Over the last years, DMO has become a standard step in seismic processing flows. The main reason for this is that it clearly improves the stack for little extra computational cost. Recently, some research showed that the improvement is even more distinct when depth-variable velocity is considered in the DMO process Artley (1992); Godfrey (1992); Meinardus and Schleicher (1993). However, these new methods are computationally expensive and a constant-velocity DMO is often enough to give an idea of the geological structures involved or improve the velocity analysis. Therefore, a parallel implementation of constant-velocity DMO is natural to reduce the processing time cost.
In a 3-D constant-velocity medium, the DMO process uses a line operator Hale (1991). It convolves the input data with elliptical impulse responses, and sums the convolved data into the final stacked volume. The single dimension of the operator as well as its dip limitation are two reasons to consider 3-D constant-velocity DMO as a fast process whatever the geometry of the data acquisition. Unfortunately, the multi-azimuthal distribution of 3-D land data is a major obstacle for parallel implementation Biondi and Moorhead (1992).
This paper details two possible parallel algorithms for 3-D constant-velocity DMO. The first algorithm involves a spiral data movement that does not use the full computational capabilities of the connection machine (CM5). The second algorithm reduces the communication cost by processing time slices but requires additional care to avoid temporal aliasing of the operator.