The Vulkan compute shader implementations are cool...particularly for FFv1 and ProRes RAW. Given that these bypass fixed-function hardware decoders entirely, I'm curious about the memory bandwidth implications. FFv1's context-adaptive arithmetic coding seems inherently sequential, yet they're achieving "very significant speedups."
Are they using wavefront/subgroup operations to parallelize the range decoder across multiple symbols simultaneously? Or exploiting the slice-level parallelism with each workgroup handling independent slices? The arithmetic coding dependency chain has traditionally been the bottleneck for GPU acceleration of these codecs.
I'd love to hear from anyone who's profiled the compute shader implementation - particularly interested in the occupancy vs. bandwidth tradeoff they've chosen for the entropy decoding stage.
If there's anything that needs audio/video automation, I've always turned to FFmpeg, it's such a crucial and indispensible tool and so many online video tools use it and are generally a UI wrapper around this wonderful tool. TIL - there's FFmpeg.Wasm also [0].
In Jan 2024, I had used it to extract frames of 1993 anime movie in 15 minutes video segments, upscaled it using Real-ESRGAN-ncnn-vulkan [1] then recombining the output frames for final 4K upscaled anime [2]. FWIW, if I had built a UI on this workflow it could've become a tool similar to Topaz AI which is quite popular these days.
Even when I don't use directly ffmpeg, I often use tools that embed ffmpeg. For instance, I've recently upscaled an old anime, ripped from a low quality DVD. I used k4yt3x/video2x, which was good enough for what I wanted, and was easy to install. It embedded libffmpeg, so I could use the same arguments for encoding:
To find the best arguments for upscaling (last line from above), I first used ffmpeg to extract a short scene that I encoded with various parameter sets. Then I used ffmpeg to capture still images so that I could find the best set.
Happy to hear that they've introduced video encoders and decoders based on compute shaders. The only video codecs widely supported in hardware are H.264, H.265 and AV1, so cross-platform acceleration for other codecs will be very nice to have, even if it's less efficient than fixed-function hardware. The new ProRes encoder already looks useful for a project I'm working on.
> Only codecs specifically designed for parallelised decoding can be implemented in such a way, with more mainstream codecs not being planned for support.
It makes sense that most video codecs aren't amenable to compute shader decoding. You need tens of thousands of threads to keep a GPU busy, and you'll struggle to get that much parallelism when you have data dependencies between frames and between tiles in the same frame.
I haven't even had a cursory look at decoders state of the art for 10+ years. But my intuition would say that decoding for display could profit a lot from GPU acceleration for later parts of the process when there is already pixel data of some sort involved. Then I imagine thet the initial decompression steps could stay on the CPU and the decompressed, but still (partially) encoded data is streamed to the GPU for the final transformation steps and application to whatever I-frames and other base images there are. Steps like applying motion vectors, iDCT... look embarrassingly parallel at a pixel level to me.
When the resulting frame is already in a GPU texture then, displaying it has fairly low overhead.
These release notes are very interesting! I spent a couple of weeks recently writing a ProRes decoder using WebGPU compute shaders, and it runs plenty fast enough (although I suspect Apple has some special hardware they make use of for their implementation). I can imagine this path also working well for the new Android APV codec, if it ever becomes popular.
The ProRes bitstream spec was given to SMPTE [1], but I never managed to find any information on ProRes RAW, so it's exciting to see software and compute implementations here. Has this been reverse-engineered by the FFMPEG wizards? At first glance of the code, it does look fairly similar to the regular ProRes.
Reupvoted you from gray because I don't think that's fair, but I also don't know how much there is to add. As far as why I'm contributing, I haven't been socially involved in the ffmpeg dev community in a decade, but, it is a very reasonable floor to assume it's 80% not full time paid contributors.
> Happy to hear that they've introduced video encoders and decoders based on compute shaders.
This is great news. I remember being laughed at when I initially asked whether the Vulkan enc/dec were generic because at the time it was all just standardising interfaces for the in-silicon acceleration.
Having these sorts of improvements available for legacy hardware is brilliant, and hopefully a first route that we can use to introduce new codecs and improve everyone's QOL.
Impressed anytime I have to use it (even if I have to study its man page again or use an LLM to construct the right incantation or use a GUI that just builds the incantation based on visual options). Becoming an indispensable transcoding multitool.
