Open ghost opened 8 years ago
For reference, they are currently:
// Epsilon is some tiny value that determines how precisely equal we want our floats to be
// This is exported and left as a variable in case you want to change the default threshold for the
// purposes of certain methods (e.g. Unproject uses the default epsilon when determining
// if the determinant is "close enough" to zero to mean there's no inverse).
//
// This is, obviously, not mutex protected so be **absolutely sure** that no functions using Epsilon
// are being executed when you change this.
var Epsilon float{32,64} = 1e-10
var MinNormal = float{32,64}(1.1754943508222875e-38) // 1 / 2**(127 - 1)
Do you have a suggestion for what values they should be for 32 and 64-bit versions of mathgl?
Also, what is MinNormal
? It's not documented beyond "1 / 2**(127 - 1)", and I don't know what that is.
So I have no clue what MinNormal is either. But according to this
https://github.com/golang/go/blob/master/src/math/const.go#L30-L33
it should be 1 / 2**(1023 - 1)
as for the default epsilon, my naive approach would be to make this n times smaller where n = MinNonZeroFloat32/MinNonZeroFloat64 but since Epsilon is already arbitrary i guess it could be any value that doesn't break the tests. (but it will obviously be much smaller)
https://www.wolframalpha.com/input/?i=(1e-10)%2F((1+%2F+2**(127+-+1+%2B+23))%2F(1+%2F+2**(1023+-+1+%2B+52))) 1e-289 i guess, but my method if most likely incorrect (it seems wayyyy too accurate)
So, I stubbed my toe on this one: https://github.com/go-gl/mathgl/blob/master/mgl64/quat.go#L440 That threshold value is way too large for its first use in the function. It's fine for the other use though.
They should quite obviously be different