In this paper, we evaluate two different approaches for the compiler-based parallelisation of a C program for MPEG-2 decoding. The first approach experiments with a commercial auto-parallelising compiler for exploiting coarse-grain parallelism on a Silicon Graphics Power Challenge, while the second approach experiments with a simple assembly code scheduler for exploiting instruction-level parallelism. Results obtained through the high-level auto-parallelisation of the code are disappointing, far from those that can be achieved through manual parallelisation. On the other hand, better results are achieved when exploiting instruction-level parallelism; the corresponding compiler technology appears to be more efficient.
Arnaud Laffitte, Rizos Sakellariou, John R. Gurd