overlay mask bug

[realfinder]

as it said here https://forum.doom9.org/showthread.php?t=183710

[pinterf]

Explanation in
https://forum.doom9.org/showthread.php?p=1961818#post1961818

[pinterf]

In my opinion this is not a bug, integer implementation may need exact explanation in documentation.
Or one would use 32 bit float format instead.

[realfinder]

thanks, so vs Overlay is slower, right?

[pinterf]

Eh, this turned into a one-week work, including reconstruction of the old ugly codebase and rewriting the all the stuff and added AVX2 as well.
Now I have benchmarks as well, it seems that the accurate new 8 bit basic (mask=yes, opacity=1.0) Overlay is slower (480fps vs 550fps - AVX2, 305 vs 550fps for pre-AVX2 processors)
But all the other cases - when opacity is <1.0 or bit depth>=10 we got a significant speed boost.

[pinterf]

Maskless blend is broken

[pinterf]

Maskless blend regression is fixed via 6884b1c