多媒體應用程式之分枝特性分析

Abstract

現代的嵌入式處理器（embedded processor），像是Intel 所發展的XScale 處理器，採用動態分支預測器（dynamic branch redictor）來減少因分支（branch）而產生的效能降低。現存的分支預測器均是針對使用在一般用途的中央處理器（general purpose processor）上的應用程式來設計，像是gcc 及gzip。然而，對智慧型手機等嵌入式系統（embedded system）來說，多媒體及通訊應用程式（multimedia/communication application）才是兩個主要的工作量（workload）。我發現一般常使用的雙模態分支預測器（bimodal branch predictor）在多媒體與通訊應用程式上的預測正確率 （prediction accuracy），比起在一般用途的中央處理器上的應用程式的表現差很多。在這篇論文中，我詳細地分析多媒體與通訊應用程式中分支的行為特性。我發現在這些應用程式中，迴圈（loop）的重複次數非常地固定，且大部分的迴圈均屬於計數迴圈（counted loop）。我同樣發現，在連續的陣列（array）元素比對上，發生許 多分支預測錯誤。最後，我發現許多控制條件分支的變數含有可預測性的數值變化。這些結果對設計嵌入式處理器中的分支預測器來說，是很重要的發現。

Modern embedded processors, like Intel's XScale, employ dynamic branch prediction to reduce performance penalty caused by branch instructions. Existing branch predictor de- signs are all based on the behavior of applications on a GPP (general purpose processor), such as gcc and gzip. However, for an embedded system, such as smart phone, multimedia and communication applications are two main workloads. I found that a commonly used bimodal branch predictor actually has lower branch prediction accuracy for multimedia/communication workloads than workloads for GPP. In this thesis, I perform detailed analysis on the branch behavior of multimedia/communication applications. I found that the iteration counts of loop branches in these applications are very regular, and most of them are counted loops. I also found that many mispredictions of if-branches occur on sequential comparisons of array elements. Finally, many of the values of control variables in if-branches have the property of data predictability. The findings of this work are important for branch predictor design for embedded processors.