Discrete Wavelet Transformation Implementation in GPU through Register Based Strategy

Abstract

The significant architectural changes made by Nvidia during the launch of Kepler architecture in 2012, upgraded its GPUs with greater register memory and rich instructions set to have communication between registers through available threads. This created a potential for new programming approach which uses registers for sharing and reusing of data in the context of the shared memory. This kind of approach can considerably improve the performance of applications which reuses implied data heavily. This work is based upon of register-based implementation of the Discrete Wavelet Transform (DWT) with the help of CUDA and openCV. DWT is the data decorrelation approach in the area of video and image coding. Results of this particular approach indicate that this technique performs at least four times better than the best GPU implementation of the DWT in past. Experimental tests also prove that this approach shows the performance close to the GPUs performance limits.