Xiph Rav1e: Rust AV1 encoder, 0.3.0

[foxyshadis]

The fastest and safest AV1 encoder.

rav1e is an AV1 video encoder. It is designed to eventually cover all use cases, though in its current form it is most suitable for cases where libaom (the reference encoder) is too slow.

https://github.com/xiph/rav1e/

Features

• Intra and inter frames
• 64x64 superblocks
• 4x4 to 64x64 RDO-selected square and 2:1/1:2 rectangular blocks
• DC, H, V, Paeth, smooth, and a subset of directional prediction modes
• DCT, (FLIP-)ADST and identity transforms (up to 64x64, 16x16 and 32x32 respectively)
• 8-, 10- and 12-bit depth color
• 4:2:0 (full support), 4:2:2 and 4:4:4 (limited) chroma sampling
• Variable speed settings
• Near real-time encoding at high speed levels

Current release: 0.3.0 (2020-02-06)

[foxyshadis]

Rav1e Release v0.3.0: FOSDEM Edition
https://github.com/xiph/rav1e/releases/tag/v0.3.0

Improvements

• Faster and better Speed 10
• Smaller binaries (Around 3MB stripped on x86_64/Linux)
• Faster build times (About 14% faster build time)
• Multi-threaded deblocking filter
• Additional x86_64 SIMD code
• More auto-vectorizable codepaths and bounds check elisions
• ⅙ less memory allocations
• Improvements on the intra-modes pruning logic in the RDO
• More float to fixed-point conversions
• Make an early-exit condition in RDO faster
• Simplify logic in Counter/Recorder store method
• Support wasm32-wasias build target.

New tools

• Intra edge filter
• Switch frame support (use -S, --switch-frame-interval <SWITCH_FRAME_INTERVAL>to enable it from the CLI)
• Fine directional intra prediction
• Still Picture support (AVIF)

Changes

• Upstream nasm-rs 0.1.7 can be used to build rav1e.
• The C header produced is now C++-compatible

[Kurtnoise]

Rav1e Release v0.4.0

https://github.com/xiph/rav1e/releases/tag/v0.4.0

Improvements

• Enable open partitions on frame boundaries (2% improvement to coding efficiency)
• Use av-metrics in CLI to compute PNSR, PSNR-HVS, SSIM, MS-SSIM, and CIEDE2000 (see --metrics)
• Enable deblocking in loop filter rate-distortion optimization (0.5% to 1.5% improvement to coding efficiency)
• Thread CDEF loop filter with tiles (1.2% reduction in encoding time with 4 tiles)
• Redesign the rate control API
• Add monochrome support
• Improve 4:2:2 support (37% reduction in encoding time, 0.8% to 5% improvement to coding efficiency)
• Add compound prediction mode variants for drl=2 and drl=3
• Enable NEAR_NEAR1MV and NEAR_NEAR2MV compound modes
• Support arbitrary-SAR anamorphic video
• Enforce a frame limit of 1 in still picture mode
• Add a quiet mode to the CLI (--quiet flag)
• Convert all motion vector predictors to full pixel precision
• Update non-broken motion estimation predictors (0.28% improvement to coding efficiency)
• Substantially rework initial motion estimation (9% reduction in encoding time)
• Optimise predictors for multipass motion estimation (0.3% to 0.4% improvement to coding efficiency)
• Optimize chroma quantizer offsets for 4:4:4 sampling
• Allow opaque extra data to be attached to frames and retrieved from encoded packets via the API
• Merge new dav1d assembly code for x86 and AArch64
• Add/improve assembly code for distortion computations
• Derive quantizers using linear models (0.7% to 1.7% improvement to coding efficiency)
• Prune intra frame prediction mode list dynamically (5.5% to 12.2% reduction in encoding time at speed level 5)
• Optimize rate-distortion optimization loop (1% reduction in encoding time)
• Reduce memory allocation count in various areas
• Optimize tile block access (1.5% reduction in encoding time)
• Allow frame sizes <16x16 in still picture mode
• Add high bit depth AVX2 assembly (9% to 31% reduction in encoding time for 10-bit video)

Bug Fixes

• Fix rebuilding with fresh assembly output
• Fix the chroma plane desyncs on narrow frames
• Abort rate controlled encoding without a bitrate target in the CLI
• Fix the -v CLI option
• Fix a crash when using 4 tiles for 1080p 4:2:2 input
• Fix intra edge filter desyncs with 4:2:2 and 4:4:4 input
• Fix a symbol redefinition error for AArch64 builds using Clang
• Fix loop restoration filter with 4:2:2 and 4:4:4 input
• Fix incorrect quantizer index clamping
• Fix cross-compiling for mingw-W64 on macOS
• Avoid a buffer underflow condition in CDEF pad_into_tmp16()
• Properly validate minimum RDO lookahead frames value
• Respect quantizer bounds with rate control enabled
• Restrict still picture mode to single-frame streams

Changes

• Bump minimum version of NASM to 2.14.02
• Update speed presets
• Enable full SGR search for levels 0-4 instead of 0-8
• Enable fine directional prediction for all speed levels
• Enable reduced transform type search for levels 6-10 instead of 5-10
• Disable transform type RDO for inter frames
• Rename "native" CPU feature level to "Rust" (use RAV1E_CPU_TARGET=rust at runtime)
• Remove in-library PSNR computation feature
• Move frame-related data structures to a separate crate (v_frame)
• Extend dump_lookahead_data feature
• Export the frame_subtype property
• Use the RAV1E_DATA_PATH environment variable to determine the output path
• Refactor CDEF to allow easier importation of dav1d CDEF assembly, as well as simplifying interaction between loop filters
• Remove leftover code ported from libaom
• Remove unused diamond motion estimation
• Reduce build time
• Disable LTO by default
• Disable code generation unit restriction
• Allow incremental builds for the release profile
• Inline various functions
• Remove large stack allocations
• Split large modules into multiple submodules
• Add an unstable channel API feature
• Prompt if the output file would be overwritten and add -y to override it.

[benwaggoner]

Looks to be a solid release that lays some good foundations for future work. Still, kind of less that I'd expect after 11 months for this new an encoder. It got hit by the Mozilla layoffs IIRC.

Are they going to merge in the SVT-AV1 work that the AOM encoder is now basing itself on?

[foxyshadis]

Originally Posted by benwaggoner

Looks to be a solid release that lays some good foundations for future work. Still, kind of less that I'd expect after 11 months for this new an encoder. It got hit by the Mozilla layoffs IIRC.

Are they going to merge in the SVT-AV1 work that the AOM encoder is now basing itself on?

There's been some cross-pollination, but as far as I've seen there's no public attempt to unify anything, so it's kind of ending up in the typical Google mode of having multiple overlapping things, none of which do all the things, and with no idea of what will be axed first.