Kostya's Boring Codec World

This post is an answer to the comment by Alex Converse on my previous post:

It’s interesting how quickly you dismiss SLS for being a hybrid coder with AAC. From a pure lossless standpoint that is a weakness but from a broader perspective it allows for a lossy layer that is widely compatible with existing hardware.

Let’s see why scalable coding is a weakness from lossless coding standpoint.

There are few hybrid lossy+lossless codecs out there which use lossy part in lossless reconstruction, e.g. MPEG-4 SLS, DTS-HD MA and WavPack. First two use two different coding techniques – MDCT or QMF for core coding and usual lossless coding for difference. In WavPack both parts are coded in the same way and correction data is stored in different block. For DCT-based codecs there are many ways of performing DCT (from trivial matrix multiplication to FFT-based implementation to decomposing DCT into smaller size DCTs) which may lead to slightly different output depending on method chosen. Thus, you should have a reference way (i.e. not the fastest one) of doing lossy stuff or you can’t guarantee truly lossless reconstruction. Also residue (the difference between original and lossy coded signal) tends to be more chaotic in this case and thus less compressible.

Another issue is what to do with correction data. If you put it into a separate file, you will have more troubles since you have to manage two files; if you put it all into single file, its size will be bigger than pure lossless coded file (unless you have very stupid method of lossless coding).

And now comes an argument that I really hate: “but it allows legacy players handle those files”. That, in my opinion, is this post title. Making it backward compatible just cripples it. In that case you need to implement new (and sometimes completely different features) in old limits and relying to new technology. So in some case it just degrades quality and/or forces you to encode something twice — for old feature set and its replacement. Another reason is that it just delays that codec adoption: old player can play it so why should I bother about this new codec support? I suspect this was a reason why we have MLP but no DTS-HD support.

The worst offender here is MP3. This codec sucks by design. It uses 36-point (or three 12-point) MDCTs which are not trivial to speed-up unlike power-of-two transforms and the output of MDCTs is used as input to QMF used in MPEG Audio layers I&II, as it claimed “to be compatible with them”. As claimed here, MP3 would perform better and since it comes from one of the leading LAME developers, I believe it. And of course MP3Pro. Most players in existence just ignore extension part and play crippled version of sound. Someone may argue that’s because it’s proprietary. Okay, look at HE-AAC where SBR is documented at least, it may still cause some confusion since it may be detected only when decoding audio frame.

In my opinion both implementing new codec support and special codec extension in general case is just single-time action with comparable effort (hacking existing code for new extension support and detection may be not that easy). And thus, adding a new codec should be preferred. MPEG-2 introduced both AAC (please look how it was called back then) and multichannel extensions to layers I-III. Guess which one works better?

A quick glance on some codec disassembly inspired me to write this post.

So today I talk about how to design perfectly awful codec (from FFmpeg decoder implementer’s point of view). Since audio and video codecs usually have some specific methods and approaches to design, it will be presented in two parts.

Video Codec Design (and why we don’t have a decoder for this codec in FFmpeg)

• Don’t care about portability. The “best” example is Lagarith — lossless video codec that uses floating point variable for arithmetic coder state. Thus, decoding it on anything but x86 requires an 8087 emulator.
• Tie it to specific API or OS. The codec mentioned at the beginning provides the best example: it stores actually a sequence of GDI commands for frame data. While storing, say, VNC protocol record may provide good lossless compression, it should be self-sufficient (i.e. it should not require external data). M$ Camcorder Video however has (and uses!) such wonderful commands as “draw text with provided font parameters (including font name)”. Thanks, I’m not going to work on decoder for that, ask those guys instead.
• Use lots of data. It really pisses decoder developer when you have to deal with lots of tables, especially with non-obvious structure. Special thanks to RealVideo 3 and 4 which stored variable-length codes data in three ways and about a hundred of codebooks.
• Use your own format. That one annoys users as well. Isn’t it nice when your video is stored in videofile.wtf that can be played only with provided player (and who knows if it can be converted at all). Sometimes this has its reasons — for game formats, for example — though this makes life of decoder developer a bit harder.

Audio Codec Design (and why nobody cares about this codec)

Let’s repeat last two items:

• Use lots of data. Yes, there are codecs that use lots of tables during decoding. The best supporters of this policy are DTS (they even decided to skip tables with more than ten thousand elements in ETSI specification, extensions require few more tables) and TwinVQ/VQF that has even more tables.
• Use your own format. Audio codec authors like to invent new formats that can be used only with their codecs. There is one example when such container format was extended to store other codecs as well. That’s infamous Ogg. If you think it’s nice then try implementing demuxer for it from the scratch.

But wait, there are more tricks!

• Containers are overrated. The best example is Musepack SV7 and earlier. That codec is known to store frames continuously and when I say “continuously”, I mean it — if one frame ends inside byte, new frame starts from the next bit. And the only way to know frame size is to decode it. And if your file is corrupted in the middle, the rest of it would be undecodable. A mild version of this is MPEG audio layer-III which stores audio data disregarding actual frame boundaries.
• Really tie codec to container. That would be Musepack SV8 now. This time they’ve designed almost sane container with only one small catch — last frame actually encodes less samples and the only way to know that would be to make demuxer somehow signal decoder number of samples to decode for each frame. If you don’t do that, you may unexpectedly get some nasty decoding errors.
• Change bitstream format often. If you throw out backward compatibility you may end with many decoders needed for each case. An example is CELT — it’s still experimental and changes bitstream format often, thus storing files in that format would be just silly since next version of decoder won’t be able to read them.
• Hack extensions into bitstream. Some codecs are known to contain extension data inside frame data for “backwards compatibility” so decoders usually have hard time finding it and verifying it’s really expected extension data instead of some garbage. Well-known examples are MP3Pro and DTS (which took it to extreme — there are extensions for both frequency and additional channels that can be present simultaneously; luckily, DTS-HD has it more structured inside an extension frame data).
• Make it unsuitable for general uses. For example, make codec take unbounded or potentially too large amounts of memory (Ogg Vorbis does that) or
• Make codec like a synonym for reference implementation. It’s good when you just make only one implementation and just change it in many subtle ways so later you need to reverse engineer the source to get specification. That was the case with binary M$ Office formats and it seems to be the case with Speex (at least I heard so).

And finally, The Lossless Audio Codec to serve an example to them all. As Måns put it, “wherever your talk about bad design of codecs, there’s always Monkey’s Audio”. Let’s see its advantages:

• container — it has custom container of course. And there’s one minor detail: it packs data into 32-bit little-endian words and frame may start at any byte of that word. This makes it somehow combine both approaches to containers.
• Bitstream format changes — check. It is known to have a lot of small tweaks making bitstream format incompatible. Some of them are actually container-related though.
• Unusable for general uses — well, it’s famous for requiring more CPU power to decode than most low-resolution (up to SD) H.264 video streams.
• One codec, one implementation — for a long time it was so until some Rockbox developer REd the source code and wrote his own decoder (FFmpeg decoder is derived from it). Also for quite a long time it was supported only on Windows. And it doesn’t support older versions — nobody bothered to extend support for them.