Automatic SIMD Parallelization of Embedded Applications Based on Pattern Recognition

This paper investigates the potential for automatic mapping of typical embedded applications to architectures with multimedia instruction set extensions. For this purpose a (pattern matching based) code transformation engine is used, which involves a three-step process of matching, condition checking and replacing of the source code. Experiments with DSP and the MPEG2 encoder benchmarks, show that about 85% of the loops which are suitable for Single Instruction Multiple Data (SIMD) parallelization can be automatically recognized and mapped.