- Don't clamp all RNG calls to [0, FRACUNIT-1]. Only does this for P_RandomFixed now.
- Use rejection sampling for any clamped RNG calls, to remove modulo bias.
- Because of this, P_RandomRange ranges >= FRACUNIT are no longer undefined behavior.
- Added P_Random/M_Random to grab RNG output directly.
- Shuffle M_Random's RNG as well, since OS rand can be [0,INT32_MAX] instead of [0,UINT32_MAX].
- Instead of picking over numskins and adjust the number if it picks an already used skin, build and shuffle a list of allowed skins and pick over that instead.
- Fixes some issues with increased weighting towards the skins immediately after your rivals' skins.
- Use PR_RULESCRAMBLE instead of PR_UNDEFINED.
- Why am I doing it in this branch? I'm getting re-acquainted with how bots are handled between matches in advance of disabling them for bonus events without losing or corrupting data.
Previously it was using P_SetRandSeed, which sets both initial and current seed to the same thing, since it's meant for level load init. It first set all RNG seeds to the current seed, and then it set all of the seeds to their initial value. The comment about this just says its "stupid and hacky", and I have no idea how it even worked before.
Now we send over both init seed and current seed independently and set them both. Hopefully this will fix the desyncs.
P_RandomChance is now a macro for something that should happen a
certain percentage of time.
P_SignedRandom was moved to a macro. Nobody cared.
# Conflicts:
# src/p_inter.c
This actually passes most diehard tests, as opposed to the old RNG.
It's also similarly fast.
Internally the PRNG generates a fixed point number from [0,1) now,
which makes P_RandomKey and P_RandomRange much easier to
calculate. P_Random is just a simple shift, as well.
Also, the lack of floating point math in P_RandomKey and
P_RandomRange now is probably for the best.
