image-png VS libjpeg-turbo

Encoding images to PNG with the "fast" compression mode is now an order of magnitude faster. This is achieved using a custom Zlib compression implementation that leverages assumptions about the patterns in PNG data to get reasonable compression ratios at phenomenal speeds.

However, the latest release of the png crate - published less than a day ago - has an ultrafast compression mode that's slightly faster than the QOI reference implementation at a comparable compression ratio. So basically you get the properties of QOI in a much more widely used image format.

In other areas, miniz_oxide is faster than miniz, Symphonia is faster than ffmpeg on most codecs, the not-yet-announced zune-png beats both libpng and the more heavily optimized libspng, and the png crate is getting considerable improvements too and also beats libpng.

gif, png, zune-jpeg are on par with their C counterparts in terms of performance

Fun fact: libpng-sys was so difficult to create safe Rust wrapper for that people just wrote their own, 100% safe Rust PNG decoder: https://github.com/image-rs/image-png

I didn't get time to finish the full spec for my attempt at a PNG decoder but it was working on common cases and was faster than image-png at the time https://users.rust-lang.org/t/new-rust-png-decoding-library/30952. I suspect it's not the case anymore since it changes it's deflate crate usage. Here is the perf issue on the image crate https://github.com/image-rs/image-png/issues/114. My code was way simpler than the one in the image crate so that might still be worth a read if you are interested.

The crate actually forbids unsafe. There's nothing necessarily wrong with unsafe code, but not having it provides a nice assurance. Plus if you are willing to tolerate (well audited) unsafe code then you may well be better off just using libpng.

> all decoders will render the same pixels

Not true. Even just within libjpeg, there are three different IDCT implementations (jidctflt.c, jidctfst.c, jidctint.c) and they produce different pixels (it's a classic speed vs quality trade-off). It's spec-compliant to choose any of those.

A few years ago, in libjpeg-turbo, they changed the smoothing kernel used for decoding (incomplete) progressive JPEGs, from a 3x3 window to 5x5. This meant the decoder produced different pixels, but again, that's still valid:

https://github.com/libjpeg-turbo/libjpeg-turbo/commit/6d91e9...

Last vestiges of this fact AFAIK were libjpeg, which had a macro NEED_SHORT_EXTERNAL_NAMES that shortens all public identifiers to have unique 6-letter-long prefixes. Libjpeg-turbo nowadays has removed them though [1].

[1] https://github.com/libjpeg-turbo/libjpeg-turbo/commit/52ded8...

While I think the move to safer code through Rust and other alternatives is a nice breath of fresh air, I doubt you can get these kinds of optimization without using unsafe code in Rust. These optimized implementations often require some kind of safety-bypassing memory modifications to work as efficiently ad they do.

There's a reason https://github.com/libjpeg-turbo/libjpeg-turbo/tree/main/sim... is filled with assembly files with conditional loading.

Sure. You'll see it very often in codec implementations. From rav1e, a fast AV1 encoder mostly written in Rust: https://github.com/xiph/rav1e/tree/master/src/x86

Large portions of the algorithm have been translated into assembly for ARM and x86. Shaving even a couple percent off something like motion compensation search will add up to meaningful gains.

Or the current reference implementation of JPEG: https://github.com/libjpeg-turbo/libjpeg-turbo/tree/main/sim...

zune-jpeg is 1.5x to 2x faster than jpeg-decoder and is on par with libjpeg-turbo.

This is interesting from the standpoint of how new JVM features can be used to improve performance (what I presume the article's main purpose to have been), but the image processing improvement itself isn't head-turning. Also, we've found that libjpeg-turbo (https://libjpeg-turbo.org/) is ~5x (IIRC, can re-run my JMH benchmark if anyone wants me to) as fast for decoding JPEGs as ImageIO, so we wouldn't even benefit from this change in 21 much.

libjpeg-turbo: https://github.com/libjpeg-turbo/libjpeg-turbo/blob/main/simd/x86_64/jsimdcpu.asm

libjpeg-turbo

For all but the simplest formats (basically BMP), the difficulty of implementing encoding/decoding from scratch is significant - well beyond a beginner's ability, and challenging/time-consuming even for senior developers. So, libraries are used in practice - e.g. libpng and libjpeg-turbo.