improve performance #32
Description
Some avenues to consider, in roughly chronological order of observation, include:
- Issues replace get_epsilon with f64 EPSILON #1, replace get_min_val with f64 MIN_POSITIVE #2, reduce use of variable shadowing (geomath) #3, replace geomath::cbrt with f64::cbrt #4, replace geomath::fmod with direct calls to % operator #5, consider simplifying polyval #6, consider changes for sincosd #7, changes to consider for atan2d #8, consider changes to ang_round #9, avoid passing GEODESIC_ORDER as argument #10, and favor type usize for variables that are primarily an array index #12 for geomath.
- Modify geomath::norm to pass values by mutable reference and update them in place rather than returning a separate value.
- Negate geodesic._gen_inverse clam12 on the same line where it's assigned, and make it non-mut.
- Explore use of f64.sin_cos() over f64.sin() followed by f64.cos().
- Explore use of mem::swap() and relatives in place of things like {let c=a; a=b; b=c;}
- Look for places that assign the same calculated value more than once in a row, like:
M12 = sig12.cos(); M21 = sig12.cos(); // instead use M21 = M12; - Explore reducing scratch array allocations, through some combination of (a) more reuse and (b) larger allocations followed by slicing.
- Review choice of declared types, to reduce need for casting.
- Make more use of operation-assign operators, like
|=. - For optional inputs, standardize (at least for each individual case) on either None or NAN as the "no value" option, rather than using a 2-step translation process (e.g. None to NAN to calculated default).
None is likely to make much impact individually, but they might add up to something together.
I'll continue adding items to this list as I explore issues with returned values, until we get around to actually trying to make performance improvements.