Actually, I need 80 optimal Huffman trees, but let’s take it one step at a time.

The VP3 video codec — the basis of Theora — employs 80 different Huffman trees. There are 16 for DC coefficients and 16 each for 4 different AC coefficient groups. An individual VP3/Theora frame gets to select 4 different Huffman trees: one for Y-plane DC, one for C-plane DC, one for Y-plane AC, and one for C-plane AC. VP3 hardcodes these tables. Theora allows more flexibility and an encoder is free to either use the default VP3 trees or create its own set and encode them into the header of the container (typical an Ogg file).

Generating an optimal Huffman tree for a particular set of input is rather well established; any introduction to Huffman codes covers that much. What I’m curious about is how one would go about creating a set of, e.g., 16 optimal Huffman trees for a given input. The first solution that comes to mind is to treat this as a vector quantization (VQ) problem. I have no idea if this idea holds water, or if it even has any sane basis in mathematics, but when has that ever stopped me from running with a brainstorm?

Here’s the pitch:

• Modify FFmpeg’s VP3/Theora decoder to print after each frame decode the count of each type of token that was decoded from the stream (for each of the 5 coefficient groups, and for each of the plane types), as well as the number of bits that token was encoded with. This will allow tallying of the actual number of bits used for encoding tokens in each frame.
• Create a separate tool to process the data by applying a basic VQ codebook training algorithm. It will be necessary to treat all of the Y-plane AC tokens as single vectors and do the same with the C-plane AC tokens, even though each AC token vector needs to be comprised of 4 separate AC group vectors. Re-use some existing E/LGB code for this step.
• Generate Huffman trees from the resulting vectors and count the number of bits per token for each.
• Iterate through the frequency vectors captured from the first step and match them to the codebooks using a standard distance algorithm.
• Tally the bits from using the new vectors and see if there is any improvement versus the default vectors (Huffman tables).

I don’t know if I’ll have time to get around to trying this experiment in the near future but I wanted to throw it out there anyway. With all of the improvments that the new Theora encoder brings to the tables, it seems that the custom Huffman trees feature is one that is left un-exercised per my reading of the documentation and source code. From studying the Big Buck Bunny Theora encodes (my standard Theora test vectors these days), I see that they use the default VP3 tables. The 1080p variant occupied 866 MB. Could there be any notable space savings from generating custom Huffman tables? Or was this a pointless feature to add to the Theora spec?

No sooner did I press “Publish” on my last post pertaining to multithreading FFmpeg’s Theora decoder, than did I receive an email from the theora-dev list regarding the release of Theora 1.1.0. It took them many, many years to release the first official version and about 10 months to get the second version out, so congratulations on that matter. This release includes the much-vaunted Thusnelda encoder which is supposed to offer substantial encoding improvements vs. the original 1.0 encoder.

So, fair warning: Be prepared for a new round of “Theora Bests H.264 / HTML5 Video Poised To Conquer Internet” type of stories.

Since I have been doing a bunch of optimizations to the FFmpeg Theora decoder this past week (a.k.a. the Theora decoder that most people actually use), I thought this would be the perfect opportunity to benchmark Theora 1.1 alongside FFmpeg’s decoder. Fortunately, libtheora has an example tool called dump_video that decodes video directly to YUV4MPEG2 format, the same way I was testing FFmpeg’s decoder.

FFmpeg command line:

ffmpeg -threads n -i big_buck_bunny_1080p_stereo.ogg
  -f yuv4mpegpipe -an -y /dev/null

Libtheora command line:

dump_video big_buck_bunny_1080p_stereo.ogg > /dev/null

The results (on my Core 2 Duo Mac Mini) were thus:

6:44 - FFmpeg, 1 thread
6:09 - FFmpeg, 2 threads *
4:51 - libtheora 1.1

* multithreaded version isn’t complete yet

Mind you, libtheora’s decoder is singly-threaded and only has basic MMX SIMD optimizations. After seeing libtheora’s relative performance, I think I blacked out. Or maybe I just went to bed since it was so late; it’s sort of a blur. I awoke in a confused stupor wondering what I’m doing wrong in the FFmpeg Theora decoder. Why is it so slow? Actually, I know why– unpack_vlcs(), which continues to dominate profiling statistics. Perhaps the question I should start with is, how does libtheora unpack VLCs so quickly? That’s a good jumping-off point for a future investigation.

As briefly mentioned in my last Theora post, I think FFmpeg’s Theora decoder can exploit multiple CPUs in a few ways: 1) Perform all of the DC prediction reversals in a separate thread while the main thread is busy decoding the AC coefficients (meanwhile, I have committed an optimization where the reversal occurs immediately after DC decoding in order to exploit CPU cache); 2) create n separate threads and assign each (num_slices / n) slices to decode (where a slice is a row of the image that is 16 pixels high).

So there’s the plan. Now, how to take advantage of FFmpeg’s threading API (which supports POSIX threads, Win32 threads, BeOS threads, and even OS/2 threads)? Would it surprise you to learn that this aspect is not extensively documented? Time to reverse engineer the API.

I also did some Googling regarding multithreaded FFmpeg. I mostly found forum posts complaining that FFmpeg isn’t effectively leveraging however many n cores someone’s turbo-charged machine happens to present to the OS, as demonstrated by their CPU monitoring tool. Since I suspect this post will rise in Google’s top search hits on the topic, allow me to apologize to searchers in advance by explaining that multimedia processing, while certainly CPU-intensive, does not necessarily lend itself to multithreading/multiprocessing. There are a few bits here and there in the encode or decode processes that can be parallelized but the entire operation overall tends to be rather serial.

So this is the goal:

…to see FFmpeg break through the 99.9% barrier in the CPU monitor. As an aside, it briefly struck me as ironic that people want FFmpeg to use as much of as many available CPUs as possible but scorn the project from my day job for being quite capable of doing the same.

Moving right along, let’s see what can be done about exploiting what limited multithreading opportunities that Theora affords.

First off: it’s necessary to explicitly enable threading at configure-time (e.g., “–enable-pthreads” for POSIX threads on Unix flavors). Not sure why this is, but there it is.

Continue reading →