Kostya's Boring Codec World

So, despite work, heat, travels, and overall laziness, I’ve managed to complete more or less full-featured Indeo 4 and 5 decoder. That means that my own decoder decodes Indeo 4 and 5 files with all known frame types (including B-frames) and transforms (except DCT because there are no known samples using it) and even transparency!

Here are two random samples from Civilization II and Starship Titanic decoded and dumped as PGM (click for full size):

I’m not going to share the code for transparency plane decoding, it’s very simple (just RLE) and the binary specification is easy to read. The only gotchas are that it’s decoded as contiguous tile aligned to width of 32 (e.g. the first sample has width 332 pixels but the transparency tile is 352 pixels) and the dirty rectangles provided in the band header are just a hint for the end user, not a thing used in decoding.

This decoder was written mostly so that I can understand Indeo better and what can I say about it: Indeo 4/5 is about the same codec with some features fit for more advanced codecs of the later era. While the only things it reuses from the previous frames are pixels and band transform mode, it can reuse decoded quantisers and motion vectors from the first band for chroma bands and luma bands 1-3 in scalability mode too. It has variable block sizes (4×4, 8×8 and 8×8 in 16×16 macroblock) with various selectable transforms and scans (i.e. you can have 2D, row or column Slant, Haar or (theoretically) DCT and scans can be diagonal, horizontal or vertical too). And there were several frame types too: normal I-, P- and B-frames, droppable I- and P-frames, and droppable P-frame sequence (i.e. P-frames that reference the previous frame of such type or normal I/P-frame). Had it had proper stereo support, it’d be still as hot as ITU H.EVC.

The internal design between Indeo 4 and 5 differs in small details, like Indeo 4 having more frame types (like B-frames and droppable I-frames) — but Indeo 5 had introduced droppable P-frame sequence; picture and band headers differ between versions but (macro)block information and actual content decoding is the same (Indeo 5 does a bit trickier stuff with macroblock quantisers but that’s all). Also Indeo 4 had transparency information and different plane reconstruction (using Haar wavelet instead of 5/7 used in Indeo 5). So, in result my decoder was split into several modules reflecting the changes: indeo4.rs and indeo5.rs for codec-specific functions, ivi.rs for common structures and types (e.g. picture header, frame type and such), ividsp.rs for transforms and motion compensation and ivibr.rs for the actual decoding functions.

As with Intel H.263 decoder, Indeo 4/5 decoders provide implementations for IndeoXParser that parse picture header, band header and macroblock information and also recombine back plane in case it was coded as scalable. In result they store not so much information, just the codebooks used in decoding and for Indeo5 the common picture information that is stored only for I-frames (in other words, GOP info).

In result, here’s how Indeo 4 main decoding function looks like:

    fn decode(&mut self, pkt: &NAPacket) -> DecoderResult<NAFrameRef> {
        let src = pkt.get_buffer();
        let mut br = BitReader::new(src.as_slice(), src.len(), BitReaderMode::LE);

        let mut ip = Indeo4Parser::new();
        let bufinfo = self.dec.decode_frame(&mut ip, &mut br)?;
        let mut frm = NAFrame::new_from_pkt(pkt, self.info.clone(), bufinfo);
        frm.set_keyframe(self.dec.is_intra());
        frm.set_frame_type(self.dec.get_frame_type());
        Ok(Rc::new(RefCell::new(frm)))
    }

with the actual interface for parser being

pub trait IndeoXParser {
    fn decode_picture_header(&mut self, br: &mut BitReader) -> DecoderResult<PictureHeader>;
    fn decode_band_header(&mut self, br: &mut BitReader, pic_hdr: &PictureHeader, plane: usize, band: usize) -> DecoderResult<BandHeader>;
    fn decode_mb_info(&mut self, br: &mut BitReader, pic_hdr: &PictureHeader, band_hdr: &BandHeader, tile: &mut IVITile, ref_tile: Option<Ref<IVITile>>, mv_scale: u8) -> DecoderResult<()>;
    fn recombine_plane(&mut self, src: &[i16], sstride: usize, dst: &mut [u8], dstride: usize, w: usize, h: usize);
}

And the nano-benchmarks:
the longest Indeo4 file I have around (00186002.avi) — nihav-tool 20sec, avconv 9sec plus lots of error messages;
Mask of Eternity opening (Indeo 5) — nihav-tool 8.1sec, avconv 4.1sec.
Return to Krondor intro (Indeo 5) — nihav-tool 5.8sec, avconv 2.9sec.
For other files it’s also consistently about two times slower but whatever, I was not trying to make it fast, I tried to make it work.

