Date of Award

2017

Document Type

Open Access Dissertation

Degree Name

Doctor of Philosophy in Electrical Engineering (PhD)

Administrative Home Department

Department of Electrical and Computer Engineering

Advisor 1

Saeid Nooshabadi

Committee Member 1

Timothy Havens

Committee Member 2

Zhuo Feng

Committee Member 3

Soner Onder

Abstract

Following the standardization of the latest video coding standard High Efficiency Video Coding in 2013, in 2014, multiview extension of HEVC (MV-HEVC) was published and brought significantly better compression performance of around 50% for multiview and 3D videos compared to multiple independent single-view HEVC coding. However, the extremely high computational complexity of MV-HEVC demands significant optimization of the encoder. To tackle this problem, this work investigates the possibilities of using modern parallel computing platforms and tools such as single-instruction-multiple-data (SIMD) instructions, multi-core CPU, massively parallel GPU, and computer cluster to significantly enhance the MVC encoder performance. The aforementioned computing tools have very different computing characteristics and misuse of the tools may result in poor performance improvement and sometimes even reduction. To achieve the best possible encoding performance from modern computing tools, different levels of parallelism inside a typical MVC encoder are identified and analyzed. Novel optimization techniques at various levels of abstraction are proposed, non-aggregation massively parallel motion estimation (ME) and disparity estimation (DE) in prediction unit (PU), fractional and bi-directional ME/DE acceleration through SIMD, quantization parameter (QP)-based early termination for coding tree unit (CTU), optimized resource-scheduled wave-front parallel processing for CTU, and workload balanced, cluster-based multiple-view parallel are proposed. The result shows proposed parallel optimization techniques, with insignificant loss to coding efficiency, significantly improves the execution time performance. This , in turn, proves modern parallel computing platforms, with appropriate platform-specific algorithm design, are valuable tools for improving the performance of computationally intensive applications.

Share

COinS