單一指令多重資料編譯器之開發

Abstract

多媒體指令集在通用的處理器中是非常普遍的。它們是由一組短向量指令組成的。為了要完全的利用單一指令多重資料(SIMD)架構的效能，我們必須要產生高效率單一指令多重資料的程式，這樣單一指令多重資料的架構下才能完全的發揮其效益。本文主要在開發單一指令多重資料編譯器,有效地利用短向量指令集自動地產生高效率單一指令多重資料的程式之技術。當你做運算時，短向量單一指令多重資料的架構並不支援不連續的記憶體位址。所以，需要把資料重新排列。在此需要解決的問題就是如何把不連貫性的記憶體位址排列成連貫及有效的處理多種不同的資料類型。

Multimedia extensions are nearly ubiquitous in today’s general-purpose processors. These extensions consist primarily of a set of short vector instructions that operate on same opcode to a vector of operands. To exploit the SIMD capabilities of these architectures, it has prompted the needs for generating efficient simdized codes that SIMD architectures can benefit from. This Thesis sets out to develop a vector SIMD compiler with techniques that target short vector instructions effectively and automatically. Operations on non-contiguous vector elements are not supported and they require explicit data realigning. Thus, one of the most common aspects of compilation is the effective management of memory alignment and coping with mixed data type. We identify several new challenges arisen in simdizing multimedia applications, and provide some solutions to these challenges. Simdizing such computation efficiently is therefore an ambitious challenge for compiler designer. We implemented an automatic simdization framework that supports effective simdization in the presence of control flow, memory misalignment, and mixed length conversion.