ultrarand VS icfpc2019

There's a pretty interesting discussion on primitive arrays where the user is trying to match java performance exactly with an array-based thing that's simple enough to port, but is still a hair off due to clojure's default toward longs even for indexing (java wants int for array indices, so casts are necessary). Ended up getting identical performance using JiSE here, although the clojure variant with uncheckedIntCasts was fairly close (just not "equivalent to Java" per the common trope).

There's an intersting thread on zulip regarding getting identical performance with java arrays. Normally if you're using unchecked math and working with primitives, then you should get with a hair of java (not 2-3x slower). The remaining differences are due to uncheckedIntCast since Clojure forces longs for indices and java wants ints. The difference is I'm in the realm of nanos but is there. You can eliminate it with custom bytecode like with the JISE library which is fairly painless. Example .

Regarding hashing, custom hash functions even for records (if all fields are longs for example) can help substantially if you are hashing points some notes here. Even then, depending on what you are doing, you may be better off using a primitive numeric representation and a dense collection (e.g. a primitive array) since the mechanical sympathy may be substantial. It is also possible to do well using nested maps (particularly for memoization) when the pieces of a compound key are already trivially hashed. E.g. clojure.core/memoize just defines a varargs wrapper around a function and caches the sequence of args in a map. For earlier reasons, this is a poor strategy for many functions and performance can be substantially improved (e.g. for a 1 or 2 arg function of numbers or keys or strings even) if we defined fixed arity paths that store keys in nested maps instead of hashing complex collections.

There's a worked example of optimizing an ICFPC2019 entry here where the readme goes through a bevy of techniques and observations.