LAV Filters - DirectShow Media Splitter and Decoders

[NikosD]

Originally Posted by VictorLS

i.e. tried Core i7-4770 not enough - stuttering present and with LAV Video Decoder impossible to know what commentator said) can decode all 50fps of only 4K but 4:2:2 h265 20180428-175510_4K ENC 3 RMAD VS LEG.ts (563 MB) https://yadi.sk/d/m0s5ztpj3XnUyw

Core i3 9100F - Win 10 1909 (18363.693)
DXVA Checker v4.2.1
LAV Filters v0.74.1-32 x64

Decoding mode: (FPS)
42/52/83 CPU usage 89% (average)

Tried MPC-HC with MPC VR and is stuttering like crazy ~ 10fps

Using EVR-CP, it plays fine at 50fps using ~90% CPU

Originally Posted by VictorLS

huhn, QBhd and other with powerful processors
Try to play with LAV Video Decoder 8K AV1 HDR 60fps 31Mbps 831MB three and half minutes (namely so long it must be played measured with independent watches):

youtube-dl -f 571 https://www.youtube.com/watch?v=d1VF4DfBPOY

- my computer can't even seek through this file (but I'm sure with only I-frame decoding it'll be possible) to see what after night with bright points in a bottom (and I haven't Internet to download with lower resolution), please. What CPU load and how many fps do you see in renderer statistic?

Tried using DXVA Checker in Decode Mode and for the first 5% of the clip I had ~08 fps on average using ~98% of my Core i3 9100F, so I just stopped it.

Originally Posted by VictorLS

One more (and beautiful Yellowstone National Park) 8K AV1 HDR 60fps 23Mbps 705MB 04min04s

youtube-dl -f 571 https://www.youtube.com/watch?v=mHUOCxVT5ro

Just like the previous 8K AV1 but ~ 11fps on average.
For 8K AV1, let's wait for nVidia Ampere or RDNA2 (?) using HW decoders.

[VictorLS]

NikosD
Thank you very much for your experiments proving my presumes (except 'Using EVR-CP, it plays fine at 50fps using ~90% CPU' - I thought i3 isn't enough, just not weakest i5 - btw what about with madVR?).

Originally Posted by NikosD

For 8K AV1, let's wait for nVidia Ampere or RDNA2 (?) using HW decoders.

I agree, at 2019 I've asked nVIDIA to add AV1 hardware acceleration in their after-TURING (I didn't know it'll be called Ampere) videoprocessor - they've promised to think about
Any news about AV1 by nVIDIA?

[clsid]

Nev, can you make a MinGW GCC 9.3 build?

[huhn]

Originally Posted by VictorLS

Intel Core i7-4770 can't play that 4K 4:2:2 h265 20180428-175510_4K ENC 3 RMAD VS LEG.ts (563 MB) https://yadi.sk/d/m0s5ztpj3XnUyw smoothly too - it's a pity my friends haven't more powerful processors.

haswell CPU are well know to handle AVX inefficiently.

Because it's offtop here can you give link I can read about?
Btw I set my vertical frequency output of videocard according the top value of Hz in madVR's Ctrl+J and usually haven't problems - usually 1 frame repeat (or drop) every (some) hours.

there is no collection that list these i know about.
smoothmotion is fixing this evne through the OSD still shows it trivial to use.
reclock does fix this and a lot of player have an alternative.
and the advanced fix is custom resolution.

Originally Posted by NikosD

Core i3 9100F - Win 10 1909 (18363.693)
DXVA Checker v4.2.1
LAV Filters v0.74.1-32 x64

Decoding mode: (FPS)
42/52/83 CPU usage 89% (average)

far worse then expected.

Originally Posted by VictorLS

NikosD
Thank you very much for your experiments proving my presumes (except 'Using EVR-CP, it plays fine at 50fps using ~90% CPU' - I thought i3 isn't enough, just not weakest i5 - btw what about with madVR?).

the name "i3 "isn't worth any thing anymore it just an i5 comapred to older CPUs now.

[retrue]

Hi, the last automatic build in https://files.1f0.de/lavf/nightly/ doesn't work.
I get a lot of mistakes and it doesn't allow me to install this LAVFilters-0.74.1-34 version.
No problem installing the previous versions.

[nevcairiel]

I'm sorry, but "doesn't work" isnt exactly something I can easily act on. There have been no changes in packaging since the -32 version.

[retrue]

I have downloaded it again and it works.
I guess my last download got corrupted.
Sorry for the bother.

[el Filou]

Edit: is it expected that the nightly builds since 0.74.1-32 are not signed anymore?

Originally Posted by nevcairiel

This is not actually true. CPU load doesn't require as much power, GPU is far worse, and its not like its running at 100%, and the display takes the top anyway.

Found some benchmarks of v0.5.0 on Android: https://medium.com/@ewoutterhoeven/a...ce-d3428221313
on the Dua Lipa YouTube test clip in 1080p, a Snapdragon 835 averages ~82 fps when using multi-core decoding. Considering 0.6.0 doesn't improve 8-bit decoding much, and a Snapdragon 865 is at best 2 times more powerful, that would give ~170 fps, so the CPU would still be constantly at 15% load on all cores, and this is the newest hardware.
It's doable, but if you watch any content for a long time it would still impact battery life much more compared to fixed-function decode of HEVC or VP9.
And this is with medium-bitrate 8-bit 1080p. Still not convinced it's worth it, but I'd love to see a battery life benchmark comparison with HEVC/VP9 on a modern device.

