two passes?

[Kidlike]

First of all, this is awesome - great work!

---

Since encoding is 20x-50x faster, would it make sense to run a 2nd pass, with a different direction, and then keep the smallest of the 2 results?

So, basically run a "horizontal" pass, then a "vertical" pass, and keep the smallest in size.

Most likely should not be the default, but if I was using QOI, I would like to have this option.

(I haven't read the code yet)

[tgonzalez89]

A vertical pass could be much slower because of cache locality, but it still could be faster than png. Maybe one could start at the "end" of the image (bottom-right) and go up (horizontally) to the top-left.

[millette]

Speaking of directions, and slightly related. I skimmed thru the code but my c is very rusty (no pun). I assume it's encoding the image line by line. If line 0 is processed left to right, is line 1 processed right to left? I would probably yield slightly better compression.

[Noxime]

I think ideally the image would be encoded in a Hilbert curve-order for, but that will trash cache quite badly. Anyone up for testing that approach and benchmarking?

[phoboslab]

The file format for QOI will not change anymore. See #37 for more info.

Ideas for a successor to QOI should be discussed here: https://github.com/nigeltao/qoi2-bikeshed/issues