GPU implementation of block transforms

Traditionally, intensive floating-point computational ability of Graphics Processing Units (GPUs) has been mainly limited for rendering and visualization application by architecture and programming model. However, with increasing programmability and architecture progress, GPUs inherent massively parallel computational ability have become an essential part of today's mainstream general purpose (non-graphical) high performance computing system. It has been widely reported that adapted GPU-based algorithms outperform significantly their CPU counterpart. The focus of the thesis is to utilize NVIDIA CUDA GPUs to implement orthogonal transforms such as signal dependent Karhunen-Loeve Transform and signal independent Discrete Cosine Transform. GPU architecture and programming model are examined. Mathematical preliminaries of orthogonal transform, eigen-analysis and algorithms are re-visited. Due to highly parallel structure, GPUs are well suited to such computation. Further, the thesis examines multiple implementations schemes and configuration, measurement of performance is provided. A real time processing display application frame is developed to visually exhibit GPU compute capability.

Name thesis:degree_name

Master of Science in Electrical Engineering - (M.S.)

Discipline thesis:degree_discipline

Electrical and Computer Engineering

Year

2012

Author dc:creator

• Zhang, Boyan

Contributors dc:contributor

• Ali N. Akansu
• Richard A. Haddad
• Edip Niver

• Karhunen-Loeve transform
• Discrete cosine transform
• Graphics processing unit architecture
• Electrical and Electronics

Repository record dc:identifier

https://digitalcommons.njit.edu/theses/149

OAI identifier oai:identifier

oai:digitalcommons.njit.edu:theses-1148