[VictorLS]

Originally Posted by huhn

haswell CPU are well know to handle AVX inefficiently.

I agree but it's much more powerful than mine Athlon(tm) II X3 460

Originally Posted by huhn

there is no collection that list these i know about.

I knew all you've mentioned - my way is custom resolution (by nVIDIA's Control Panel) equal or multiple of played file (stream). About "1 frame drop (repeat) every X seconds (minutes, hours)" in madVR's Ctrl+J statistic after ~ minute of playing is well sign all work fine - if many minutes or hours playing is well, if few seconds - something (few CPU and/or GPU power etc) wrong.

Originally Posted by huhn

far worse then expected

Can you provide your results (and average CPU load) with Ryzen 7 3700X while playing 4K 4:2:2 h265 20180428-175510_4K ENC 3 RMAD VS LEG.ts (563 MB) https://yadi.sk/d/m0s5ztpj3XnUyw and 8K AV1 HDR 60fps 31Mbps 831MB 3min30s files (and by the way X from madVR in i.e. graphstudionext64 in small window to avoid any postprocessing)?

Originally Posted by huhn

the name "i3" isn't worth any thing anymore it just an i5 comapred to older CPUs now.

I agree too.

[huhn]

~100 FPS at 56% CPU usage.

i'm not testing AV1 it's to early for this.

(and by the way X from madVR in i.e. graphstudionext64 in small window to avoid any postprocessing)?

no irrelevant.

[VictorLS]

Originally Posted by huhn

~100 FPS at 56% CPU usage.

Thanks.

Originally Posted by huhn

i'm not testing AV1 it's to early for this.

When you support nevcairiel in not adding not all frame decoding in LAV Video Decoder I logically presume you're sure as minimum your Ryzen 7 3700X don't need this function for almost all (comparatively often using) files (streams) including 8K AV1 HDR 60fps - after NikosD's experiments with such files if you sure your 3700X ~ 6 times faster than Core i3 9100F it worth to try otherwise stop writing not all frame decoding is useless function, please

Originally Posted by huhn

irrelevant.

So I presume vertical frequency was not 50Hz while you're playing 20180428-175510_4K ENC 3 RMAD VS LEG.ts (or you don't play this file to the end without seeking) but in case of 8K AV1 HDR 60fps you don't need to change vertical frequency because 60Hz is standard.

[huhn]

i don't show the video for performance test. i just do a decoding test.

the AV1 decoder is still not fully optimisedyet and it's useless 8k at an eye watering 31 mbit. i just take a different version which will most likely look better to the AV1 encoder has ways to go too.

[NikosD]

@nevcairiel

Hello.
It's been a year since LAV filters v0.74.1 release (19/03/2019)

Maybe it's time for v0.75 ?

They are not all familiar with nightly releases.

Just a thought.

[ryrynz]

Originally Posted by NikosD

Maybe it's time for v0.75 ?

Based on the commits I think it qualifies more as 0.74.2 :P

[NikosD]

Originally Posted by VictorLS

When you support nevcairiel in not adding not all frame decoding in LAV Video Decoder I logically presume you're sure as minimum your Ryzen 7 3700X don't need this function for almost all (comparatively often using) files (streams) including 8K AV1 HDR 60fps - after NikosD's experiments with such files if you sure your 3700X ~ 6 times faster than Core i3 9100F it worth to try otherwise stop writing not all frame decoding is useless function, please

Performance difference between Core i3 9100F and 3700X is heavily dependent on workload.
But on average we can safely support that 3700X is around 3 times faster.
Regarding video decoding and judging by the reported figures of average FPS and CPU usage during HEVC 4K 4:2:2 decoding, I could presume that the first 8K AV1 clip "Omagari Fireworks" could be decoded by 3700X at ~20 fps.

[VictorLS]

Originally Posted by NikosD

But on average we can safely support that 3700X is around 3 times faster...I could presume that the first 8K AV1 clip "Omagari Fireworks" could be decoded by 3700X at ~20 fps.

I thing ~ same but interesting real experiment with 0.74.1.34 (may be it's really faster than previous for AV1 decoding on modern processors)

[NikosD]

Originally Posted by VictorLS

I thing ~ same but interesting real experiment with 0.74.1.34 (may be it's really faster than previous for AV1 decoding on modern processors)

I have no issue testing my Core i3 9100F using latest 0.74.1-34 and dAV1d v0.6.0, but huhn should do that too.
For now, he is just evading the matter, using as an excuse the immaturity of dAV1d decoder.

[foobar2000]

why it's in beta for a year

[VictorLS]

Originally Posted by NikosD

I have no issue testing my Core i3 9100F using latest 0.74.1-34 and dAV1d v0.6.0

With same 8fps on FireWork and 11fps on Yellowstone (or some more fps)?

[NikosD]

Originally Posted by VictorLS

With same 8fps on FireWork and 11fps on Yellowstone (or some more fps)?

I haven't done the tests yet, but i really don't think that dAV1d v0.6.0 has any significant difference for 8 bit videos compared to v0.5.1