oa Optimizing SRME: Exploiting the Power of GPUs and High-Core Count CPUsBEI

In this work, we review the optimization compromises made for running Surface-Related Multiple Elimination (SRME) on GPUs and high core count CPUs. The most important optimization is reducing disk reads by caching input data in the fastest levels of memory. Other optimizations include running asynchronous tasks, buffering the pipeline of data movement, and recycling redundant calculations when outputting results in common-shot order. However, care must be taken not to introduce an excessive memory footprint for large datasets. Therefore, we prioritize total memory for reducing the number of disk reads by serializing some steps in the data movement pipeline and splitting large shots with an offset-based binning scheme.