A performance model of fast 2D-DCT parallel JPEG encoding using CUDA GPU and SMP-architecture

The performance of image compression algorithms for big data can be enhanced using parallel computations. JPEG algorithm is a lossy compression method that uses DCT to eliminate high-frequency components. In this paper, we describe a cross-compatible design of JPEG on SMD and GPU architectures. To achieve maximal efficiency, we exploit the substantial parallelism to design an optimized version of JPEG based on thread model. A fair algorithm's evaluation on 24-bit BMP, using several performance metrics, is run on the fully optimized GPU using CUDA and SMP using SESC simulator. Our cross-architectural evaluation results revealed a 25.49 speedup in SESC and 21 in GPU and that CPU outperformed GPU for the JPEG.