Generation of a Fast JPEG 2000 Encoder using SPIRAL

The thesis deals with different levels of representation of the bit-plane coding algorithm and its transformation and optimization that exploit the structure of the algorithm to better match the hardware platform. The main focus is on extensions of high level language (Operator Language) used in SPIRAL and the presentation and generation EBCOT algorithm in the of the JPEG 2000 encoder. The code compiler is also extended to generate the optimized code. Summary Domain specific program generators reduce the need for tuning or rewriting performance libraries when new platforms are released. This thesis extends SPIRAL's framework to automatically generate the entropy coding part in JPEG 2000 encoder-EBCOT (Embedded Block Coding with Optimal Truncation), which is the most time-consuming component in the entire encoding process. The EBCOT algorithm is initially expressed in the high level representation of Operator Language (OL), which is an extension of the Signal Processing Language (SPL). Breakdown rules and manipulation rules are then applied to generate and optimize formula on the implementation level. The formula is then translated into C-like code and code level optimization is applied. Finally, the C-like code is optimized by the spiral code compiler and is unparsed into C code. For encoding JPEG 2000 images, parallelization is used due to the fact that blocks are independently encoded, this exploits the multi-core functionality of the most recent commercial processors. Vectorization is also used as an option to fully use the capabilities of short vector instruction set (Streaming SIMD Extensions) existing on the latest Intel Processors. A special adapted version of EBCOT is used to replace the open source reference implementation JasPer and the encoding performance compares and even outperforms that of Intel Performance Primitive. Acknowledgements I want to thank all person who supported me to make this work possible. In particular, I want to thank my advisor Christian W. Probst at DTU for his input and feedback. I also want to thank my advisors Franz Franchetti and Markus Püschel at CMU for giving me the chance of working on my thesis with the SPIRAL team in Pittsburgh and their advice and ideas throughout the project. Additionally, I want to thank Frédéric de Mesmay and Yevgen Voronenko from the SPIRAL team for their essential help on the SPIRAL platform. I would also like to thank Karl Sigurjónsson for his comments on the report. I want to thank my family and especially my parents for always believing …