dav1d accelerated AV1 decoder

[hajj_3]

Originally Posted by Blue_MiSfit

To the contrary, 4k TVs are widely popular, and they all have 10 bit HDR & WCG panels.

I'll agree that the vast majority of PC / Mac systems do not have HDR / 10 bit displays

hardly anyone with an x86 desktop/laptop is connected to a tv therefore it isn't a priority for netflix.

[soresu]

Originally Posted by hajj_3

hardly anyone with an x86 desktop/laptop is connected to a tv therefore it isn't a priority for netflix.

Speak for yourself, not everyone has perfect vision and can use a sub 30 inch monitor to read with.

I'm using a 40 inch HDTV and even with that I still need to scale up the text/dpi just to read without getting a splitting headache.

I'm just hoping that when I finally upgrade to a UHDTV that all the relevant software I use has dynamically scaling UI now as many didn't when I first started noticing my vision loss 9 years ago.

[soresu]

Originally Posted by Blue_MiSfit

To the contrary, 4k TVs are widely popular, and they all have 10 bit HDR & WCG panels.

I'll agree that the vast majority of PC / Mac systems do not have HDR / 10 bit displays

Oddly not all 4K TVs are either 10 bit or HDR capable, something that took sometime to figure out as my dad kept getting a 'downscaling' message on his 2015 Panasonic 4k TV when he was playing a 4K bluray.

As it turned out his TV had been made before the HDR part of the 4K UHD Bluray standard had been set in stone, so it basically just does 2160p resolution, no 10 bpc (native or 8 bpc + FRC) or HDR to be had.

It does have a Displayport connector though as well as the standard HDMI, which is a considerable oddity for a TV model.

As 4K TV's were being made well before my dad bought his I can only guess that before the full 4k UHD standard was hammered out it was somewhat like the pre 'Full HD' phase of HDTV's that had 720p and other such sub 1080p resolutions that poor people were duped into en masse.

[Blue_MiSfit]

Sure, there's some outliers. 2015-2016 is when things really solidified. 4K UHD TVs before 2015 were extremely rare and expensive. I'd say your dad's is in that edge territory

[foxyshadis]

Originally Posted by hajj_3

hardly anyone with an x86 desktop/laptop is connected to a tv therefore it isn't a priority for netflix.

I wouldn't be surprised if they'd love to flip a switch and have HDR on all PCs, if it wasn't for gpu drivers that can't HDR their way out of a wet paper bag, and create new bugs every version since they don't use video as a regression test case at all. It's a bit of a catch-22, since gpu driver teams only put any effort into anything other than games when it becomes critical mass and they can't ignore it anymore.

[soresu]

Originally Posted by Blue_MiSfit

Sure, there's some outliers. 2015-2016 is when things really solidified. 4K UHD TVs before 2015 were extremely rare and expensive. I'd say your dad's is in that edge territory

Or just plain unreliable.

The TV broke less than 6 months shy of the 5 year warranty going kaput last year - thankfully the new motherboard inside not only works fine but boots the TV 3-4 times faster from standby than the old one

I'm pretty sure that it was also a 2014 model on very reduced price sale to clear stock, so not quite so expensive - I'm hoping to get that myself for a Samsung Q80T 55 inch to finally upgrade from my basic HDTV I'm using for a monitor.

[soresu]

dav1d is finally getting some x86 10 bpc AVX2 SIMD by porting some recent work for rav1e, no idea what sort of gains it will get yet but I'll leave the gitlab issue links in case the info gets posted there.

https://code.videolan.org/videolan/d..._requests/1110

https://code.videolan.org/videolan/d..._requests/1111

[soresu]

Mostly more ARM32 NEON assembly.

At this point 1080p should be pretty viable on more recent streaming devices like Fire TV and Chromecast.*

*that's Chromecast 4 with A55 cores mind you.

[Spyros]

Some 10-bit assembly from rav1e was merged in dav1d, seemingly resulting in 25-30%+ better performance in AVX2 systems



Source + more details

[Blue_MiSfit]

Excellent news!

[benwaggoner]

Originally Posted by Spyros

Some 10-bit assembly from rav1e was merged in dav1d, seemingly resulting in 25-30%+ better performance in AVX2 systems



Source + more details

What's the system these results are being measured on?

