MPEG-4高效率音訊編解碼器之增速及其在TI DSP平台上的實現

Abstract

由於數位音訊編碼技術的成熟發展，與音訊相關的產品，如MP3播放器、手機，在我們生活中扮演了一個很重要的角色。而所有的音訊壓縮標準中，MPEG-4 高效能音訊編碼(HE-AAC)提供了非常高的壓縮效率與不錯的音訊品質。在本篇論文中，針對HE-AAC編碼器的模組，我們提供了較快速的演算法，並且符合DSP系統的加速。我們也把HE-AAC編碼器實現在德州儀器公司(TI)的數位訊號處理器(DSP)上。 我們首先在DSP系統上針對HE-AAC編碼器分析其複雜度。我們發現QMF bank(正交鏡像濾波器)、降頻濾波器、暫態訊號偵測器、心理學模式、量化模組在DSP系統上耗費了大部分的運算量。因此我們針對這些模組去研究，並且提供了快速演算法來降低其複雜度。 針對頻帶複製(SBR)編碼器部分，我們提供了快速的暫態訊號偵測器與有效率的降頻濾波器結構。針對AAC編碼器部分，我們提供了查表的方式來減少心理學模式的複雜度，並且也提出了簡化的視窗轉換、簡化的TNS、快速的量化模組與簡化的短視窗分組方法。此外針對DSP架構，我們使用了很多加速的方法像是定點數運算、TI DSP的特殊指令群、單個指令存取多筆資料、迴圈的分解與巨集指令。最後修改過的HE-AAC編碼器版本在DSP系統上執行的週期，在相同的最佳化設定之下，比最原始的改善了大約55%並且還能維持相同的音訊品質。

Due to the recent advances of digital audio coding technology, audio coding related devices play an important role in our daily life such as MP3 players and mobile phone hand set. MPEG-4 High Efficient AAC (HE-AAC) provides a very high compression ratio and a good audio quality among all known audio coding standards. In this thesis, we propose several fast algorithms to speed up the MPEG-4 HE-AAC encoder for the DSP platform. Their implementations on the Texas Instrument (TI) TMS320C6416T fixed-point DSP are also presented. We first analyze the complexity of HE-AAC encoder on a DSP system. We find that the QMF bank, transient detector, downsampling filter, psychoacoustic model (PAM), and quantization modules require the most operational cycles on DSP. Hence, we study and suggest several fast algorithms to reduce their complexities. For the SBR encoder part, we propose a fast transient detector and an efficient decimation structure. For the AAC encoder part, we use a look-up table method to reduce the complexity of PAM and propose simplified block switching, simplified TNS, fast quantization and simplified short window grouping methods. We also adopt several DSP techniques to speed up the DSP, such as using fixed-point operation, intrinsic function, single instruction multiple data (SIMD), loop unrolling and macro function. Comparing to the original 3GPP HE-AAC encoder, the modified HE-AAC encoder save about 55% operational cycles under the same compiler optimization level and still maintains about the same audio quality.