Intel SVT-AV1

[Kurt.noise]

Originally Posted by BlueSwordM

It's been fully fixed now.

yes, confirmed. Thank you.

[Kurt.noise]

Guys, could you benchmark 2.3.0-B windows x64 using different compilers please ?

https://www.mediafire.com/file/o7gmm..._clang.7z/file
https://www.mediafire.com/file/almit..._-_gcc.7z/file
https://www.mediafire.com/file/lzjwr..._intel.7z/file

Seems intel one is sightly faster on my machine.

[oibaf]

Better late than never: SVT-AV1 v2.2.x Deep Dive

[oibaf]

The End of SVT-AV1-PSY

[BlueSwordM]

Originally Posted by oibaf

The End of SVT-AV1-PSY

Yeah, we decided on quite the title to get people to read it.

No need to worry though, as most of the team members are still there and somewhat working hard.

I'm working on a guide on how to best take advantage of svt-av1-psy and its new featureset to best effect, as while `--psy-rd` was modeled like what x264/x265 do it, it was extended further by making it touch many more things overall.

[oibaf]

Originally Posted by BlueSwordM

Yeah, we decided on quite the title to get people to read it.

No need to worry though, as most of the team members are still there and somewhat working hard.

I'm working on a guide on how to best take advantage of svt-av1-psy and its new featureset to best effect, as while `--psy-rd` was modeled like what x264/x265 do it, it was extended further by making it touch many more things overall.

Thanks for the feedback and for all the effort on this!

Also very interesting that the patches are being submitted in SVT-AV1 (so that everyone can benefits)!
Actually, 8 patches out of 9 submitted are already merged:
https://gitlab.com/AOMediaCodec/SVT-...ch=SVT-AV1-PSY
and will be included in SVT-AV1 version following latest 2.3.0.

[ShortKatz]

Originally Posted by oibaf

The End of SVT-AV1-PSY

Ah, this explains why they started porting features from SVT-AV1-PSY to SVT-AV1, I was already wondering.

Originally Posted by oibaf

Also very interesting that the patches are being submitted in SVT-AV1 (so that everyone can benefits)!
Actually, 8 patches out of 9 submitted are already merged:
https://gitlab.com/AOMediaCodec/SVT-...ch=SVT-AV1-PSY
and will be included in SVT-AV1 version following latest 2.3.0.

The patch for noise-norm-strength was closed without merging. So it's 7 of 9. Only the sharpening PR is now missing. I am wondering if they are planning to port also other feature in the future.
Edit: Now sharpness is also merged. :-)

[oibaf]

Originally Posted by ShortKatz

Ah, this explains why they started porting features from SVT-AV1-PSY to SVT-AV1, I was already wondering.
I am wondering if they are planning to port also other feature in the future.

From the blog post:

We've authored over 20 highly effective features, aimed at improving perceptual quality and ease of use

20+ total.
8 merged (+1 discarded, + 1 compiler fix).

So, excluding the compiler fix, 20+ - (8 + 1), there are still 11+ specific features still in SVT-AV1-PSY only?

[Z2697]

What's spy-rd?

[Kurt.noise]

Originally Posted by Z2697

What's spy-rd?

it's a psychovisually-based metric focused on enhancing sharpness and detail retention, with a slight trade-off in blocking and banding artifacts, without deeply affecting RD/RDO decisions like psy-rd. The default is 0 (disabled).

It can be used independently or alongside psy-rd.

[Kurt.noise]

SVT-AV1 3.0.0 has been released.

[3.0.0] - 2025-2-20

API updates

• Refreshed API cleaning up unused fields, use stdbool type and cleanup redundant parameter in svt_av1_enc_init_handle
• Repositioned the presets and removed one preset resulting in a max preset of M10 in the current version
• Added temporal layer and averageQP fields in output picture structure, along with an option to specify a QP offset for the startup gop
• The API changes are not backwards compatible, more details about the changes can be found in issue 2217

Encoder

• Improved mid and high quality presets quality vs speed tradeoffs for fast-decode 2 mode:
• ~15-25% speedup for M3-M10 at the same quality levels - (!2376 and !2343)
• ~1% BD-rate improvement for presets M0-M2 - (!2376 and !2343)
• Repositioned the fast-decode 1 mode to produce ~10% decoder cycle reduction vs fast-decode 0 while reducing the BD-rate loss to ~1% (!2376)
• Further Arm Neon and SVE2 optimizations that improve high bitdepth encoding by an average of 10-25% for 480p-1080p resolutions beyond the architecture-agnostic algorithmic improvements since v2.3.0
• Ported several features from SVT-AV1-SPY fork to further improve the perceptual quality of tune 0 mode
• Added an avif mode to reduce resource utilization when encoding still images

