Archive for the ‘NihAV’ Category

NihAV: updated for Rust 1.69

Thursday, July 27th, 2023

Since I had nothing better to do I decided to optimise my H.264 decoder a bit more, and that required a rather recent version of rustc that supports sym construct in asm!{} (so I can reference data tables in the inline assembly). Why this specific version though? I picked whatever was both recent enough to support the aforementioned feature (and older version had multiple micro version releases which hints on some problems with them) and not too recent either (again, I’m no beta tester of the compiler and I don’t need other shiny features).

And while at it I decided to make the code a bit more up to date. cargo-clippy is still annoying with its default warning about all-caps names and some lints that changed names and their suppressors no longer work. Getting rid of some leftover hints for the old versions of the compiler (like explicit drop()s for the objects borrowing code and some type hints) was nice though. Inline assembly is still only halfway done, especially considering that using const in it won’t be possible in stable for a long time and sym sucks compared to GCC inline assembly (it provides just a symbol name and you should magically know for yourself how the target platform works in order to make it possible to load it correctly; on AMD64 it’s rather simple but on aarch64 and on 32-bit ARMs that depends on target OS and PIC mode). Who would’ve thought that assembly may be platform-dependent! Looks like the current solution to that problem is to expose current configuration to the user so it’s up to you to check all environment variables and write the appropriate code. And of course even that solution will be available some time in the future since the developers haven’t thought about it at all.

Anyway, now my H.264 decoder features some more assembly optimisations and decodes video even faster than before. Though I fear it still takes too much CPU for the comfortable playback of my typical content so I’ll have to dabble in the hardware video acceleration. NihAV is a learning project after all.

NihAV: another step to being self-sufficient

Tuesday, June 13th, 2023

I’ve mentioned previously that I played with my H.264 decoder trying to make it multi-threaded. Now I went a bit further and plugged it into my video player. So now instead of hopelessly lagging on 720p video it can play it in real time just fine—so after improving my player even further (and enabling assembly optimisations when Rust compiler is good enough for that) I can use it to play most of the videos I care about without resorting to the external decoders or players. And in theory using it more will lead to fixing and polishing it more thus forming a stable loop.

Anyway, the code is not public yet as I hacked this new encoder in a separate crate and I still need to merge it back and clean up a bit, but I’d like to describe the interfaces and my reasons behind them.

So, multi-threaded decoder has a separate interface (for obvious reasons). I thought about writing a wrapper for single-threaded decoders to behave like multi-threaded ones but decided against it (at least for now). NADecoderMT has the following methods:

  • init()—initialises the decoder. One of the parameters is number of threads to use. IMO it’s the caller that decides how many threads it can spare as the decoder does not know what will be done in parallel (maybe there’s another multi-threaded decoder or two are running);
  • can_take_input()—queries if the decoder is ready to queue the next frame for decoding. Of course you can call queue_pkt() and check if it accepted the input but it may not always be desired (e.g. if we need to retrieve an input packet and then hold it waiting until the decoder is ready to accept it);
  • queue_pkt()—tries to queue the next frame for decoding;
  • has_output()—checks if the decoder has produced some frames for the output. Since get_frame() is waiting for a frame to be decoded this function is necessary unless you want to block the thread calling the decoder;
  • get_frame()—waits until at least one frame is decoded and returns it (or a special error if there are no frames to be decoded);
  • flush()—stops decoding all frames and clears the state (e.g. after seek).

Another peculiarity of this decoder interface is that it operates on pairs of a frame and its sequential number. The reason is simple: you get decoded frames out of order so you need to distinguish them somehow (and in case of a decoding error we need to know which frame caused it).

This also leads to a special frame reorder mechanism for such codecs. I’ve created MTFrameReorderer that requires you to “register” frame for decoding (providing you with an ID that is fed to the decoder along with frame data) and to “unregister” frame on error (that’s one of the places where returned frame ID comes in handy). Unfortunately it’s not possible to create a generic reorderer that would a) work completely codec-agnostic b) not require a whole file (or an indefinitely long sequence of frames) to be buffered before output and c) produce monotone increasing sequence of frames. Considering how H.264 has no real concept of frames and can build a pyramid of referenced frames adding layer by layer (and mind you, some frames may have an error during decoding and thus not present in output). I simply gave up and made a heuristic that checks if we have enough initial frames decoded and outputs some of them if it’s possible. At least it seems to work rather fine on the conformance suite (except for a couple of specially crafter files but oh well).

