Improving Multimedia Applications through Specialization of DCT/IDCT Kernels

The DCT/IDCT kernels play an important role in multimedia applications and image processing. There are several attempts to improve the execution speed of these kernels at different levels of code. The optimizations in these implementations are either aimed at reducing the number of operations, or at exploiting various features provided by modern processor architectures. In this article, we suggest a novel approach of improving the performance of such kernels through code specialization. We generate multiple specialized versions and transform them at object-code level to generate another optimized version. The experiments have been performed over IA-64 architecture with icc compiler. We are able to achieve a good speed-up of more than 2 and 1.48 for IDCT and DCT respectively for different implementations of multimedia application.