|
计算机科学 2007
SIMD Code Generation for Multimedia Processors
|
Abstract:
The SIMD (Single Instruction Multiple Data extensions appeared in general-purpose microprocessors provide new architectural support for improving performance of multimedia applications, but current compiler techniques cannot exploit them well. Based on the idea of integrating program analysis of the compiler front-end and machine information of the back-end, this paper proposes a new algorithm for SIMD code generation, which can identifies program parallelism effectively to generate SIMD code adopting enhanced tree parsing techniques. The algorithm has been implemented in SUIF (Stanford University Intermediate Format compiler infrastructure, and experimental results for a group of multimedia kernels show that the algorithm proposed can reduce 47% cycles on average compared to non-SIMD code.