Computer Science Colloquia
Friday, May 4, 2012
Advisor: Andrew Grimshaw
Attending Faculty: Kevin Skadron, Chair;
9:00 AM, Rice Hall, Rm. 242
Master's project presentation
Implementing Super-Resolution Ultrasonic image Reconstruction Using Matlab on the GPU
TONE (Time-Domain Optimized Near-Field Estimator) is a modern algorithm to reconstruct ultrasonic images. In this project, we try to port a CPU-based algorithm TONE to GPU using the latest GPGPU functionality of Matlab and Jacket (a third party Matlab GPU computing platform). A comparison of Matlabs own GPGPU computing and Jacket will also be compared. The parallelized version of TONE can process one image roughly under 500 seconds on ITC cluster. This is not fast enough because we want image processing to be near real-time. The possible means to improve the efficiency is to port the algorithm on GPU. One problem of using Matlab with GPU is the memory management. In Matlab, memory allocation and deallocation is controlled by Matlab itself, rather than by the programmer using low-level languages (C, CUDA for example). In our experiments, the image raw data is close to the capacity of graphics cards memory. This, without a close attention to memory management, will lead to out-of-memory errors. A graphics card with larger memory will bypass this issue. However, CPUs usually have a larger memory than GPUs. This report also explores the way to make a smaller memory graphics card possible to process a relatively larger raw image data.