Accelerating huffman decoding of seismic data on GPUs

Huffman coding algorithm is widely used for seismic data compression because it offers good performance in terms of compression ratio. The algorithm compresses the data by assigning shorter code-words to the most frequent symbols, while the other symbols have longer code-words. It is difficult to accelerate the decoding process by exploiting the parallel architectures because the variable length of the code-words makes this process highly sequential. We propose a strategy that uses packets with headers to save the encoded data. This strategy forces the alignment of code-words at packet boundaries, allowing us to parallelize the decoding process. The parallel Huffman decoder was implemented on a GeForce GTX660 GPU and tested using different seismic datasets supplied by an oil company. Comparisons in terms of throughput (i.e. decoded data per second) suggest that our work is superior to other implementations. Experimental results allowed us to establish how the proposed strategy affects the compression ratio and how the number of threads per block affects the performance of the algorithm. Additionally, we show how the throughput is related with the compression ratio.