Repository navigation
[4.3.1] - 2026-11-18
Encoder (general)
- Faster encoding across use cases, compared with v4.3.0:
- ~14-15% less CPU for random access VOD at preset 5
- ~8-11% fewer CPU cycles for RTC (presets 7-12) and mobile-build random access (presets 9-11)
- ~5-10% fewer CPU cycles for still images (presets 5-9)
- Bit-identical output to v4.3.0 for still images and mobile random access preset 9; elsewhere within noise for CRF
- Speed optimizations in mode decision (light and regular PD1 reuse, class-bucketed MDS0 candidates, cached txb contexts), RDOQ and coefficient rate (precomputed 2D contexts, zero-run skip), TPL (N2/N4-only quantization, fused 16x16 transform), motion estimation, MV and DV cost tables, CDEF, loop filter, padding and entropy coding
- Random access: full-width ME/TF and TPL segments and a right-sized ME worker pool at low thread counts
- Align picture buffer rows and plane origins to 64 bytes, and defer reference picture border initialization to first use
- Pool small allocations on Android, and free the bitstream of output packets never released by the application
- RTC CBR: faster rate control recovery after key frames (larger qindex drops on the following inter frames, a resolution-scaled initial correction factor, and faster correction updates for 30 frames), and the top 8 qindex steps are kept for buffer overflow
- VBR: intermediate temporal layers blend their qindex toward a CRF-style cascade anchored on the recent leaf qindex
- Added
--roi-map-block-size(roi_map_block_size) for 16x16 ROI QP offset maps; ROI maps with more than eight distinct offsets are now grouped into eight segments instead of being rejected
x86
- Added AVX-512 kernels for the 64x64 forward transform, open-loop and full-pel motion search, half-pel search, zone-2 intra prediction, coefficient rate, and VNNI variances
- Optimized AVX2 variance, and dispatch AVX-512 SAD to AVX2 on odd heights
Arm
- Added Neon int16 64x64 forward DCT, sparse inverse-transform coefficient packing and a DC-only fast path, shared rectangular forward-transform kernels, and a single quantize helper
- Faster Neon and Neon dot-product full-pel SAD search, CDEF masks and nz-map contexts
- Fixed the Neon scaled compound first predictor
Bug fixes and build
- Fixed
svt_av1_enc_stream_header()returning uninitialized fields - Public encoder APIs now reject null arguments with
EB_ErrorBadParameterinstead of crashing, and releasing a null output packet is a no-op - Fixed an out-of-bounds copy of luma DC quantization data in the transform depth search (no change in output)
- Fixed cppcheck findings, and the Xcode
-Wshorten-64-to-32and GCC AArch64-Wmaybe-uninitializedwarnings - Refreshed the documentation: parameter defaults and ranges, per-use-case preset limits and examples, build options
- Added ELF function annotations to assembly symbols, and fixed Windows, mingw-w64 and MSVC warnings
- Added
SVT_AV1_TEST_VECTORS_ONLYCMake option - Moved CI to GitHub Actions (Linux, macOS, Windows, BSD, Android, musl, sanitizers, Valgrind); stylecheck now compares against the GitHub repository