Cleanup Build and bug fixes and documentation

• third_party: Removed vendored cpuinfo, will attempt to use one provided by the system. For those without cpuinfo, it will be pulled and compiled into the library, similar to before
• Improved the unit test coverage for Arm Neon and SVE2 code
• Updated documentation

Known issue

Mismatch in bitstreams between Arm and x86 platforms that is known not to be due to Arm or x86 optimizations. (#2247)

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

Some fresh compiles that I made (only windows x64 plateform) if you need to play with :
- https://www.mediafire.com/file/iuwlx...0_msvc.7z/file
- https://www.mediafire.com/file/l5vb8..._intel.7z/file
- https://www.mediafire.com/file/bqnme...langCL.7z/file

[ShortKatz]

Thanks. Will also soon be implemented in HandBrake to play with.
Edit: SVT-AV1 3.0 is now in HandBrake master.

[rwill]

Originally Posted by Z2697

What's spy-rd?

In contrast to others I actually looked that one up in the source. When spy-rd is enabled certain modes, those that are suspected to smooth the picture, have their RD cost scaled up. So it is not a metric. It narrows down the available modes for the Mode Decision to pick a Mode from. So a huge intervention to RDO decisions without any analysis.

From my experience this will result in spending more bits for slightly less objective quality.

Just avoid.

[ShortKatz]

I played a bit with the new sharpness parameter and I am not a huge fan of it. For me it makes noise unsteady and emphasises it more. But maybe I need to play with it a bit more to find its benefits.

[oibaf]

Couple of new releases:

[3.0.2] - 2025-3-21
Encoder

• More Arm simd improvements (!2401, !2402, !2403, !2405, !2409, !2410)
• Fixed mising initalization of lossless and avif (#2255, !2404)

Documentation

• Add missing --luminance-qp-bias documentation (!2407)

[3.0.1] - 2025-3-10
Encoder cleanup and bug fixes

• Further Arm improvements along with fixing arm vs x86 output mismatches (!2393, !2399, !2400, #2247)
• Fixed memory leak in compute_global_motion (!2395, #2248)
• Fixed integer overflow in subpel search to prevent an assertion (!2396, #2250)
• Clean up some undefined behavior (!2398)
• API change fixes now available in FFmpeg and GStreamer (#2249, #2252)

Known issue

• Hard to reproduce SIGTRAP being raised on macOS m1 with libavif's avifsvttest (#2251)

[oibaf]

Comparing Video Encoders: https://giannirosato.com/blog/post/comparing-encoders/

[rwill]

Originally Posted by oibaf

Comparing Video Encoders: https://giannirosato.com/blog/post/comparing-encoders/

Sorry, I only trust encoder comparisons I have manipulated myself.

[oibaf]

svt-av1-psy v3.0.2: Supernova: https://github.com/psy-ex/svt-av1-ps...ses/tag/v3.0.2

This will be the last BIG release of svt-av1-psy for a long while.

I'll be starting to port over all of the relevant features to mainline svt-av1.

However, even after all relevant features will have been ported, I won't be working on "mainline" svt-av1-psy anymore; all future feature additions, bug fixes and optimizations will be done on an svt-av1-psy "fork".

Some members wish to distance themselves from the project for reasons I won't discuss here; they're not harmful to the svt-av1-psy project, those members just want to leave their work on svt-av1-psy as history.

Getting back on topic, we've decided to release our current coup-de-grâce efforts, svt-av1-psy 3.0.2: Supernova. We're going out with a bang!

[ShortKatz]

There is also a new readme, announcing the real end of the project.

I posted a blog post about this in November of 2024, but on April 20th 2025 I can officially announce the actual end of SVT-AV1-PSY.

I stepped away from SVT-AV1-PSY in December of 2024. Our other amazing team member Julio also left shortly afterwards, and a third awesome team member Clybius began to roll back his efforts as well.

https://github.com/psy-ex/svt-av1-ps...ster/README.md

[BlueSwordM]

That is indeed correct, although each one left for different reasons.
I'll start porting svt-av1-psy features to mainline starting from next week.