Software solution for Entropy Decoding on TM32 cores

Delft University of Technology Faculty of Electrical Engineering, Mathematics and Computer Science CE-MS-2003-13 Entropy Decoding is an essentially sequential task. Executing this task on a processor that benefits from Instruction Level Parallelism (ILP), Data Level Parallelism (DLP) or both requires an efficient implementation of Entropy Decoding. Entropy Decoding forms the part of MPEG-2 Decoding that exploits the least parallelism. Creating more parallelism in Entropy Decoding is expected to optimize MPEG-2 Decoding significantly. When writing an efficient MPEG-2 Decoder the result needs to be conform to the Berkeley MPEG-2 Decoder. In this thesis the Entropy Decoding process is optimized by compressing the data-stream from the Lookup Table to the Variable Length Decoder. In addition, the number of branches in the Entropy Decoding process is reduced. This stimulates Instruction Level Parallelism (ILP), that is exploited even more by using preloading data. The result of these three optimization steps is evaluated on the TriMedia32 core. Here a reduction of 30% in the number of cycles needed to decode a coefficient is achieved when going from 23 to 16 instruction cycles. In addition, an exceptionally high VLIW slot occupancy of more than 85% shows that the Entropy Decoding process is highly optimized. This result is integrated in the MPEG-2 Decoder written for SpaceCAKE. The resulting MPEG-2 Decoder has the same properties in terms of DLP compared to the original decoder. Despite the patch needed to conform to the data management in the multi-threaded MPEG-2 decoder, the improvement still is significant. On average an improvement of 12% is measured, with a peak improvement of 20%. These figures represent the lower bound of performance improvement. It is required to rewrite the data management structure of the Berkeley MPEG-2 Decoder to allow for a fully performing optimized Entropy Decoding process in this decoder. In conclusion, the optimized MPEG-2 Decoder gives a significant performance improvement based on remarkable results achieved by increasing the efficiency of the compiled Entropy Decoder source code, thus optimizing the Entropy Decoder process. Software solution for Entropy Decoding on TM32 cores