Implementation of QOI encoding and decoding in Go.
For more info about QOI, see:
- the specification at https://qoiformat.org/qoi-specification.pdf
- the reference implementation at https://github.com/phoboslab/qoi
Even though this implementation is very straightforward and not particularly optimised (there is not much to optimise with such a simple format anyway), I found that it is much faster than the existing Go implementations:
| compared to... | decoding | encoding |
|---|---|---|
| takeyourhatoff/qoi | 4.3x faster | 12.4x faster |
| xfmoulet/qoi | 1.43x faster | 7.1x faster |
| arian/go-qoi | 7.5x faster | 7.5x faster |
Full benchmark comparisons can be found here.
I think a lot of the difference is caused by these libraries making huge amounts of unneeded allocations.
Compared to the reference C implem, I didn't make a full benchmark, but I have some evidence that this package would be about 20-50% slower. Encoding especially takes a hit because it needs to compare pixels a lot and Go doesn't have (C-style) union types, so 4 bytes cannot be cast to an uint32 for free to make a single comparison, it needs to do 4 compares.
The only thing I did that I would call "optimisation" (i.e. apart from doing sensible allocations) simply consists of replacing the check for whether we can use an OP_DIFF in the encoder (and similarly for OP_LUMA):
if dr >= -2 && dr <= 1 && dg >= -2 && dg <= 1 && db >= -2 && db <= 1 { ...
with this equivalent version:
if uint8((dr+2)|(dg+2)|(db+2))&0xfc == 0 { ...
This avoids having too many branches and for photographic images where OP_DIFF and OP_LUMA are plentiful, this can speed up encoding by more than 10 %.