Skip to content

Latest commit

 

History

History
48 lines (41 loc) · 2.88 KB

README.md

File metadata and controls

48 lines (41 loc) · 2.88 KB

High Performance Computing on Graphics Processing Units

Mean Shift Image Segmentation

Mean shift is a non-parametric feature-space analysis technique for locating the maxima of a density function, a so-called mode-seeking algorithm. Application domains include cluster analysis in computer vision and image processing.

Master Thesis

Example from the Thesis with Segmentation

Thesis.pdf

Today’s Graphic Processing Units (gpus) are not only good for gaming and graphics processing their highly parallel structure is predestined for a range of complex algorithms. They offer a tremendous memory bandwidth and computational power. Contrary to Central Processing Units (cpus), gpus are accelerating quickly and advancing at incredible rates in terms of absolute transistor count. Implementing a massively parallel, unified shader design, its flexibility and programmability makes the gpu an attractive platform for general purpose computation. Recent improvements in its programmability, especially high level languages (like C or C++), gpus have attracted developers to exploit the computational power of the hardware for general purpose computing. Several gpu programming interfaces and Application Programming Interfaces (apis) represent a graphics centric programming model to developers that is exported by a device driver and tuned for real time graphics and games. Porting non-graphics applications to graphics hardware means developing against the graphics programming model. Not only the diffculties of the unusual graphics centric programming model but also limitations of the hardware makes development of nongraphics applications a tedious task.

Therefore nVidia Corporation developed the Common Unified Device Architecture (cuda) that is a fundamentally new computing architecture that simplies sofware development by using the standard C language. Using cuda this thesis will show on the basis of an massively parallel application in which extent gpus are suitable for general purpose computation. Special attention is paid to performance, computational concepts, efficient data structures and program optimization.

The result of this work is the demonstration of feasibility of General Purpose Computation on gpus (gpgpu). It will show that gpus are capable of accelerating specific applications by an order of magnitude.

This work will represent a general guideline for suggestions and hints as well as drawbacks and obstacles when porting applications to gpus.

CUDA (multi gpu) Implementation of Mean-Shift

cuda-meanshift

OpenCL Implementation of Mean-Shift

opencl-meanshift