Community VideoLAN and FFmpeg has published the second release (0.2) of the library dav1d with the implementation of alternative free decoder format video encoding AV1. Project code is written in C language (C99) the assembler inserts (NASM/GAS) and is distributed under the BSD license. Supports x86, x64, ARMv7 and ARMv8, and Linux operating systems, Windows, macOS, Android and iOS.
Dav1d library supports all the features AV1, including advanced types of subsampling, and all specifications and management parameters color depth (8, 10 and 12 bits). The library is tested on a large collection of files format AV1. Dav1d key feature is the focus on achieving the highest possible output, decoding and providing quality work in a multithreaded mode. In the decoding process dav1d consumes 4 times less memory.
In the new version of the work on increase of productivity when processing streams with 8-bit encoding of color on older computers and mobile devices. Added assembler optimization are implemented with the use of SSSE3 and NEON instructions (previously, the optimization was limited to use of AVX2 instructions). The performance of most functions in C increased in the range of 2 to 20 times. Compared to the previous edition decoding performance in single-threaded mode on average increased 2.25 times, and in the case of multi – threaded 2.5 times.
In General dav1d now provides performance sufficient to decode 1080p video with a frame rate of 30fps on almost all processors support SSSE3.
Compared to the reference decoder aomdec on systems that support SSSE3, but no instructions AVX2, dav1d works in 2.5 – 3 times faster for threads with 8-bit color depth and color sub-sampling 4:2:0 (aomdec uses 4 threads, and dav1d – 8 streams for separate processing personnel and 4 for stitching fragments). In single thread performance at the same level.
When using SSE4 instructions.1 (supported on 95,82% of the systems according to the service Steam), dav1d while slightly behind the encoder aomdec in single-threaded mode, but is twice faster in multi-threaded mode.
When you use AVX2 instructions, compared with last year productivity increased by 1-2%. While dav1d remains faster aomdec 40% in single-threaded mode and 2.5 to 5 times in a multithreaded.
On ARM64 systems (Snapdragon 835) instructions NEON performance rose on average by 38% in single-threaded mode and 53% in multithreaded. On a 32-bit ARMv7 processors the performance increase was 62% and 46%, respectively.
Recall that the AV1 video codec developed by Open Media Alliance (AOMedia), which represents companies such as Mozilla, Google, Microsoft, Intel, ARM, NVIDIA, IBM, Cisco, Amazon, Netflix, AMD, VideoLAN, CCN and Realtek. AV1 is positioned as a public and do not require payment of royalties free format video encoding, which is significantly ahead of H. 264 and VP9’s compression level. For the entire range of tested solutions in average AV1 provides the same level of quality while reducing the bit rate by 13% compared to VP9 and 17% compared to HEVC. At high bitrates the payout increases to 22-27% for VP9 and up to 30-43% for HEVC. In tests, Facebook AV1 overtook the level of compression main profile H. 264 (x264) at 50.3%, high profile H. 264 at 46.2%, and VP9 (libvpx-vp9) at 34.0%.
Additionally, you can mark the publication of the draft specifications for the new format of encoding images AVIF, which involved technology intraframe compression of AV1. After reviewing the draft by the technical Committee of the AOM Alliance (Alliance) is scheduled to publish the final version of the specification (the specification is at the stage of candidate releases – 1.0.0-rc). AVIF supports images in HDR (High Dynamic Range) and color space Wide-gamut and standard dynamic range (SDR). Container for distribution of compressed data in AVIF is completely similar to HEIF.
According to the materials: www.opennet.ru