[question] Is it deterministic with respect to different hardware platforms?

[ip]

I'm particularly interested in using this library for deterministic calculations in a real-time online game. Maybe anyone had this use case?

I.e. I need a library which produces exactly the same results (think memcmp()) on x86 and ARM processors (to cover mobile and desktops).

[PetteriAimonen]

I think they should be deterministic, as only integers are used. But I'm not aware of anyone actually having tried that out.

[grimaldini]

using fix16_from_float is not quite deterministic right? as you are doing a floating point arithmetic operation in the function's definition, therefore if you have something like fix16_t a = fix16_from_float(-2.1f) you can expect the value of a to be different in different architectures, is that right, @PetteriAimonen ?

[PetteriAimonen]

Well, floats are also supposed to be deterministic except for rounding of intermediate products. As fix16_from_float() is calculated in compile time it has pretty good chances of always yielding the IEE754 standard result. But who knows, I haven't tried it.

And anyway, a difference could only occur if the resulting fix16 value is very close to the rounding boundary. -2.1f is 2.0999999046325684 or 2.1000003814697266 which both are close to -137625.6 and round to -137626, so no chance of it rounding wrong. But if it was -2.0999984741210787 it might round to either -137625 or -137626.

[grimaldini]

so the only way around issues like this would be using fix16_t a = fix16_from_str("-2.1") anytime constants are defined, right? because the issue with -2.0999984741210787 might even happen when using FIXMATH_NO_ROUNDING if say we want to just truncate certain things and prefer consistency over accuracy.

[PetteriAimonen]

You can also do (fix16_one * 21 / 10) which would be all integer math and thus deterministic.