Towards Automatic Parallelisation for Multi-Processor DSPs

This paper describes a preliminary compiler based approach to achieving high performance DSP applications by automatically mapping C programs to multi-processor DSP systems. DSP programs typically contain pointer based memory accesses making automatic parallelisation difficult. This paper presents a new method to convert a restricted class of pointer-based memory accesses into array accesses with explicit index functions suitable for parallelisation. Different parallelisation approaches suitable for multi-processor DSPs are considered. We implemented our pointer conversion algorithm in the prototype Octave compiler where experimental results demonstrated that our technique increases the number of parallelisable loops from 6 to 24 for 11 of the DSPstone benchmarks. Furthermore our technique is shown to also improve the actual performance of DSP codes on single processor systems decreasing execution time by up to 33%.