GPU architectures for wavelet-based image coding acceleration /
- Autores: Pablo Enfedaque
- Directores de la Tesis: Francesc Aulí Llinàs (dir. tes.), Juan Carlos Moure López (codir. tes.)
- Lectura: En la Universitat Autònoma de Barcelona ( España ) en 2017
- Idioma: español
- Tribunal Calificador de la Tesis: Manuel Ujaldón Martínez (presid.), Joan Bartrina Rapesta (secret.), Mikel Lujan (voc.)
- Programa de doctorado: Programa de Doctorado en Informática por la Universidad Autónoma de Barcelona
- Materias:
- Enlaces
- Tesis en acceso abierto en: DDD
- Resumen
Modern image coding systems employ computationally demanding techniques to achieve image compression. Image codecs are often used in applications that require real-time processing, so it is common in those scenarios to employ specialized hardware, such as Field-Programmable Gate Arrays (FPGAs) or Application-Specific Integrated Circuits (ASICs). GPUs are throughput-oriented, highly parallel architectures that represent an interesting alternative to dedicated hardware. They are software re-programmable, widely available, energy efficient, and they offer very competitive peak computational performance.
Wavelet-based image coding systems are those that employ some kind of wavelet transformation before the data coding stage. Arguably, JPEG2000 is the most representative of those systems. Many research projects have tried to develop GPU implementations of JPEG2000 to speed up the coding pipeline. Although some stages of the pipeline are very suitable for GPU computing, the data coding stage does not expose enough fine-grained parallelism. Data coding is the most computationally demanding stage (75% of the total execution time) and represents the bottleneck of the pipeline. The research presented in this thesis focuses on the GPU computing of the most critical stages of wavelet-based image coding systems: the wavelet transform and the data coding stage.
This thesis proposes three main contributions. The first is a GPU-accelerated implementation of the Discrete Wavelet Transform. The proposed implementation achieves speedups up to 4x with respect to the previous state-of-the-art GPU solutions. The second contribution is the analysis and reformulation of the data coding stage of JPEG2000. We propose a new parallel-friendly high performance coding engine: Bitplane Image Coding with Parallel Coefficient Processing (BPC-PaCo). BPC-PaCo reformulates the mechanisms of data coding, without renouncing to any of the advanced features of traditional data coding. The last contribution of this thesis presents an optimized GPU implementation of BPC-PaCo. It compares its performance with the most competitive JPEG2000 implementations in both CPU and GPU, revealing speedups up to 30x with respect to the fastest implementation.