Maybe in the future I’ll try more multi-threaded decoders but for now even one decoder is enough, especially such practical one. Still, I need to find something more interesting to do.

Further Cinepak experiments

Monday, June 5th, 2023

For having nothing better to do I kept experimenting with Cinepak encoder.

I considered implementing some variant of codebook decomposition scheme suggested by Tomas in the comments to the previous post but I’m still not sure if I should bother even if it looks promising. So I tried the old thresholds-based scheme instead.

And what do you know, it speeds things up considerably: my usual test sample gets encoded in 27-35 seconds (depending on thresholds) instead of 44 seconds in the usual mode. But since I don’t know what would be good thresholds I did the opposite and added a refinement mode: after deciding which codebook to use for which block I re-generate codebook using only those blocks that belong to it. Of course it increases processing time, for example that file it takes 75 seconds to encode with refinement—which is still 70% more time but still less than double (for comparison, in full ELBG mode it’s an increase from about 160 seconds to 270 seconds).

So by rough estimate selecting only relevant blocks for codebook generation shaves 20-40% off the encoding time. And splitting data into partitions and generating a codebook by parts made the process about three times faster. I suspect that with a proper approach to clustering vector quantisation can be made two-three times faster but I don’t think I want to experiment with that. I should call it a day and move to something else instead.

Quick experiments with Cinepak encoder vector quantisation

Saturday, June 3rd, 2023

Out of curiosity I decided to check how partitioning input before creating a codebook affects encoding speed. So I’ve added a mode to Cinepak encoder that partitions vectors by luma variance and creates a part of common codebook just for them. The other two modes are median cut (the simplest one but also with mediocre output) and ELBG (that uses median cut to create the initial codebook—also if it’s not full that means we have all possible entries and do not need to perform ELBG at all).

Here are rough results on encoding several different files (and using different number of strips): median cut worked for 11-14 seconds, ELBG took 110-160 seconds, new mode (I decided to call it fast) takes 43-62 seconds. I think even such approximate numbers speak for themselves. Also there’s an interesting side effect: because of the partitioning it tends to produce smaller codebooks overall.

And while we’re speaking about quantisation results, here’s the first frame of waterfall sample encoded in different modes:

median cut

fast

full ELBG

As you can see, median cut produces not so good images but maybe those artefacts will make people think about the original Cinepak more. Fast mode is much nicer but it still has some artefacts (just look at the left edge of the waterfall) but if you don’t pay too much attention it’s not much worse than full ELBG.

Are there ways to improve it even further? Definitely. For starters, the original encoder exploits the previous codebook to create a new one while my encoder always generates a new codebook from scratch (in theory I can skip median cut stage for inter strips but I suspect that ELBG will work much longer in this case). The second way is to fasten up the ELBG itself. From what I could see it spends most of the time determining to which cluster each of the points belong. By having some smarter structure (something like k-d tree and some caching to skip recalculating certain clusters altogether) it should be possible to speed it up in several times. Unfortunately in this case I value clarity more so I’ll leave it as is.

P.S. I may also try to see how using thresholds and block variance to decide its coding mode affects the speed and quality (as in this case we first decide how to code blocks and then form codebooks for them instead of forming codebooks first and then deciding which mode suits the current block better; and in this case we’ll have smaller sets to make codebooks from too). But I may do something different instead. Or nothing at all.

NihAV experiments: multi-threaded decoder

Thursday, June 1st, 2023

In my efforts to have an independent player (that relies on third-party libraries merely for doing input and output while the demuxing and decoding is done purely by NihAV) I had to explore the way of writing a multi-threaded H.264 decoder. And while it’s not working perfectly it’s a good proof of a concept. Here I’ll describe how I hacked my existing decoder to support multi-threading.
(more…)

rv4enc: probably done

Saturday, May 13th, 2023

In one of the previous posts I said that this encoder will likely keep me occupied for a long time. Considering how bad was that estimation I must be a programmer.

Anyway, there were four main issues to be resolved: compatibility with the reference player, B-frame selection and performing motion estimation for interpolated macroblocks in them, and rate control.

