Density function precision issues

[jacobsjo]

Deepslate density functions aren't always accurate when floating point precision has an effect. I'll look into this more in the future.

Discord discussion: https://discord.com/channels/738046951236567162/750811307925831841/1318399656148860970

• splines use float, we we should use Math.fround() to handle precision accurately
• some casts to int are apparently wrong

[unnecessarymb]

https://discord.com/channels/738046951236567162/750811307925831841/1291393498658111629

At

deepslate/src/math/noise/PerlinNoise.ts

Line 82 in 36dfb49

return value - Math.floor(value / 3.3554432E7 + 0.5) * 3.3554432E7

, you use Math.floor. However, in the MC source code, a utility function is used that clamps the result to [-2^63+1, 2^63-1]. At very high firstOctave levels, this causes the ImprovedNoise calls to diverge, due to extreme precision loss in MC’s case. This divergence is of particular use to me, as it allows me to specify a cube/square/horizontal region at any offset using only two noise queries. I’m pinging here instead of writing a full issue because I’m not entirely sure to what degree a fix would be in scope, as that code hasn’t been touched in two years and this hasn’t come up before me, and I would welcome input.