The performance impact of data placement for wavelet decomposition of two-dimensional image data on SIMD machines

Wavelet transform is a mathematical tool through which 2D spatial image data can be mapped into wavelet space for compact representation and for various signal analyses. The highly regular structure of the wavelet decomposition algorithm makes it well-suited for fine-grained parallelization. Most existing parallelization approaches focus on how to map computing functions to processors, but pay little attention to the problem of data placement. We investigate the impact of different data placement schemes on their achievable speedups in a MasPar MP-2 parallel computer. Our experimental results show that data communication is a dominating factor which can influence the overall system performance. We drastically speed up the computation of the wavelet transform by maximally localizing interprocessor communications for data exchange.<
>