This parallel processing data flow poses some unique problems to anti-aliasing compared to more convention trace-sequential scalar or vector processing strategies. In a trace-sequential migration, several copies of a single input trace can be filtered at different high-cut frequencies and then used appropriately during anti-aliasing depending on the local migration operator dip. In a parallel data flow, I cannot afford to keep an extra copy of my input data in memory, since it is as large as my output image volume, and nearly half the total available memory in order to optimize parallelism in the computations. I have experimented with a more conservative memory-use algorithm structure for the input data, but the underlying principle of not easily maintaining multiple copies of input data in memory remains the same.
Fortunately, the local triangle lowpass filtering approach to anti-aliasing can be done on the fly without storing multiple copies of input data. All the input traces are first given a full 3-D time derivative (as opposed to the half-time derivative of 2-D migration), and then integrated once causally, and once acausally. Each process is implemented in parallel over the surface coordinate direction, each trace residing serially in-processor. The triangle filtering for any triangle length N is then just a 3-point filter operation on the doubly integrated traces [see Claerbout (1992)], which is also implemented in parallel over all input traces.