Kostya's Boring Codec World

(luckily there’s not much left of this month of intermediate codecs)

So I’ve looked at another intermediate codec, post title hints on both its name and design. Coding scheme is rather simple: you code lines either in raw form or with prediction (from the left neighbour for the top line or (3 * L + 3 * T - 2 * TL) >> 2 for other lines, prediction error is coded with fixed Huffman codes.

Simple, right?

Here’s the catch: there is an insane number of formats it supports, both for storage and output and there’s an insane number of decoding functions for decoding format X into format Y.

So quite probably no decoder — not interesting and too tedious.

I’ve finally brought myself into looking at alpha plane decoding support for ProRes. It was a bit peculiar but rather easy to reverse engineer. Now I only need to update my ConsumerRes decoder to support it.

And that’s probably enough for the month of intermediate codecs.

The adjective is referring to the hype that the company that made this codec is run by designers (unlike some other companies where even design is made by developers or — even worse — marketers). And let’s call it AWIC or iNtermediate codec for short. Let’s not mention its name at all.

It is a rather old codec and it codes 8-bit YUV420 in 16×16 macroblocks with DCT, quantisation and static codes. Frame is divided into slices in such way so that there are not more than 32 slices on one line (and slice height is one macroblock). The main peculiarity is having scalable mode — every macroblock is partitioned into 8×8 sub-macroblock (i.e. 8×8 luma block and two 4×4 chroma blocks) with the following data for the rest of the block, and this is exploited for decoding frames in half-width, half-height or half-width half-height modes.

Maybe I’ll write a decoder for it after all.

Astrologers proclaim month of intermediate codecs. Number of blog posts about intermediate codecs doubles! (from zero)

Let’s look at Canopus codecs and their development.

Canopus Lossless

This is very simple lossless video codec: you have code tables description and coded difference from the left neighbour (or top one for the first pixel) for each component. For RGBA and YUV there are slight improvements but the overall coding remains the same.

Canopus HQ

This is an ordinary intermediate codec employing IDCT with 16×16 macroblocks in 4:2:2 format and interlacing support.
It has predefined profiles with frame sizes (from 160×120 to 1920×1080), number of slices and macroblock shuffling order (yes, like DV it decodes macroblocks in shuffled order).

Block coding is nothing special but quantising is. For every macroblock there can be one of sixteen quantiser sets selected. And for each block one of four quantising matrices can be selected from that set. Of course there are different quantiser matrices for luma and chroma (i.e. 128 quantising matrices in total, about 80 of them are unique).

Interlacing is signalled per block (in case it’s enabled for the frame).

Canopus HQA

This is Canopus HQ with alpha support. The main differences are flexible frame size (no hardcoded profiles), alpha component in macroblocks and coded block pattern. Coding and tables seem to be the same as in HQ.

Coded block pattern specifies which of 4 luma blocks are coded (along with corresponding alpha and chroma blocks). Uncoded blocks are filled with zeroes (i.e. totally transparent).

Canopus HQX

This codec combines both previous codecs and extends them for more formats support. While HQ was 4:2:2 8-bit, this one can be 4:2:2 or 4:4:4 and 9-, 10- or 11-bit support (with or without alpha).

There are changes in overall and block coding.

Frame is now partitioned into slices of 480 macroblocks and every 16 macroblocks are shuffled.

Blocks now have more adaptive coding. DCs are coded as the differences to the previous ones (inside macroblock component) and instead of being coded as 9-bit number they are now Huffman-coded and table is selected depending on component bit depth. Quantisation is split: now there’s a selectable quantiser and two quantiser matrices (for luma/alpha and chroma). AC codes are selected depending on quantiser selected for the block. So there are less quantiser matrices (two instead of seventy eight) but more VLC tables (CBP + 3 DC + 6 AC tables instead of CBP and single AC table).

Conclusion

Reverse engineering all those formats was obvious because they are not complex, obfuscated or C++ (which is usually both).

Shall I write decoders for them? Unlikely. The codecs are not too interesting (I’ve seen only one Canopus HQ sample and no HQA or HQX samples at all) and rather tedious to implement because of all those tables. And we have Canopus Lossless decoder already.

So, as it happens to me sometimes, I’ve looked at some random codec before fixing myself a breakfast.

What do we have here?

• 8×8 DCT
• lots of static VLCs (not as much as Real codecs have though)
• several coding modes — 4:2:2, 4:4:4, both with alpha or not
• block-level interlacing

Macroblock consists of 4 luma blocks, 4 or 8 chroma blocks and optionally 4 alpha blocks. In the latter case not all blocks might be coded and CBP is present.

There are separate VLC tables for DC depending on bit depth (10, 11 or 12 bits per component) and for quantised ACs (for quantisers 0-7, 8-15, 16-31, 32-63, 64-127 and the rest).

The only interesting scheme is quantiser selection: there is a table with predefined quantiser quadruplets. Every macroblock can select any quadruplet and every block can use any quantiser from it.

Oh, and I think codec’s name was Canopus HQX.