Function div_255

pub(crate) const fn div_255(val: u16) -> u16

Perform an approximate division by 255.

There are three reasons for having this method.

1. Divisions are slower than shifting + adding, and the compiler does not seem to replace divisions by 255 with an equivalent (this was verified by benchmarking; doing / 255 was significantly slower).
2. Integer divisions are usually not available in SIMD, so this provides a good baseline implementation.
3. There are two options for performing the division: One is to perform the division in a way that completely preserves the rounding semantics of a integer division by 255. This could be achieved using the implementation (val + 1 + (val >> 8)) >> 8. The second approach (used here) has slightly different rounding behavior to a normal division by 255, but is much faster (see https://github.com/linebender/vello/issues/904) and therefore preferable for the high-performance pipeline.

Four properties worth mentioning:

• This actually calculates the ceiling of val / 256.
• Within the allowed range for val, rounding errors do not appear for values divisible by 255, i.e. any call div_255(val * 255) will always yield val.
• If there is a discrepancy, this division will always yield a value 1 higher than the original.
• This holds for values of val up to and including 65279. You should not call this function with higher values.