasks ; these parallel tasks are split or merged with ease using the estimates of execution time cost of each task and thus automatically generating optimal parallel tasks. These schemes were tested and verified on the Titan computer purchased for this project.