Optimise usage of pow using fast equivalents and exp2f

[Rossmaxx]

No description provided.

[LostRobotMusic]

Don't use fastPow in these cases, that approximation is extremely inaccurate and should only be used when precision isn't required. This is likely a major backwards compatibility break, always perform null tests with these things.

(Note that fastPow10 in contrast doesn't use an approximation and is therefore perfectly safe, despite the similar name.)

[DomClark]

I'd rather we used std::exp2, which will operate at the appropriate precision for its argument, than std::exp2f, which causes its argument to be implicitly converted to a float.

Also, in addition to Lost Robot's comment on the inaccuracy of the fastPow approximation, I would add that a lot of these cases are actually in UI code or file IO code, where performance doesn't matter nearly as much, and the cost of the computation is most likely dwarfed by other operations. As well as testing for regressions in behaviour, you should also profile the application before and after the changes to ensure we actually see a gain in performance in compensation for the reduction in accuracy.

[Rossmaxx]

I'd rather we used std::exp2,

Noted

, in addition to Lost Robot's comment on the inaccuracy of the fastPow approximation, I would add that a lot of these cases are actually in UI code or file IO code, where performance doesn't matter nearly as much, and the cost of the computation is most likely dwarfed by other operations.

I have reverted the of calls to fastPow. What exists now is calls to fastPow10f (mind the 10f). That function is accurate AND fast.

As well as testing for regressions in behaviour, you should also profile the application before and after the changes to ensure we actually see a gain in performance in compensation for the reduction in accuracy.

I don't have profiling set up in this branch. I'll try something tho.