Closed tkoolen closed 5 years ago
This really feels like cheating the benchmark too much and adapting to the specific input data. I think we can still get better performance without "exploiting" using e.g. a SoA approach and vectorizing horizontally instead of vertically as we do now.
2.86 s -> 2.37 s.
With these changes, it's no longer beneficial to make
Vec
backed by a 4-tuple.Not sure if this is allowed. Happy to make it a separate file instead of changing
nbody-fast.jl
if that's preferred.