I haven't looked at the source, but the way to write one of these that's fast is to use incremental computation. [0] It tends to take a lot of RAM, but the spe

I haven't looked at the source, but the way to write one of these that's fast is to use incremental computation. [0] It tends to take a lot of RAM, but the speedups possible are 3-5 orders of magnitude in the projects I've studied. Here's a parallel C++ library to do so that requires relatively minor changes to the original algorithm. [1] [2] I also believe it would be possible to develop a compiler framework to do the transformation mostly automatically, using an MLIR dialect. [0] https://en.wikipedia.org/wiki/Incremental_computing https://en.wikipedia.org/wiki/Incremental_computing [1] https://dl.acm.org/doi/pdf/10.1145/3409964.3461799 https://dl.acm.org/doi/pdf/10.1145/3409964.3461799 [2] Source code: https://github.com/cmuparlay/psac https://github.com/cmuparlay/psac