I think building some processing off of Vulkan 1.3 was the right move. (Aside, I also just noticed yesterday that Asahi Linux on Mac supports that standard as well.)
Yeah, you can give an LLM queries like “make this smaller with libx265 and add the hvc1 tag” or “concatenate these two videos” and it usually crushes it. They have a similar level of mastery over imagemagick, too!
Yeah, LLMs have honestly made ffmpeg usable for me, for the first time. The difficulty in constructing commands is not really ffmpeg's fault—it's just an artifact of the power of the tool and the difficulties in shoehorning that power into flags for a single CLI tool. It's just not the ideal human interface to access ffmpeg's functionality. But keeping it CLI makes it much more useful as part of a larger and often automated workflow.
fwiw, `tar xzf foobar.tgz` = "_x_tract _z_e _f_iles!" has been burned into my brain. It's "extract the files" spoken in a Dr. Strangelove German accent
Better still, I recently discovered `dtrx` (https://github.com/dtrx-py/dtrx) and it's great if you have the ability to install it on the host. It calls the right commands and also always extracts into a subdir, so no more tar-bombs.
If you want to create a tar, I'm sorry but you're on your own.
I used tar/unzip for decades I think, before moving to 7z which handles all formats I throw at it, and have the same switch for when you want to decompress into a specific directory, instead of having to remember which one of tar and unzip uses -d, and which one uses -C.
"also always extracts into a subdir" sounds like a nice feature though, thanks for sharing another alternative!
For anyone curious, unless you are running a 'tar' binary from the stone ages, just skip the gunzip and cat invocations. Replace .gz with .xz or other well known file ending for different compression.
Examples:
tar -cf archive.tar.gz foo bar # Create archive.tar.gz from files foo and bar.
tar -tvf archive.tar.gz # List all files in archive.tar.gz verbosely.
tar -xf archive.tar.gz # Extract all files from archive.tar.gz
-l, --check-links
(c and r modes only) Issue a warning message unless all links to each file are archived.
And you don't need to uncompress separately. tar will detect the correct compression algorithm and decompress on its own. No need for that gunzip intermediate step.
What value does tar add over plain old zip? That's what annoys me about .tar files full of .gzs or .zips (or vice versa) -- why do people nest container formats for no reason at all?
I don't use tape, so I don't need a tape archive format.
A tar of gzip or zip files doesn't make sense. But gzipping or zipping a tar does.
Gzip only compresses a single file, so .tar.gz lets you bundle multiple files.
You can do the same thing with zip, of course, but...
Zip compresses individual files separately in the container, ignoring redundancies between files. But .tar.gz (and .tar.zip, though I've rarely seen that combination) bundles the files together and then compresses them, so can get better compression than .zip alone.
The zip directory itself is uncompressed, and if you have lots of small files with similar names, zipping the zip makes a huge difference. IIRC in the HVSC (C64 SID music archive), the outer zip used to save another 30%.
The problem is it's very non-obvious and thus is unnecessarily hard to learn. Yes, once you learn the incantations they will serve you forever. But sit a newbie down in front of a shell and ask them to extract a file, and they struggle because the interface is unnecessarily hard to learn.
LLMs and complex command line tools like FFmpeg and ImageMagick are a perfect combination and work like magic…
It’s really the dream UI/UX from sience fiction movies: “take all images from this folder and crop 100px away except on top, saturate a bit and save them as uncompressed tiffs in this new folder, also assemble them in a video loop, encode for web”.
Had to do exactly that with a bunch of screenshots I took but happened to include a bunch of unnecessary parts of the screen.
A prompt to ChatGPT and a command later and all were nicely cropped in a second.
The dread of doing it by hand and having it magically there a minute later is absolutely mind blowing. Even just 5 years ago, I would have just done it manually as it would have definitely taken more to write the code for this task.
it can work but it's far from science fiction. LLMs tend to produce extremely subpar if not buggy ffmpeg code. They'll routinely do things like put the file parameter before the start time which needlessly decodes the entire video, produce wrong bitrates, re-encode audio needlessly, and so on.
If you don't care enough about potential side effects to read the manual it's fine, but a dream UX it is not because I'd argue that includes correctness.
LLMs are a great interface for ffmpeg. There are tons of tools out there that can help you run it with natural language. Here's my personal script: https://github.com/jjcm/llmpeg
Has anyone made a good GUI frontend for accessing the various features of FFMPEG? Sometimes you just want to remux a video without doing any transcoding, or join several video and audio streams together (same codecs).
