Intel SVT-AV1

[hajj_3]

v1.6.0 - 2023-06-18

Encoder:

Improve the tradeoffs for the random access mode across presets M1-M13:
Speeding up the higher quality presets by 30-40%
Improving the BD-rate by 1-4% for the faster presets
Improve the tradeoffs for the low delay mode for both screen content and non-screen content encoding modes
Add a toggle to remove the legacy one-frame buffer at the input of the pipeline allowing the low delay mode to operate at sub-frame processing latencies
Add a new API allowing the user to specify quantization offsets for a region of interest per frame

Build, cleanup and bug fixes:

Various cleanups and functional bug fixes
Fix the startup minigop size BD-rate loss
Add ability to run the ci-testing offline

[ShortKatz]

Version 1.8.0

Encoder

Improve the tradeoffs for the random access mode across presets:
Speedup CRF presets M6 to M0 by 17-53% while maintaining similar quality levels
Re-adjust CRF presets M7 to M13 for better quality with BD-rate gains ranging from 1-4%
Improve the quality and speed of the 1-pass VBR mode
More details on the per preset improvements can be found in MR !2143
Add API allowing to update bitrate / CRF and Key_frame placement during the encoding session for CBR lowdelay mode and CRF Random Access mode
ARM Neon SIMD optimizations for most critical kernels allowing for a 4.5-8x fps speedup vs the c implementation
Cleanup and bug fixes and documentation

Various cleanups and functional bug fixes
Update the documentation for preset options and individual features

[Marsu42]

Originally Posted by ShortKatz

Version 1.8.0

Does anyone happen to have a Windows binary to share?

Strangely, except for being built into Handbrake (nightly) and ffmpeg I cannot finde recent "SvtAv1EncApp.exe" anywere for use with av1an or custom command lines.

And yes, I know about just-learn-to-compile-it-yourself so with me asking this isn't the answer.

[Boulder]

There are some community builds available at the SVT-AV1 Discord channel.

[nakTT]

Originally Posted by ShortKatz

Version 1.8.0

Tested the ARM version of SVT-AV1 1.8.0 and it is as good as advertised. It is WAYYYY faster than the last one I tested (Version 1.4.1 released on 2022-12-12). HUGE progress made on the ARM front of the SVT-AV1 encoder within the space of just ONE year.
Many thanks to the AV1 development team.

[Blue_MiSfit]

I highly recommend you get media-autobuild-suite up and running to cut binaries for you

[ShortKatz]

Does somebody know if there are plans to add support for YUV422 to SVT-AV1 ?
The only (outdated) information I found is this: https://www.phoronix.com/news/Intel-...-Decode-FFmpeg

[nakTT]

Version 2.0.0

Changelog

[2.0.0] - 2024-03-13

Major API updates

- Changed the API signaling the End Of Stream (EOS) with the last frame vs with an empty frame
- OPT_LD_LATENCY2 making the change above is kept in the code to help devs with integration
- Removed the 3-pass VBR mode which changed the calling mechanism of multi-pass VBR
- Moved to a new versioning scheme where the project major version will be updated everytime API/ABI is changed


Encoder

- Improve the tradeoffs for the random access mode across presets:
- Speedup presets MR by ~100% and improved quality along with tradeoff improvements across the higher quality presets (!2179)
- Improved the compression efficiency of presets M9-M13 by 1-4% (!2179)
- Simplified VBR multi-pass to use 2 passes to allow integration with ffmpeg
- Continued adding ARM optimizations for functions with c_only equivalent
- Replaced the 3-pass VBR with a 2-pass VBR to ease the multi-pass integration with ffmpeg
- Memory savings of 20-35% for LP 8 mode in preset M6 and below and 1-5% in other modes / presets


Cleanup and bug fixes and documentation

- Various cleanups and functional bug fixes
- Update the documentation to reflect the rate control changes

[ShortKatz]

Did someone experiment already with the new variance boost feature, especially with the values of variance-boost-strength and variance-octile?

[Boulder]

Originally Posted by ShortKatz

Did someone experiment already with the new variance boost feature, especially with the values of variance-boost-strength and variance-octile?

The default values are very good already and prevent boosting areas unnecessarily. For darker material, changing strength to 3 and/or octile to 4 can be useful.

[ShortKatz]

OK, thanks. I used the default values for some time and started with variance-boost-strength=3 today, so not much experience yet.

[benwaggoner]

Originally Posted by Boulder

The default values are very good already and prevent boosting areas unnecessarily. For darker material, changing strength to 3 and/or octile to 4 can be useful.

I imagine different parameters would be optimal for gamma versus PQ. Algorithms based on code values often fall apart when applied to PQ if they don't account for the big perceptual differences in PQ code values. For example, gamma needs lower QPs near black, but PQ does not. However PQ doesn't have a defined max brightness like gamma does, and can vary hugely between titles and shots.

[ShortKatz]

There is now a handy documentation for the new variance boost feature:
https://gitlab.com/AOMediaCodec/SVT-...iance-Boost.md

[benwaggoner]

Originally Posted by ShortKatz

There is now a handy documentation for the new variance boost feature:
https://gitlab.com/AOMediaCodec/SVT-...iance-Boost.md

Great link with some helpful illustrations.

[BlueSwordM]

SVT-AV1-PSY 2.0.0-A Release:
https://github.com/gianni-rosato/svt...s/tag/v2.0.0-A

This time, we've decided to change the way we name these updates as they've been difficult to keep track of versioning against mainline SVT-AV1 and our own. We'll still follow the cadence of mainline SVT-AV1, but with letter releases added in between for additional clarity.