The next post should be either about the things that irritate me in Rust and make it not so good for codec implementing or about cooking.

Obviously it moves very slowly: I spend most of my time on work, sleep, cooking and travelling around. Plus it was too hot to think or do anything productive.

Anyway, I’ve completed IMC/IAC decoder for NihAV. In case you’ve forgotten or didn’t care to find out at all, the names stand for Intel Music Coder and Indeo Audio software with IAC being slightly upgraded version of IMC that allows stereo and has tables calculated for every supported sample rate instead of the set of them precalculated for 22kHz. And despite what you might think it is rather complex audio codec that took a route of D*lby AC-3, G.722.1/RealAudio Cooker and CELT—parametric bit allocation codecs. It’s the kind of audio codecs that I dislike less than speech codecs but more than the rest because they have large and complex function that calculates how many bits/values should be spent on each individual coefficient or subband. In IMC/IAC case it gets even worse since the codec uses floating point numbers so the results are somewhat unstable between implementations and platforms (a bit more on that later). Oh, and this codec has I- and P-frames since some blocks are coded as independent and others are coded using information from the previous block.

Rust does not have much to do with C so you cannot simply copy-paste code and expect it to work and it’s against the principles of the project anyway. Side note: the only annoying Rust feature so far is array initialisation, I’d like to be able to fill array in a loop before using it without initialising array contents to some default value (which I can’t do for some types) or resorting to mem::uninitialized() and ptr::write(). Anyway, I had to implement my own version of the code so it’s structured a bit differently, has different names, uses bitstream reader in MSB16LE mode instead of block swapping and decodes most files I could test without errors unlike libavcodec—so it’s NIH all the way!

I wasted time mostly on validating my code against the binary specifications so this version actually decodes most files as intended while libavcodec fails to do that. To describe the problem briefly, it all comes from the same place: the codec first produces bit allocation for all bits still available then determines how to read flags for skipping coefficients in some bands, reads those flags and adjusts bit allocation for the number of bits freed by this operation; the problem is that bit allocation may go wrong and in result skip flags take more bits than the coefficients that would be coded otherwise and decoder would fail to adjust bit allocation for that case (it’s not supposed to do that in the specification) and will read more bits than the block contains. For the thirty-something IMC and IAC in AVI samples only one fails now for me because in bit allocation the wrong band gets selected for coefficient length decreasing. And the reason is the difference in the fourth or fifth digit after the decimal point in one array of constants that makes the wrong value minimum (and thus selected for coefficients length decreasing). Since it takes several minutes with gdb+mplayer2 to get information at this point (about at 10-second position in 14-second audio) I decided not to dig further.

Also I had to write other pieces of code like split-radix FFT, byte writer and WAV dumper that accepts audio packets and writes them with the provided ByteWriter.

P.S. Nanobenchmarks ahoy: decoding the longest IMC stream that I had (a bit more than two minutes) takes 0.124s with avconv and 0.09s with nihav-tool. Actual decoding functions take about the same time though Rust implementation is still faster by couple percents and my FFT implementation is slower (but on the other hoof it’s called for every frame since it decodes that file without errors).

P.P.S. So next is Indeo 4/5 with all wonderful features like scalable decoding, B-frames and transparency (that reminds me that Libav and ScummVM had a competition who would be the last to implement proper transparency support for Indeo 4, now they both might win). And then I’d probably go back to implementing the features I wanted: being able to tell the demuxer to discard and don’t demux certain streams, better streams reporting from the demuxer, seeking and decoder reset, frame reordering functionality, maybe WAV support too. And then maybe back to decoders. I want to have several codec families fully implemented, like RAD (Smacker, Bink and Bink2), Duck/On2 (TM1, TM-RT, TM2, TM2X, TM VP3, VP4, VP5, AVC, VP6 and VP7) and RealMedia (again). But I’m not in a hurry.

P.P.P.S. I’m not going to publish all source code but bits of it may be either posted when relevant or leaked to rust-av, its developer(s) has(have) shown some interest, so enquire there.