Seconded, HandBrake[0] is great for routine tasks / workflows. The UI could be simplified just a tad for super duper simple stuff (ex. ripping a multi-episode tv show disc but don't care about disc extras? you kind of have to hunt and poke based on stream length to decide which parts are the actual episodes. The app itself could probably reliably guess and present you with a 1-click 'queue these up' flow for instance) but otherwise really a wonderful tool!
For Mac users, ffWorks [1] is an amazing frontend for FFmpeg that surfaces most of the features but with a decent GUI. It’s batchable and you can setup presets too. It’s one of my favorite apps and the developer is very responsive.
Handbrake and Losslssscut are great too. But in addition to donating to FFmpeg, I pay for ffWorks because it really does offer a lot of value to me. I don’t think there is anything close to its polish on other platforms, unfortunately.
I have found the best front-end to be ChatGPT. It is very good at figuring out the commands needed to accomplish something in FFmpeg, from my natural description of what I want to do.
I haven't used a GUI I like, but LLMs like ChatGPT have been so good for solving this for me. I tell it exactly what I need it to do and it produces the ffmpeg command to do it.
FFMpeg is probably not as up high since video processing only needs to be done on the servers that receive media. I doubt most phones are running FFMpeg on video.
LLMs have really made ffmpeg implementations easy-- the command line options are so expansive and obscure it's so nice to just tell it what you want and have it spit out a crazy ffmpeg command.
I remember saving my incantation to download and convert a youtube playlist (in the form of a txt file with a list of URLs) and this being the only way to back up Chrome music bookmark folders.
Then it stopped working until I updated youtube-dl and then that stopped working once I lost the incantation :<
Yeah, basically anytime a video or audio is being recorded, played, or streamed its from ffmpeg. It runs on a couple planets [0], and on most devices (maybe?)
FFMpeg is definitely fairly ubiquitous, but you are overstating its universality quite a bit. There are alternatives that utilize Windows/macOS's native media frameworks, proprietary software that utilizes bespoke frameworks, and libraries that function independently of ffmpeg that offer similar functionality.
That being said, if you put down a pie chart of media frameworks (especially for transcoding or muxing), ffmpeg would have a significant share of that pie.
Not necessarily. A lot of video software either leverages the Windows/MacOS system codecs (ex. Media Player Classic, Quicktime) or proprietary vendor codecs (Adobe/Blackmagic).
Linux doesn't really have a system codec API though so any Linux video software you see (ex. VLC, Handbrake) is almost certainly using ffmpeg under the hood (or its foundation, libavcodec).
This seemed to be interesting to users of this site. tl;dr they added support for whisper, an OpenAI model for speech-to-text, which should allow autogeneration of captions via ffmpeg
Heads up: Whisper support depends on how your FFmpeg was built. Some packages will not include it yet. Check with `ffmpeg -buildconf` or `ffmpeg -filters | grep whisper`. If you compile yourself, remember to pass `--enable-whisper` and give the filter a real model path.
This is because blurays ship their subtitles as a bunch of text images. So pirates have 3 options:
1. Just copy them over from the Bluray. This lacks support in most client players, so you'll either need to download a player that does, or use something like Plex/Jellyfin, which will run FFMpeg to transcode and burn the picture subtitles in before sending it to the client.
2. Run OCR on the Bluray subtitles. Not perfect.
3. Steal subtitles from a streaming service release (or multiple) if it exists.
Nice! Anyone have any idea how and when this will affect downstream projects like yt-dlp, jellyfin, etc? Especially with regard to support for HW-acceleration?
I'm not using it yet, I'm using SRT for LAN streaming, and it was hard to reduce latency. I manged to bring it down to just a bit below 1 second, but supposedly WHIP can help to make it very low which would be neat.
What changed? I see the infrastructure has been upgraded, this seems like a big release, etc. I guess there was a recent influx of contributors? A corporate donation? Something else?
Not an insider, but I noticed that there is now a filter for using Whisper (C++) for audio transcription [1]. It looks like you provide the path to a model file [2].
ffmpeg is a treasure to the open source and audio technology communities. The tool cuts right through all kinds of proprietary and arcane roadblocks presented by various codecs and formats and it's clear a tremendous amount of work goes into keeping it all working. The CLI is of course quite opaque and the documentation for various features is often terse, but it's still the only tool on any platform anywhere that will always get you what you need for video and audio processing without ever running up against some kind of commercial paywall.
