A codec does not really exist until everyone can decode it.
Today, we announce dav2d, a fast decoder for the new AV2 codec, developed by members of the VideoLAN community.
A few weeks ago, we opened the repository and started development in public. Since then, AV2 itself has reached its first official specification release, making this a good moment to explain what dav2d is, why we started it, and where the project stands today.
dav2d is the continuation of the work we started with dav1d, our AV1 decoder.
The goal is similar: provide a small, fast, portable and correct decoder, suitable for real applications, media players, browsers, test tools and operating systems.
AV2 is the successor to AV1 and the latest royalty-free video codec from the Alliance for Open Media.
The specification is now publicly available at:
AV1 was finalized in 2018 and became one of the most successful video codecs ever deployed. Today it is available in browsers, mobile devices, operating systems, televisions, streaming services and video applications around the world.
AV2 builds on that success. The codec introduces new coding tools across prediction, transforms, entropy coding, filtering and chroma processing, while continuing the goal of improving compression efficiency.
The reported gains vary depending on the test conditions, but improvements around 25% compared to AV1 are commonly seen, with some evaluations reporting even larger gains.
AV2 decoding is roughly five times more complex than AV1 decoding. In practice, that means software running on today’s hardware will struggle to decode AV2 in real time without careful, architecture-specific optimization.
This is why we started dav2d early rather than waiting for the specification to stabilize.
The origins of dav2d go back to the beginning of dav1d.
When AV1 was being finalized, we pushed for a fast software decoder, because we did not believe hardware decoding would become available quickly enough, or on enough devices.
Not everyone agreed with that assessment. Some members of the AOM community felt that hardware implementations and the reference decoder would be sufficient.
We thought otherwise. Browsers, media players, operating systems and mobile devices would need a production-quality decoder long before dedicated hardware became commonplace.
In the end, AOM itself funded part of the initial development work and some members of the Alliance eventually joined that effort.
The result was dav1d.
In hindsight, the need for a fast software decoder proved larger than many people expected.
Today, dav1d is the most widely deployed AV1 software decoder.
It is used in VLC, FFmpeg, mpv, Firefox, Chrome, Safari, Android, Windows, Linux and many other applications and platforms.
The project has also become the reference AV1 decoder implementation for many developers working on AV1 deployment, testing and optimization.
You can read the full history of dav1d on this blog: Introducing dav1d, the road to the first release, First release, dav1d 1.2 and 1.5 “Sonic”.
With AV2, we are trying to start that work earlier.
A codec specification is important, but it is not enough. Developers need a decoder that can be built, tested, benchmarked, integrated and compared against other implementations.
This is what dav2d is meant to provide.
The current dav2d tree already contains a feature-complete AVM v15 decoder supporting both 8-bit and 10-bit decoding.
Most major parts of the codec are already implemented and are now being optimized, including:
This is still early work, and the AV2 ecosystem itself is still young, but the decoder is already functional and far beyond an empty announcement repository.
A growing part of the work is now focused on correctness, conformance, optimization and platform support.
One reason the project has progressed so quickly is that dav2d does not start from scratch. AV2 shares many concepts with AV1, and dav1d already solved a number of architectural questions around threading, SIMD organization, testing, portability and API design.
While AV2 requires substantial new decoder code, a lot of the experience accumulated over years of dav1d development transfers directly to dav2d.
The performance work has already started.
On x86, dav2d already contains AVX2 code for several inverse transform sizes, as well as work around CCTX, deblock, intra prediction and CfL-related paths.
On ARM, there is already AArch64 NEON work for entropy decoding, SAD, intra prediction, palette prediction, DC predictors, smooth predictors and motion-related functions. Some arm32 work has also started.
There is also early RISC-V work, mostly around re-enabling and adapting existing intra prediction and motion compensation assembly.
This is the same kind of progression we had with dav1d: first a clean C implementation, then validation infrastructure, then architecture-specific optimized code for the most important hot paths.
One important difference compared to the early days of dav1d is tooling.
During the development of dav1d, we created checkasm, a framework used to validate and benchmark optimized implementations against their C equivalents.
dav2d benefits from that infrastructure from day one.
Combined with the architectural experience gained from dav1d, this has allowed the project to progress considerably faster than dav1d did at a comparable stage.
The current tree already contains checkasm coverage for several areas, including inverse transforms, motion compensation, film grain, CfL and reference motion-vector code.
This should make future optimization work both faster and safer.
Like dav1d before it, dav2d is developed as an open source project.
The decoder is released under the same BSD-style license as dav1d, making it easy to integrate into open source and proprietary applications alike.
As with most VideoLAN projects, development happens in public from day one:
We believe open implementations are essential for the healthy deployment of new media technologies. They provide interoperability, independent validation of specifications, easier experimentation, and a common foundation for the ecosystem.
There is still a lot of work ahead.
We need to continue tracking the AV2 specification, improve conformance, extend test coverage, optimize further x86 and ARM, work on RISC-V, improve high bit-depth performance, improve threading, reduce memory usage and prepare future releases.
But the foundations are already there: the tooling, the architecture and the experience gained from dav1d, with additional improvements.
dav1d helped make AV1 practical long before hardware support became ubiquitous.
We intend to do the same for AV2.
Let dav2d be. From VideoLAN, with love.