5 Commits
Author SHA1 Message Date
swaits 737f533f7e perf: add benchmark, optimize
Realized about a 10x speedup.

- add benches/million-updates.rs
- optimization: use rand instead of nanorand (because of thread_rng() initialization being muchf aster)
- update README

Signed-off-by: Stephen Waits <steve@waits.net>
2023-02-09 16:36:24 -07:00
swaits da6a70e08d build: enable and fix stricter clippy linting- add clippy to .build.yml to run on sr.ht- add missing docs- use core:: instead of std:: where possible- use hypot() and mul_add() for better numeric stability- make all float tests use asser_approx_eq() and explicitly check for NaN/Inf- add missing semicolons here and there- add #[must_use] to ensure caller uses returned value- add const fn where appropriate- change visibility from pub(crate) to pub (in private modules)Signed-off-by: Stephen Waits <steve@waits.net> 2023-02-09 10:54:42 -07:00
swaits e86e211bed test: add property tests to HeightVector
Which also helped me fix a bug (using is_sign_negative() when we actually are okay with -0.0).

Signed-off-by: Stephen Waits <steve@waits.net>
2023-02-07 21:37:21 -07:00
swaits 26a7b243dc fix: add numeric stability
Made `HeightVector` numerically safe; while `Vector` remains numerically unsafe (i.e. it can divide by 0, NaN, Inf, etc).

When HeightVector runs into a problem, it just re-randomizes itself. Since the algorithm converges relatively quickly, this seems to work fine.

Also fixed a bug so that now height will always start as a positive number (and stay that way, else get re-randomized).

Added safety to `NetworkCoordinate` which now checks that its error didn’t pull a `HeightVector` into a negative height; in which case it also re-randomizes.

Clarified the `random` construction, tweaked the DEFAULT_ERROR constant, and updated tests accordingly.

Signed-off-by: Stephen Waits <steve@waits.net>
2023-02-07 20:25:10 -07:00
swaits 441f83cc6f test: add proptest, make vec mul/div scalar safer
Signed-off-by: Stephen Waits <steve@waits.net>
2023-02-07 16:29:01 -07:00