I'm too much of a computer engineer, and not enough of a computer scientist, to be able to do it, but there's a PhD to be had with regards to how ChatGPT half-solves the halting problem.
There's actually a pattern of these accounts that look like they were once real people, who stopped commenting, and multiple years later were necromanced as a spam-zombie. You'll notice it clearly if you start looking at the histories of spammers after you [flag] them.
I've complained several times to the mods about it, so I'm sure they're aware too.
Is it too big a leap for me to assume someone is going around using password spraying or whatever to compromise neglected accounts for use as spam bots?
Are they using wavefront/subgroup operations to parallelize the range decoder across multiple symbols simultaneously? Or exploiting the slice-level parallelism with each workgroup handling independent slices? The arithmetic coding dependency chain has traditionally been the bottleneck for GPU acceleration of these codecs.
I'd love to hear from anyone who's profiled the compute shader implementation - particularly interested in the occupancy vs. bandwidth tradeoff they've chosen for the entropy decoding stage.
If there's anything that needs audio/video automation, I've always turned to FFmpeg, it's such a crucial and indispensible tool and so many online video tools use it and are generally a UI wrapper around this wonderful tool. TIL - there's FFmpeg.Wasm also [0].
In Jan 2024, I had used it to extract frames of 1993 anime movie in 15 minutes video segments, upscaled it using Real-ESRGAN-ncnn-vulkan [1] then recombining the output frames for final 4K upscaled anime [2]. FWIW, if I had built a UI on this workflow it could've become a tool similar to Topaz AI which is quite popular these days.
[0]: https://github.com/ffmpegwasm/ffmpeg.wasm
[1]: https://github.com/xinntao/Real-ESRGAN-ncnn-vulkan
[2]: https://files.horizon.pics/3f6a47d0-429f-4024-a5e0-e85ceb0f6...
> Only codecs specifically designed for parallelised decoding can be implemented in such a way, with more mainstream codecs not being planned for support.
It makes sense that most video codecs aren't amenable to compute shader decoding. You need tens of thousands of threads to keep a GPU busy, and you'll struggle to get that much parallelism when you have data dependencies between frames and between tiles in the same frame.
I wonder whether encoders might have more flexibility than decoders. Using compute shaders to encode something like VP9 (https://blogs.gnome.org/rbultje/2016/12/13/overview-of-the-v...) would be an interesting challenge.
When the resulting frame is already in a GPU texture then, displaying it has fairly low overhead.
My question is: how wrong am I?
The ProRes bitstream spec was given to SMPTE [1], but I never managed to find any information on ProRes RAW, so it's exciting to see software and compute implementations here. Has this been reverse-engineered by the FFMPEG wizards? At first glance of the code, it does look fairly similar to the regular ProRes.
[1] https://pub.smpte.org/doc/rdd36/20220909-pub/rdd36-2022.pdf
I'm curious wrt how a WebGPU implementation would differ from Vulkan. Here's mine if you're interested: https://github.com/averne/FFmpeg/tree/vk-proresdec
This is great news. I remember being laughed at when I initially asked whether the Vulkan enc/dec were generic because at the time it was all just standardising interfaces for the in-silicon acceleration.
Having these sorts of improvements available for legacy hardware is brilliant, and hopefully a first route that we can use to introduce new codecs and improve everyone's QOL.
I think building some processing off of Vulkan 1.3 was the right move. (Aside, I also just noticed yesterday that Asahi Linux on Mac supports that standard as well.)
FFmpeg arguments, the original prompt engineering
One would use gemini-cli (or claude-cli),
- and give a natural language prompt to gemini (or claude) on what processing needs to be done,
- with the correct paths to FFmpeg and the media file,
- and g-cli (or c-cli) would take it from there.
Is this correct?
ffmpeg right after
fwiw, `tar xzf foobar.tgz` = "_x_tract _z_e _f_iles!" has been burned into my brain. It's "extract the files" spoken in a Dr. Strangelove German accent
Better still, I recently discovered `dtrx` (https://github.com/dtrx-py/dtrx) and it's great if you have the ability to install it on the host. It calls the right commands and also always extracts into a subdir, so no more tar-bombs.