## PSY Updates

### Features
- The CRF range, previously capped at 63, has been extended to a maximum value of 70. It can also be incremented in quarter increments of 0.25
- New option: `--enable-dlf 2` for a slower, more accurate deblocking loop filter
- New option: `--qp-scale-compress-strength` *(0 to 3)* which sets a strength for the QP scale algorithm to compress values across all temporal layers. Higher values result in more consistent video quality
- New option: `--frame-luma-bias` *(0 to 100)* enables experimental frame-level luma bias to improve quality in dark scenes by adjusting frame-level QP based on average luminance across each frame

### Quality & Performance
- A temporal layer chroma qindex offset has been added to Tune 3 for more consistent quality across frames when encoding
- Minor speed bump to Preset 8
- Encoder innovation:
Dynamic delta_q_res switching implemented to help reduce signaling overhead, which should improve quality especially at CRF ≥40!
- Other general improvements to Tune 3

### Documentation
- PNG images have been replaced with smaller lossless WebP images, resulting in faster loading & repository cloning times.
- More consistent & thorough [PSY Development page](https://github.com/gianni-rosato/svt...Development.md), including build instructions

### Bug Fixes
- Help menu formatting adjusted for less frequent underlining
- `--progress 2` no longer reports the same information as `--progress 3`

### Other
- Introducing **PSY Micro-releases**! Each micro-release will be marked with a letter, bringing a bundle of new features & improvements. The release letter will reset back to the initial `A` each time our mainline version is updated. More info can be found in this project's README & the PSY Development page

Yes, this is a near copy paste of what I've posted on other public forums

[ShortKatz]

Version 2.1 with variance boost has been released.

[2.1.0] - 2024-05-17
API updates

One config parameter added within the padding size. Config param structure size remains unchanged
Presets 6 and 12 are now pointing to presets 7 and 13 respectively due to the lack of spacing between the presets
Further preset shuffling is being discussed in #2152


Encoder

Added variance boost support to improve visual quality for the tune vq mode
Improve the tradeoffs for the random access mode across presets:
Speedup of 12-40% presets M0, M3, M5 and M6 while maintaining similar quality levels
Improved the compression efficiency of presets M11-M13 by 1-2% ( !2213)
Added ARM optimizations for functions with c_only equivalent

Cleanup, build, bug fixes and documentation

Use nasm as a default assembler and yasm as a fallback
Fix performance regression for systems with multiple processor groups
Enable building SvtAv1ApiTests and SvtAv1E2ETests for ARM
Added variance boost documentation
Added a mailmap file to map duplicate git generated emails to the appropriate author

https://gitlab.com/AOMediaCodec/SVT-AV1/-/releases

[BlueSwordM]

SVT-AV1-PSY 2.1.0-A release:
https://github.com/gianni-rosato/svt...s/tag/v2.1.0-A

-> MAIN CHANGES:
- New parameter: `--max-32-tx-size`. This restricts block transform sizes to a maximum of 32x32 pixels. This can be useful in very specific scenarios for improving overall visual quality where the AV1 spec can allow for bad decisions decided by SAD/SSD/PSNR.

- New option: `--hdr10plus-json`. This parameter, added by quietvoid, allows for HDR10+ HDR support. In order to build a binary with support for HDR10+, see our PSY Development page

- New option: `--adaptive-film-grain`. This parameter which helps remedy perceptually harmful grain patterns caused by extracting grain from blocks that are too large for a given video resolution. This parameter is enabled by default.

-> TUNING:
- Disabled SSIM-cost transform decisions while keeping SSIM-cost mode decisions, generally improving Tune 3's efficiency & consistency. Future changes might contain additional tuning at this level.

-> PERFORMANCE:
- Additional NEON optimizations for ARM platforms, providing a speed increase

-> BUG FIXES:
- Temporarily disabled quantization matrices (--enable qm 0) for presets 5 and higher due to a visual consistency bug: https://github.com/gianni-rosato/svt-av1-psy/issues/56

[oibaf]

It looks like from 2.1.1 they removed the decoder:

[2.1.2] - 2024-06-27
Cleanup, bug fixes:

Fixed profile-guided-optimization build by removing the remaining decoder path


[2.1.1] - 2024-06-25
Cleanup, bug fixes, and documentation improvements:

Removed the SVT-AV1 Decoder portion of the project. The last version containing the decoder is SVT-AV1 v2.1.0.
Updated the folder structure and library build order to reflect the removal of the decoder.
Renamed all files (except for API files) to remove the "Eb" prefix and changed them to camel_case format.
Updated the gtest version to v1.12.1.
Added CI support for ARM-based macOS machines.
Improved documentation for accuracy and completeness.

[oibaf]

Originally Posted by oibaf

It looks like from 2.1.1 they removed the decoder:

Here is the rationale: https://gitlab.com/AOMediaCodec/SVT-AV1/-/issues/2182

[Spyros]

SVT-AV1 2.2.0 was released two days ago. (Release)

Encoder

Improve the tradeoffs for the random access mode across presets:
Speedup of ~15% across presets M0 - M8 while maintaining similar quality levels (!2253)
Improve the tradeoffs for the low-delay mode across presets (!2260)
Increased temporal resolution setting to 6L for 4k resolutions by default
Added ARM optimizations for functions with c_only equivalent yielding an average speedup of ~13% for 4k10bit

Cleanup Build and bug fixes and documentation

Profile-guided-optimized helper build overhaul
Major cleanup and fixing of Neon unit test suite
Address stylecheck dependence on public repositories