| Summary: | 碩士 === 國立彰化師範大學 === 資訊工程學系 === 105 === With the age of big data coming, more and more researchers use General-purpose computing on graphics processing units (GPGPU) to accelerate data parallel processing on different algorithms. Convolution filtering is an important operation to graphic process, which can be applied in noise filtering, edge detection, image sharpening or image blurring. The genetic algorithm provides an efficient method for training the weight value of the filter, which makes input signal filtered and calculates respective input signal for the fitness value. However, system has to evaluate the fitness values repeatedly when performing image processing, resulting in taking longer time for training comparatively. Therefore, it is possible to use the genetic algorithm to accelerate fitness evaluation in parallel on the GPGPU in a short time. There are four methods in the previous research. Sub-images-based method (SBM) is which provides the configuration and the best performance on GPGPU. Our approach is to improve the SBM method by reusing registers according to their characteristics. Although the fastest storage unit in the computer architecture is register, the capacity is restricted. Excessive use of the registers will cause the throughput to be higher. We propose a method is to reduce the use of registers for the limited temporary storage, in order to achieve the best performance. We evaluated the performance of different sizes of images on TITAN X, and analyzed the results of our method implementation in different memory configurations. The approach we proposed has up to a maximum of 3.5 times improvement.
|
|---|
