DLP acceleration on general purpose cores

Abstract

High-performance and power-efficient multimedia computing drives the design of modern and increasingly utilized mobile devices. State-of-the-art low power processors already utilize chip multiprocessors (CMP) that add dedicated DLP accelerators for emerging multimedia applications and 3D games. Such heterogeneous processors deliver desired performance and efficiency at the cost of extra hardware specialized accelerators. In this paper, we propose dynamically-tuned vector execution (DVX) by morphing one or more available cores in a CMP into a DLP accelerator. DVX improves performance and power efficiency of the CMP, without additional costs for dedicated accelerators.