Implementation of CT algorithms on the Cell Broadband Engine

Midtvejs PhD-forsvar - Mikkel Engbo: Implementation of CT algorithms on the Cell Broadband Engine

In the last few years, CT-scanning has become popular, not only in the medical industry, but also in other areas. The Danish Meat Research Institute has undertaken a project to investigate the feasibility of CT-scanning of pigs prior to slicing, gaining absolute knowledge of the carcass at hand. This knowledge can be used for classification of the pigs, guidance to the robots on the slaughter-line, etc.

The project involves investigating the physical requirements for installing such a device in a
slaughterhouse and analyzing of scanned volumes, investigating into what information can be extracted. Another goal is to analyze the computational requirements of the reconstruction and subsequent analysis of the volumes. Due to the large quantities of data and the complexity of the problems, implementing the algorithms on a standard desktop PC will not suffice. To this end, the Cell processor has been suggested as a possible alternative due to the low cost and high potential.

The goal of this thesis has been to implement two algorithms for CT on the Cell in form of a PS3, investigating whether the parallelism of the Cell can be exploited in order to gain speedup in the different algorithms. The first problem, classification of tissue involves four subproblems, each with different properties to accommodate when parallelizing and vectorizing. The result very satisfactory with a throughput of over 370fps, or 6.9 times as
fast as the generic implementation.

Censor Christian S Pedersen, DAIMI
Supervisor Brian Vinter, DIKU, eScience