And is anyone keeping track of how PC power utilization is impacted by using SW AV1 versus a HW decoder? Back in the Silverlight days, I saw ~20 watts extra on a beefy laptop with H.264 SW decode versus HW decode. I imagine the gap is lower now, but it could be a pretty significant net environmental impact if 10M people are watching YouTube at an extra 10 watts each. Everything that can push that down is helpful.

[unlord]

Originally Posted by benwaggoner

What's the system these results are being measured on?

Author of the AVX2 patches (and graph) here. This experiment was conducted on a 3970x with a single core running at 3.7GHz. I can provide the sequences if you'd like to repeat the test on another system.

[benwaggoner]

Originally Posted by unlord

Author of the AVX2 patches (and graph) here. This experiment was conducted on a 3970x with a single core running at 3.7GHz. I can provide the sequences if you'd like to repeat the test on another system.

Thank you.

Why single-core for testing, just curious? There aren't any AVX2 CPUs without at least 4 cores IIRC.

[unlord]

Originally Posted by benwaggoner

Why single-core for testing, just curious? There aren't any AVX2 CPUs without at least 4 cores IIRC.

Conducting a multi-threaded decoder comparison across implementations requires controlling for more variables. It is outdated now, but here is a comprehensive multi-threaded configuration study I ran in May 2020 comparing just dav1d to libgav1:

https://docs.google.com/spreadsheets...htcqj1_4js9jSo

Note the Thread Configurations table.

[hajj_3]

dav1d v0.8.2:

- ARM32 optimizations for ipred and itx in 10/12bits, completing the 10b/12b work on ARM64 and ARM32
- Give the post-filters their own threads
- ARM64: rewrite the wiener functions
- Speed up coefficient decoding, 0.5%-3% global decoding gain
- x86: rewrite the SGR AVX2 asm
- x86: improve msac speed on SSE2+ machines
- ARM32: improve speed of ipred and warp
- ARM64: improve speed of ipred, cdef_dir, cdef_filter, warp_motion and itx16
- ARM32/64: improve speed of looprestoration
- Add seeking, pausing to the player
- Update the player for rendering of 10b/12b
- Misc speed improvements and fixes on all platforms
- Add a xxh3 muxer in the dav1d application

[benwaggoner]

Originally Posted by unlord

Conducting a multi-threaded decoder comparison across implementations requires controlling for more variables. It is outdated now, but here is a comprehensive multi-threaded configuration study I ran in May 2020 comparing just dav1d to libgav1:

https://docs.google.com/spreadsheets...htcqj1_4js9jSo

Note the Thread Configurations table.

Makes sense, and thank you!

Do you have any rough estimate for the gap in perf of a 10-bit and an 8-bit decode with equally optimized decoders?

[savage747]

Originally Posted by benwaggoner

There aren't any AVX2 CPUs without at least 4 cores IIRC.

My notebook with an i5 4300U would like to make its existence known ;-)

(2 Cores, 4 Threads, AVX2, 1.9 to 2.9 GHz)

Intel *really* loved selling dual-core mobile CPUs with similar nametags as quad-core desktop CPUs...

[LigH]

Hyper-Hyper threading FTW

[benwaggoner]

Originally Posted by savage747

My notebook with an i5 4300U would like to make its existence known ;-)

(2 Cores, 4 Threads, AVX2, 1.9 to 2.9 GHz)

Intel *really* loved selling dual-core mobile CPUs with similar nametags as quad-core desktop CPUs...

I stand corrected!

There weren't any single core AVX2 processors at least, right?

[savage747]

Originally Posted by benwaggoner

There weren't any single core AVX2 processors at least, right?

Well, not really, as far as I know.

The closest to a single-core processor with AVX2 I know of would be the AMD A6-9500 for the current AM4 platform or the AMD A6-7480 for the old FM2+ platform. Those are marketed as two cores and two threads, but they're using AMD's ill-fated Bulldozer module architecture - so there is only one "module" (and thus only one FPU/vector unit to process AVX instructions) with two integer cores sharing the FPU/vector unit and parts of the frontend and memory system.