Boosting the Performance of Multimedia Applications Using SIMD Instructions

Modern processors’ multimedia extensions (MME) provide SIMD ISAs to boost the performance of typical operations in multimedia applications. However, automatic vectorization support for them is not very mature. The key difficulty is how to vectorize those SIMD-ISA-supported idioms in source code in an efficient and general way. In this paper, we introduce a powerful and extendable recognition engine to solve this problem, which only needs a small amount of rules to recognize many such idioms and generate efficient SIMD instructions. We integrated this engine into the classic vectorization framework and obtained very good performance speedup for some real-life applications.

• Type: Conference
• Year: 2005
• Where: CC
• Authors: Weihua Jiang, Chao Mei, Bo Huang, Jianhui Li, Jiahua Zhu, Binyu Zang, Chuanqi Zhu
• Keywords: Automatic Vectorization Support, CC 2005, Efficient Simd Instructions, Programming Languages, SIMD ISAs