xiph/rav1e v0.4.0-alpha

on GitHub

This is a new big release of rav1e after 7 months making the encoder sensibly faster and better.

From 0.3 round there have been new 435 commits with around 50,000 additions and 17,000 deletions from 29 contributors.

Improvements

• Enable Open-Partition on frame boundary, gives ~2% rd gains.
• Use av-metrics in CLI to compute PNSR, PSNR-HVS, SSIM, MS-SSIM,
  CIEDE2000 (see --metrics)
• Unwaffle Rebase for Loop-filter: Now deblocking is enabled to loopfilter RDO
  giving 0.5 to 1.5% gains
• Thread CDEF with tiles giving ~1.2% performance using 2x2 Tiles
• New Rate Control API that is less error-prone to use.
• Full Monochrome Support
• Enabling CDEF, Restoration Filter for 4:2:2, decreasing encoding time by ~37%
  and making overall improvements substantial between 0.8 to 5%
• Added compound prediction mode variants for drl=2 and drl=3
• Enable NEAR_NEAR1MV, NEAR_NEAR2MV Compound mode
• Support arbitrary SAR anamorphic video
• Enforce a frame limit of 1 in STILL_PICTURE_MODE
• Quiet Mode in CLI with -q or --quiet
• Ensure all mv predictors are converted to fullpel
• Update non-broken Motion Estimation Predictors giving ~0.28% gains
• Substantially rework initial motion estimation: 9% improved performance
• Optimise Preditors for multipass motion estimation giving 0.3-0.4% gains
• Optimize Chroma quantizer offsets for subset3 4:4:4 giving 31% for Luma Metrics
  and 14% BD-Rate Improvement for CIEDE2000 for 4:4:4 clips
• Opaque data can be pinned to frames and retrieved from the matching packet.
• Merge of dav1d 0.6.0 dav1d 0.7.0, 0.7.1 Assembly for both x86 and AArch64
• Naive x86_64 intrinsics for get_satd HBD
• Added NEON assembly for dist::get_sad on aarch64 giving ~66% improved encoding time
• Integration of around 200+ 16BPC AArch64 Functions from dav1d resulting in an
  overall speedup of around ~9.5%
• Added x86 SIMD for weighted SSE computation giving 5-7% speedup on PSNR
• Derive quantizers using linear models giving ~0.7 to 1.7% gains in metrics for
  4:2:0
• Pruned Intra Mode list by SATD reducing encoding time between 5.5% to 12.2%
  at default speed level
• Optimization of rdo_loop_decision reducing total allocation count by 25% and
  1% for encoding time
• Removal of Initial Allocation for lookahead_intra_costs
• Avoid temporary allocation for inter pruning resulting in a reduced allocation
  significantly
• Reduce manual indexing in for_each in TileBlocks giving 1.5% speedup

Bug Fixes

• Fixed the rebuild with fresh assembly output
• Fixed the Chroma Desync for narrow-frames
• Abort pass encoding without a bitrate target in CLI
• Fixed the -v cli option
• Fixed a crash when using 4 tiles for 1080p 4:2:2 input
• Fixed the 4:2:0 assumption in IEF block context selection
• Fixed the symbol redefinition error for AArch64 builds using Clang
• Fixed for LRF choosing different LRU sizes in Y and UV when not 4:2:0
• Fixed the broken borrow checker for tile_blocks
• Fixed the quantizer index clamping
• Fixed the Cross-compiling from macOS to mingw-W64
• Avoids a buffer underflow condition in CDEF pad_into_tmp16()
• Properly validate minimum rdo_lookahead_frames value

Changes

• Bumped minimum version of NASM to 2.14.0
• Updated Speed Preset Settings
  • Full SGR Search is enabled for Speed Levels till 4 instead of 8
  • Enabled Fine Directional Intra Preset for all speed levels
  • Removed Diamond Motion Estimation
  • Reduced TX_Set preset is now enabled from Speed 6 instead of Speed 5
  • Disabled TX-Type RDO for inter frames.
• Rename of Native CPU Feature level to Rust: Use RAV1E_CPU_TARGET=rust from rav1e
  0.4.0-alpha instead of RAV1E_CPU_TARGET=NATIVE
• Removed in-library psnr computation facility
• Moved Frame related data structures to a separate crate (v_frame)
• Extended dump_lookahead_data
  • Now the frame_subtype is exported
  • Use the RAV1E_DATA_PATH env to place the output file.
• Major Refactoring in CDEF is both towards allowing easier import of dav1d CDEF
  assembly, as well as simplifying bitdepth and [re-]buffering requirements in LR.
• Remove of leftover libaom code
• Remove unused diamond motion estimation
• Reduced Build Time:
  • do not enable LTO by default,
  • use as many codegen unit
  • allow incremental builds for the release profile
  • in-lined various functions
  • removed large stack allocation, improved HBD SATD for x86 targets
  • split large modules in multiple submodules

Unstable features

• Channel-based API
• A mean to use a pre-allocated threadpool, and share it across multiple encoders.