Congratulations to the Halide folks on wpd, their new decoder. The performance numbers are impressive. If wpd (a Rust library) works for you, great, you should use it!

But since Wuffs was mentioned, I'll just drop a few selling points for why you'd still consider Wuffs.

1. Wuffs' implementation is transpiled to C code (and that in-C-form is checked into the repository, as well as into the leaner google/wuffs-mirror-release-c Github repo). If your existing project is C/C++, not a Rust one, then it's very easy to add Wuffs as a dependency. It's like adding any other third-party C library. It's just not hand-written .c code. It's hand-written .wuffs code that gets transpiled to a single-file C library, as easy to integrate as the STB libraries but memory-safe.

1.a. Similarly, if you're a Python project, or Java, or whatever, if you can wrap C code, you can wrap the Wuffs library (in its C form) and still get in-process, memory-safe image decoding without having to add a new toolchain to your build process.

2. Wuffs is a zero-capability language. It's a language for writing (safe) libraries, that only compute. It's not a language for writing applications. The Wuffs language cannot open files or write to the network. It can't even dynamically allocate memory. That means that Wuffs code can operate under `SECCOMP_MODE_STRICT` sandbox that prohibits basically everything except reading from stdin and writing to stdout.

2.a. For out-of-process, extremely memory-safe image decoding, the example/convert-to-nia/convert-to-nia.c program in the Wuffs repository reads an image (JPEG, PNG, WebP, etc) on stdin and writes NIA (a trivial image format, similar to Farbfeld) on stdout. Even if you don't want to audit the Wuffs language and toolchain itself, the security review for that convert-to-nia program is also absolutely trivial, because one of the first things that the main function does is to self-impose a `SECCOMP_MODE_STRICT` sandbox.

3. Wuffs uses intrinsics for SIMD, like C/C++, but memory safety is enforced on all loads and stores. And the toolchain also enforces that Wuffs general code can't call into Wuffs AVX2-using code unless your cpuid is AVX2-capable. The memory safety story of all of that is a bit better than just "we do SIMD via assembly in unsafe blocks". Wuffs (the language) doesn't even have an "unsafe" keyword.

We got a pull request (https://github.com/google/wuffs/pull/168) in February to add lossy WebP support. Lossless WebP (which in some sense is an entirely different format, just reusing the WebP "brand") has had a Wuffs implementation for a couple of years now.

Anyway, the PR included SIMD acceleration and performance was on par with libwebp (C code).

The PR's code was, as far as I could tell, somewhat or mostly AI assisted. While that's great in terms of features, I still have more confidence in hand-crafted code.

I have since been working to manually rewrite the PR. I'd also like to add animation support, and last month I landed some Wuffs tooling changes re animated PNG, to be better able to (as a comparison baseline) decode and test animated WebP.

A lot of that manual rewrite has been committed, but the SIMD parts haven't landed yet. So yes, for what's on the main branch (not the PR), performance is not as good as libwebp yet, but landing the SIMD parts should fix that.

> wuffs is not bit-identical to libwebp

That's news to me. PR 168 says it produces pixel-identical output to libwebp. And what's in the main branch aims to be pixel-identical, e.g. for YUV to RGB conversion, it implements libwebp's formulae, not libjpeg's formulae. Both use BT.601, but libwebp uses studio range and libjpeg uses full range.

Can you link to some example .webp images that are not bit-identical?

1. Decode to 16x16. This is done by the Handsum library.

2. Upscale 2x. This is done by the caller of the Handsum library (or, for the Wasm demo, by the HTML renderer).

If you want 48x48 or 64x64, just raise the 2x, in step 2, to 3x or 4x.

