Entropy Decoding on TriMedia/CPU64

The paper describes a software implementation of an MPEG–compliant Entropy Decoder on a TriMedia/CPU64 processor. We first outline entropy decoding basics and TriMedia/CPU64 architecture. Then, we describe the reference implementation of the entropy decoder, which consists mainly of a software pipelined loop. On each iteration, a set of look-up tables partitioning the VariableLength Codes (VLC) table defined by the MPEG standard are accessed in order to retrieve the run-level pair, or detect an end-of-block or error condition. An average of 21.0 cycles are needed to decode a DCT coefficient according to this reference implementation. Then, we focus on software techniques to optimize the entropy decoding software pipelined loop. In particular, we propose a new way to partition the VLC table such that by exposing the loop prologue to the compiler, testing each of the end-of-block and error conditions within the prologue becomes superfluous. This is based on the observation that either an end-of-block or error condition will never occur within the first table look-up. For the proposed implementation, the simulation results indicate that an average of 16.9 cycles are needed to decode a DCT coefficient. That is, our entropy decoder is more than 20% faster than its reference counterpart.