dav1d 0.8 Released With More Optimizations - More AMD Performance

Dav1d 0.8 was released this weekend (and subsequently 0.8.1 too) as the latest major release for this CPU-based AV1 decoder hosted by the VideoLAN project. Dav1d continues to be about offering the best AV1 decode speed and with the v0.8 series are even faster results -- so here are some of our initial data points as well from some weekend benchmarking.

Dav1d 0.8 offers up more optimizations. One of the main optimizations this cycle is the introduction of a picture buffer pool. Under Windows at least the usage of this picture buffer pool can improve performance by up to 10%. There are also more AVX2 and SSE/SSSE3 optimizations too.

Dav1d 0.8 also brings several ARM32 and ARM64 optimizations. But beyond the ARM optimizations is also initial support for Apple Silicon with their ARM-based Apple M1 SoC.

Dav1d 0.8.1 was released as a minor bug fix release to fix a regression stemming from the picture buffer pool. But also a few more ARM32/ARM64 optimizations were also added.

Dav1d 0.8.x can be downloaded via the VideoLAN.org Gitlab.

As a result the Phoronix Test Suite / OpenBenchmarking.org dav1d test profile has been updated for the new release. This weekend I have been running some comparison benchmarks of dav1d 0.8.1. Below is a look at a few Intel and AMD systems comparing dav1d 0.7 to 0.8.1. More benchmarks including on ARM with the Ampere Altra coming up in the days ahead.




The AMD (Zen 2 / Zen 3) systems tested had the most benefit compared to the Intel boxes. More data on that quick comparison via this OpenBenchmarking.org result file. More benchmarks coming out soon as well as via the dav1d test profile page.