I gave up on the compatibility. The reference player is unwieldy and I’d rather not run it at all let alone debug it. Nowadays the majority of players use my decoder anyway and the produced videos seem to play fine with it.

The question of motion vector search for interpolated macroblocks was discusses in the previous post. The solution is there but it slows down encoding by several times. As a side note, by omitting intra 4×4 mode in B-frames I’ve got a significant speed-up (ten to thirty percent depending on quantiser) so I decided to keep it this way by default.

The last two issues were resolved with the same trick: estimating frame complexity. This is done in a relatively simple way: calculate SATD (sum of absolute values of Hadamard-transformed block) of the differences between current and some previous frame with motion compensation applied. For speed reasons you can downsample those frames and use a simpler motion search (like with pixel-precision only). And then you can use calculated value to estimate some frame properties.

For example, if the difference between frames 0 and 1 is about the same as the difference between frames 1 and 2 then frame 1 should probably be coded as B-frame. I’ve implemented it as a simple dynamic frame selector that allows one B-frame between reference frames (it can be extended to allow several B-frames but I didn’t bother) and it improved coding compared to the fixed frame order.

Additionally there seems to be a correlation between frame complexity and output frame size (also depending on the quantiser of course). So I reworked rate control system to rely on those factors to select the quantiser for I- and P-frames (adjusting them if the predicted and the actual sizes differ too much). B-frames simply use P-frame quantiser plus constant offset. The system seems to work rather well except that it tends to assign too high quantisers for some frames, resulting in rather crisp I-frame followed by more and more blurry frames.

I suppose I’ll play with it for a week or two, hopefully improving it a bit, and then I shall commit it and move to something else.

P.S. the main goal of NihAV is to provide me with a playground for learning and testing new ideas. If it becomes useful beside that, that’s a bonus (for example, I’m mostly using nihav-sndplay to play audio nowadays). So RealVideo 4 encoder has served its purpose by allowing me to play more with various concepts related to B-frames and rate control (plus there were some other tricks). Even if its output makes RealPlayer hang, even if it’s slow—that does not matter much as I’m not going to use it myself and nobody else is going to use it either (VP6 encoder had some initial burst of interest from some people but none afterwards, and nobody cares about RV4 from the start).

Now the challenge is to find myself an interesting task, because most of the tasks I can think about involve improving some encoder or decoder or—shudder—writing a MOV/MP4 muxer. Oh well, I hope I’ll come with something regardless.

Starting yet another failure of an encoder

Thursday, April 6th, 2023

As anybody could’ve guessed from Cook encoder, I’d not want to stop on that and do some video encoder to accompany it. So here I declare that I’m starting working on RealVideo 4 encoder (again, making it public should prevent me from chickening out).

I can salvage up some parts from my VP7 encoder but there are several things that make it different enough from it (beside the bitstream coding): slices and B-frames. Historically RealMedia packets are limited to 64kB and they should not contain partial slices (grouping several slices or even frames in the packet is fine though), so the frame should be split during coding. And while Duck codecs re-invent B-frames to make them still be coded in sequence, RealVideo 4 has honest B-frames that should be reordered before encoding.

So while the core is pretty straightforward (try different coding modes for each macroblock, pick the best one, write bitstream), it gives me enough opportunity to try different aspects of H.264 encoding that I had no reason to care about previously. Maybe I’ll try to see if automatic frame type selection makes sense, maybe I’ll experiment with more advanced motion search algorithms, maybe I’ll try better heuristics for e.g. quantiser selection.

There should be a lot to keep me occupied (but I expect to spend even more time on evading that task for the lack of inspiration or a sheer amount of work to do demotivating me).

NihAV: towards RealMedia encoding support

Friday, March 10th, 2023

Since I have nothing better to do with my life, I’ve decided to add a rudimentary support for encoding into RealMedia. I’ve written a somewhat working RMVB muxer (it lacks logical stream support that is used to describe the streaming capabilities, it also has some quirks in audio and video support but it seems to produce something that other players understand) and now I’ve made a Cook audio encoder as well.

Somebody who knows me knows that I fail spectacularly at writing non-trivial lossy audio encoders but luckily here we have a format which even I can’t botch much. It is based on parametric bit-allocation derived from band energies. So all it takes is to perform slightly modified MDCT, calculate band energies, convert them to scale factors, perform bit allocation based on them, pack band contents depending on band categories and adjust the categories until all bands fit the frame. All of these steps are well-defined (including the order in which bands should be adjusted) so making it all work is rather trivial. But what about determining the frame size and coupling mode?

