-
Notifications
You must be signed in to change notification settings - Fork 198
Open
Labels
Description
There remains one source of indeterminacy in game computations that we haven't thought about: incomplete specification of floating point operations. Many transcendental functions in floating point are only specified to plus/minus one ulp of precision. Furthermore, optimizations often cause intermediate floating-point results to retain extra precision in some cases, and not others. It's not clear to what extent these things will cause problems in real-world games.
There are several stages to this problem:
- Confirm that it's really a problem. That would mean collecting data across many platforms and browsers about whether floating point arithmetic (including complex nested expressions) gives different answers. We could, in theory, do this for visitors to code.world, and collect divergences from a reference value in files on the server.
- If it is a problem, find solutions. We can replace many transcendental math functions with deterministic variants (e.g., those at https://deterministic.js.org/; but keep in mind that we don't want to completely load deterministic.js, since it kills access to the system clock, among other things.)