Comments URL: https://news.ycombinator.com/item?id=48837153

Points: 11

# Comments: 0

https://github.com/rsc/fpfmt/blob/main/bench/uscalec/ftoa.c

There's also a "wget some JSON and pipe that to what Wuffs calls example/jsonptr" example at https://nigeltao.github.io/blog/2020/jsonptr.html#sandboxing

> Kaitai Struct is in a similar space, generating safe parsers for multiple target programming languages from one declarative specification. Again, Wuffs differs in that it is a complete (and performant) end to end implementation, not just for the structured parts of a file format. Repeating a point in the previous paragraph, the difficulty in decoding the GIF format isn't in the regularly-expressible part of the format, it's in the LZW compression. Kaitai's GIF parser returns the compressed LZW data as an opaque blob.

Taking PNG as an example, Kaitai will tell you the image's metadata (including width and height) and that the compressed pixels are in the such-and-such part of the file. But unlike Wuffs, Kaitai doesn't actually decode the compressed pixels.

---

Wuffs' generated C code also doesn't need any capabilities, including the ability to malloc or free. Its example/mzcat program (equivalent to /bin/bzcat or /bin/zcat, for decoding BZIP2 or GZIP) self-imposes a SECCOMP_MODE_STRICT sandbox, which is so restrictive (and secure!) that it prohibits any syscalls other than read, write, _exit and sigreturn.

(I am the Wuffs author.)

> Software has a Peter Principle. If a piece of code is comprehensible, someone will extend it, so they can apply it to their own problem. If it’s incomprehensible, they’ll write their own code instead. Code tends to be extended to its level of incomprehensibility.

https://github.com/nigeltao/qoir has some numbers comparing QOIR (which is QOI-inspired-with-LZ4) vs PNG.

QOIR has better decode speed and comparable compression ratio (depending on which PNG encoder you use).

QOIR's numbers are also roughly similar to ZPNG.

    pub status "#blah"
    pub struct foo(etc etc)
    pri func foo.bar(etc etc)

    rg -N ^p   std/jpeg/*.wuffs

    rg -N ^pub std/jpeg/*.wuffs

https://nigeltao.github.io/blog/2022/gamma-aware-ordered-dit...

https://nigeltao.github.io/blog/2021/json-with-commas-commen...

JWCC is also what Tailscale call HuJSON, as in "JSON for Humans", which as amusingly also what json5 claims to be.

https://github.com/tailscale/hujson

This is possibly getting too academic, but IIUC for a progressive JPEG, e.g. encoded by cjpeg to have 10 0xDA Start Of Scan markers, it's actually legitimate to post-process the file, truncating to fewer scans (but re-appending the 0xD9 End Of Image marker). The shorter file is still a valid JPEG, and so still relevant for discussing whether all decoders will render the same pixels.

I might be wrong about validity, though. It's been a while since I've studied the JPEG spec.

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...

Still, Wuffs doesn't use autotools, and if you're pulling the library from the https://github.com/google/wuffs-mirror-release-c repository then that repo doesn't even contain any binary-data test files.

    $ git clone --quiet --depth=1 https://github.com/google/wuffs.git
    $ gcc -O3 wuffs/example/mzcat/mzcat.c -o my-mzcat
    $ ./my-mzcat     < linux-6.8.2.tar.xz | sha256sum 
    d53c712611ea6cb5acaf6627a84d5226692ae90ce41ee599fcc3203e7f8aa359  -
    $ /usr/bin/xz -d < linux-6.8.2.tar.xz | sha256sum 
    d53c712611ea6cb5acaf6627a84d5226692ae90ce41ee599fcc3203e7f8aa359  -

Decoding takes a compressed image file as input, which have complicated formats. Roughly speaking, encoding just takes a width x height x 4 pixel buffer, with very regular structure. It's much easier to hide something malicious in a complicated format.

Higher priority means that, when deciding whether to work on a Wuffs PNG encoder or a Wuffs JPEG decoder next, when neither existed at the time, I chose to have more decoders.

(I work at Google, and am the Wuffs author, but have nothing to do with Jpegli. Google is indeed a big company.)