推断与推测技术在现代高性能微处理器设计中的应用研究

The Application Research of Predication & Speculation for Modern High Performance Microprocessor Design

在现代高性能微处理器设计中 ,推断和推测成为开发指令级并行性ILP(InstructionLevelPara llelism)的两种重要技术途径 .推断的目的是打破程序间固有的控制流程 ,将控制相关转变为数据相关 ,使指令级并行性识别从一个基本块扩大为一个超块 .推测执行是为打破分支或访存引起的相关问题而进行的操作 ,进一步分为控制推测和数据推测 .控制推测的目的是打破分支和其他操作间的相关性 ,进而由编译器在一个超块内识别并行性 ,减小控制相关的高度 .数据推测则是消去访存相关 ,提高指令级并行度 .该文首先对推断和推测本身进行分析 ,然后在此基础上进一步将推断、推测技术相结合 ,并应用于高性能媒体处理器的设计中 .性能评价和比较结果显示 ,两种技术相结合将比任何一种技术都更加行之有效 .

Predication&speculation are key techniques in the design of modern high- performance microprocessors in order to expose sufficient amounts of instruction level parallelism to the hardware. The predication whose purpose is to change the original program control flow can be able to convert control dependency to data dependency. The predication can also be used to identify the instruction level parallelism from a basic block to a hyperblock. The speculation can eliminate dependency problem caused by branch or memory access. Control speculation and data speculation are chief components of the speculation. Control speculation can eliminate dependency problem between branch and other operations. In this way, compiler can be able to identify instruction level parallelism in a hyperblock and decrease the height of control dependency. Data speculation can remove the dependency of memory access and increase the instruction level parallelism. Both of these two techniques have individually been analyzed in this paper. In the mean time, we focus on the study of the combination with each other. Both of these two techniques are used in our high performance media microprocessor design. Performance evaluation and comparison show that both of the two techniques can be elegantly combined to achieve much better performance than each individual of them.