INTERNATIONAL ORGANISATION FOR STANDARDISATION
ORGANISATION INTERNATIONALE DE NORMALISATION
ISO/IEC JTC1/SC29/WG11
CODING OF MOVING PICTURES AND AUDIO

ISO/IEC JTC1/SC29/WG11
N2957
October 1999 / Melbourne

Source: Audio Subgroup
Title: Revised Report on Complexity of MPEG-2 AAC Tools
Authors: S. R. Quackenbush, Y. Toguri, and Jürgen Herre

Revised Report on Complexity of MPEG-2 AAC Tools

This web page contains only excerpts of the document. The complete document is available as PDF file.

Introduction

This document is based on document and N2005, which presented the complexity of the tools in MPEG-2 AAC Main, Low Complexity and Sampling Scalable Rate Profiles. In addition, an analysis of the complexity of a one-band AAC Scaleable Sampling Rate (SSR) Profile decoder is provided in an annex of this document.

We desire to quantify the complexity of the tools in the MPEG-2 Advanced Audio Coding (AAC) decoder. They are:

huffman decoding
inverse quantization and scaling
M/S dematrixing
intensity stereo
coupling channel
backward adaptive prediction
temporal noise shaping (TNS)
inverse modified discrete cosine transform (IMDCT)
gain control and hybrid filter bank (inverse polyphase quadrature filter (IPQF)+IMDCT)

Unless otherwise indicated, complexity is specified in terms of

machine instructions required to realize the tool’s computations, as run on a typical (but unspecified) programmable digital signal processor
read/write storage locations
read-only storage locations

We assume that:

the target machine uses only IEEE floating point arithmetic, so that all floating point data require four bytes of storage. All storage is specified in terms of 32-bit words.
the coder block size is 1024 input samples, equivalent to 1024 spectral coefficients per channel.
an audio signal is sampled at 48 kHz, 16-bits per sample
the compressed bit rate is 64000 bits per second per audio channel

Furthermore, we only indicate storage that is required by a tool and cannot be shared or re-used by other tools. Specifically, we do not count temporary, stack-based scratch storage ("automatic" variables), as such storage is implicitly shared across tools.

Unless explicitly indicated, all complexity figures are for one audio channel.

Overview

One should consider two important categories of AAC decoder implementations: software decoders running on general-purpose processors, and hardware decoders running on single-chip ASICs. For these two categories the data presented in this document, augmented by demonstrated real-time software decoder implementations, can be summarized in the following table:

Decoder	Complexity
2-channel Main profile software decoder	40 % of 133 MHz Pentium
2-channel Low Complexity profile software decoder	25 % of 133 MHz Pentium
5-channel Main profile hardware decoder	90 sq. mm die, 0.5 micron CMOS
5-channel Low Complexity profile hardware decoder	60 sq. mm die, 0.5 micron CMOS
2-channel Scaleable Sampling Rate profile software decoder	not estimated
5-channel Scaleable Sampling Rate profile hardware decoder	53 sq. mm die, 0.5 micron CMOS (3 band decoder)

In this Table, the 3-band SSR decoder complexity is shown. The 4-band SSR decoder has almost the same complexity as Low Complexity Profile.

Heiko Purnhagen 08-Feb-2000