high performance computing on graphics processing units: hgpu.org

Stereoscopic rendering methods generate two images for binocular viewing. Exploiting the spatial coherence present in the surfaces rendered, it is possible to reduce the rendering time of a stereoscopic image pair significantly. Ray tracing is a frequently used rendering method for realistic image synthesis applications. With the tremendous increase in computational power of Graphics Processing Units and their availability for general purpose programming, ray tracing methods suitable for Graphics Processing Units have been developed. In this thesis, a parallel stereoscopic ray tracing method that reuses view independent information in one of the stereoscopic image pairs for generating the other image is presented. The results of the method are analyzed using a perception based error detection mechanism. The main contribution lies in the new parallel algorithm that takes advantage of the massive parallel processing power of the modern Graphics Processing Units efficiently. Results show that about 20 times speed-up is achieved over a previous work which is based on a sequential algorithm in the worst case among the experiments carried out. According to the perception based quality analysis performed, lower than only 2% of the pixels in high resolution images is expected to be perceived to be in error.