Parallel implementation of medium level algorithms on a monolithic ASIMD multiprocessor

The efficient implementation of algorithms with irregular data access or control-flow on a parallel SIMD processor requires specific architectural measures. This paper demonstrates the parallelization of medium-level algorithms on the HiPAR-DSP, a programmable RISC processor for real-time image processing with 4 or 16 parallel data paths. We show the efficient use of memory and ASIMD control capabilities of the processor for the parallel execution of a memory efficient Huffman decoding algorithm. Performance figures for a selection of further medium level algorithms are given, demonstrating that even a SIMD architecture can obtain high utilization for algorithms with data dependent control flow.