As it turns out, RealMedia supports only certain codec profiles (or “flavors” in its terminology), so Cook has about 32 different flavours defined for different combinations of sample rates, number of channels and bitrates. And each flavour has an internally bitrate parameters (frame size and which coupling parameters to use) for each channel pair so you just pick a fitting profile and go on with it. In theory it’s possible to add a custom profile but it’s not worth the hassle IMO.

And now here are some fun facts about it. Apparently there are three internal revisions of the codec: the original version for RealMedia G2, version 2 for RealMedia 8 and multichannel encoder (introduced in RealMedia 10, when they’ve switched to AAC already). Version 2 is the one supporting joint-stereo coding. The codec is based on G.722.1 (the predecessor of CELT even if Xiph folks keep denying that) but, because Cook frames are 256-1024 samples instead of fixed 320-sample frames, they’ve introduced gains to adjust better for volume changes inside the frame (but I haven’t bothered implementing that part). That is probably the biggest principal change in the format (not counting the different frame sizes and joint stereo support in the version 2). And finally I should probably mention that there are some flavours with the same bitrate that differ by the frequency range and where the joint stereo part starts (some of those are called “high response music” whatever that means).

Time to move to the video encoder…

Indeo 3 encoder: done

Thursday, February 16th, 2023

After fixing some bugs I think my encoder requires no further improvement (there are still things left out there to improve but they are not necessary). The only annoying problem is that decoding some videos with the original VfW binary gives artefacts. Looks like this happens because of its peculiar way to generate and then handle codebooks: corrector pairs and quads are stored as single 32-bit word and its top bit is also used to signal if it’s a dyad or a quad. And looks like for some values (I suspect that for large negative ones) the system does not work so great so while e.g. XAnim module does what is expected, here it mistakes a result for another corrector type and decodes the following bytestream in a wrong way. Of course I’m not going to deal with that annoyance and I doubt anybody will care.

Also I’ve pruned one stupid bug from my MS Video 1 encoder so it should be done as well. The third one, Cinepak, is in laughable state (well, it encodes data but it does not do that correctly let alone effectively); hopefully I’ll work on it later.

For now, as I see no interesting formats to support (suggestions are always welcome BTW), I’ll keep writing toy encoders that nobody will use (and hopefully return to improving Cinepak encoder eventually).

Revisiting MSVideo1 encoder

Wednesday, February 8th, 2023

Recently somebody asked me a question about my MS Video 1 encoder (not the one in NihAV though) and I’ve decided to look if my current encoder can be improved, so I took Ghidra and went to read the binary specification.

Essentially it did what I expected: it understands quality only, for which it calculates thresholds for skip and fill blocks to be used immediately, clustering is done in the usual K-means way and the only curious trick is that it used luminance for that.

So I decided to use that idea for improving my own encoder. I ditched the generic median cut in favour of specially crafted clustering in two groups (I select the largest cluster axis—be it luma, red, green or blue—split 4 or 16 pixels into two groups by being above average or not and calculate the average of those two groups). This made encoding about two times faster. I’ve also fixed a bug with 8-colour blocks so now it encodes data properly (previously it would result in a distorted block). And of course I’ve finally made quality affect encoding process (also by generating thresholds, but with a different formula—unlike the original my encoder uses no floating-point maths anywhere).

Also I’ve added palette mode support. The idea is simple: internally I operate on pixel quads (red, green, blue, luma) so for palette mode I just need to replace an actual pixel value with the index of the most similar palette entry. For that task I reused one of the approaches from my palettiser (it should be faster than iterating over the whole palette every time). Of course the proper way would be to map colour first to have the proper distortion calculated (because the first suitable colour may be far from perfect) but I decided not to pursue this task further, even if it results in some badly-coded features sometimes. It’s still not a serious encoder after all.

Now this member of the early 90s video codecs ruling triumvirate should be good enough. Cinepak encoder is still rather primitive so I’ll have to re-check it. Indeo 3 encoder seems to produce buggy output on complex high-motion scenes (I suspect it’s related to the number of motion vectors exceeding the limit) but it will have to wait until I rewrite the decoder. And then hopefully more interesting experiments will happen.