If you want to create a tar, I'm sorry but you're on your own.
"also always extracts into a subdir" sounds like a nice feature though, thanks for sharing another alternative!
The only options you ever need are tar -x, tar -c (x for extract and c for create). tar -l if you wanna list, l for list.
That's really it, -v for verbose just like every other tool if you wish.
Examples:
You never need anything else for the 99% case.Surely you mean -t if you wanna list, t for lisT.
l is for check-Links.
And you don't need to uncompress separately. tar will detect the correct compression algorithm and decompress on its own. No need for that gunzip intermediate step.Whoops, lol.
> on its own
Yes.. I'm aware, but that's more options, unnecessary too, just compose tools.
I don't use tape, so I don't need a tape archive format.
Gzip only compresses a single file, so .tar.gz lets you bundle multiple files. You can do the same thing with zip, of course, but...
Zip compresses individual files separately in the container, ignoring redundancies between files. But .tar.gz (and .tar.zip, though I've rarely seen that combination) bundles the files together and then compresses them, so can get better compression than .zip alone.
Principle of least surprise and all that.
I wasn't expecting the downvotes for an xkcd reference
It’s really the dream UI/UX from sience fiction movies: “take all images from this folder and crop 100px away except on top, saturate a bit and save them as uncompressed tiffs in this new folder, also assemble them in a video loop, encode for web”.
A prompt to ChatGPT and a command later and all were nicely cropped in a second.
The dread of doing it by hand and having it magically there a minute later is absolutely mind blowing. Even just 5 years ago, I would have just done it manually as it would have definitely taken more to write the code for this task.
If you don't care enough about potential side effects to read the manual it's fine, but a dream UX it is not because I'd argue that includes correctness.
https://youtu.be/9kaIXkImCAM?si=b_vzB4o87ArcYNfq
It's a great tool. Little long in the tooth these days, but gets the job done.
Past that, I'm on the command line haha
[0] https://handbrake.fr
Handbrake and Losslssscut are great too. But in addition to donating to FFmpeg, I pay for ffWorks because it really does offer a lot of value to me. I don’t think there is anything close to its polish on other platforms, unfortunately.
[1]: https://www.ffworks.net/index.html
https://www.shotcut.org/
https://www.mltframework.org/
Could be an interesting data source to explore that opinion.
Then it stopped working until I updated youtube-dl and then that stopped working once I lost the incantation :<
[0] - https://xkcd.com/2347/
[0] https://link.springer.com/article/10.1007/s11214-020-00765-9
That being said, if you put down a pie chart of media frameworks (especially for transcoding or muxing), ffmpeg would have a significant share of that pie.
Linux doesn't really have a system codec API though so any Linux video software you see (ex. VLC, Handbrake) is almost certainly using ffmpeg under the hood (or its foundation, libavcodec).
It also was originally authored by the same person who did lzexe, tcc, qemu, and the current leader for the large text compression benchmark.
Oh, and for most of the 2010's there was a fork due to interpersonal issues on the team.
It’s exceedingly good software though, and to be fair I think it’s gotten a fair bit of sponsorship and corporate support.
This seemed to be interesting to users of this site. tl;dr they added support for whisper, an OpenAI model for speech-to-text, which should allow autogeneration of captions via ffmpeg
yep, finally the deaf will able to read what people are saying in a porno!
This could streamline things
1. Just copy them over from the Bluray. This lacks support in most client players, so you'll either need to download a player that does, or use something like Plex/Jellyfin, which will run FFMpeg to transcode and burn the picture subtitles in before sending it to the client.
2. Run OCR on the Bluray subtitles. Not perfect.
3. Steal subtitles from a streaming service release (or multiple) if it exists.
I am curious about adoption and features that would make big difference to users :)
Secondly, just curious: any insiders here?
What changed? I see the infrastructure has been upgraded, this seems like a big release, etc. I guess there was a recent influx of contributors? A corporate donation? Something else?
[1]: https://github.com/ggml-org/whisper.cpp
[2]: https://git.ffmpeg.org/gitweb/ffmpeg.git/commit/13ce36fef98a...
https://news.ycombinator.com/item?id=44886647 ("FFmpeg 8.0 adds Whisper support (ffmpeg.org)"—9 days ago, 331 comments)
I've complained several times to the mods about it, so I'm sure they're aware too.