Parallel huffman decoding: presenting a fast and scalable algorithm for increasingly multicore devices

Huffman encoding provides a simple approach for lossless compression of sequential data. The length of encoded symbols varies and these symbols are tightly packed in the compressed data. Thus, Huffman decoding is not easily par-allelisable. This is unfortunate since it is desirable to have a parallel algorithm which scales with the increased core count of modern systems. This paper presents a parallel approach for decoding Huffman codes which work by decoding from every location in the bit sequence then concurrently combining the results into the uncompressed sequence. Although requiring more operations than serial approaches the presented approach is able to produce results marginally faster, on sufficiently large data sets, then that of a simple serial implementation. This is achieved by using the large number of threads available on modern GPUs. A variety of implementations, primarily OpenCL, are presented to demonstrate the scaling of this algorithm on CPU and GPU hardware in response to cores available. As devices with more cores become available, the importance of such an algorithm will increase.