L-SMASH Source

[videoh]

Originally Posted by MeteorRain

I've got a 8K 60fps HEVC 10bit TV broadcasting stream that my freshly bought Ryzen 3600 can barely playback at 5 fps full CPU load. You want to see how $1200 GPU works compared to your $40 card? Try those. It can decode, downscale, process the image and then output to AviSynth at > 60 fps, comparing to, let's say, 3 fps on a 6 core CPU sold this year.

Amen, bro!

[Atak_Snajpera]

Originally Posted by MeteorRain

You are so funny.

Unfortunately I can't say the same about your lies. They are not funny at all.

[videoh]

Cool down Atak. You did use the epithet "pathetic". MeteorRain was merely analogizing: obsessing over a video decoder when the card brings so much more to the table is like obsessing over a closet in a house. I suppose you did understand that but calling someone a liar is so much cooler, right? From the guy that is too old to be childish!

To be honest, I'm having trouble understanding what your basic point is here. Is it that using GPU power is useless? Is it that 2080 Ti etc. are poor value? DG tools suck? DG sucks? What is your overall point?

[DJATOM]

Originally Posted by MeteorRain

People buy a house not because they want to have a slightly bigger closet in their room than your condo. People buy a RTX 2080Ti not because they want to show how 1% of its processing power improves 2% of the portion.

Your $1200 GPU argument makes no sense because its just a card he has. I have a $60 worth of GTX 950 and I can get exactly the same decoding speed that he got, and thus the same result. Saying people are pathetic buying a 1 million dollar house by comparing how big the closet is, is IMHO so funny that I literally laughed out load when I saw your words.

If you really want to see how big the difference that graphics card can do, decode and scale down some UHD to 1080P or 720P and check how big of the portion those card can take over. Even worse, I've got a 8K 60fps HEVC 10bit TV broadcasting stream that my freshly bought Ryzen 3600 can barely playback at 5 fps full CPU load. You want to see how $1200 GPU works compared to your $40 card? Try those. It can decode, downscale, process the image and then output to AviSynth at > 60 fps, comparing to, let's say, 3 fps on a 6 core CPU sold this year.

Yeah, good real word scenario. I've tried to decode such stream on my card and it's playing smoothly without any stutters while stuttering a lot with SW decoder and 3900X (40-60% utilization, so a believe it hits PCIe Gen3 x16 max throughput capacity).

MI from that video: https://pastebin.com/xB8CHFiS.

[videoh]

Originally Posted by DJATOM

so a believe it hits PCIe Gen3 x16 max throughput capacity

That's a great point. If GPU processing has an achilles heel, that is it. We can look forward to greater bandwidth with future generations. Also, to mitigate this I have been experimenting with my CUDASynth framework, which allows many full frame transfers over PCIe to be eliminated for a script with multiple filters. You can read about it at the DG forum.

[MeteorRain]

Originally Posted by DJATOM

SW decoder and 3900X (40-60% utilization, so a believe it hits PCIe Gen3 x16 max throughput capacity).

That means the video renderer can't keep up because of the limited PCIe bandwidth? Very interesting.

Maybe try reducing the frame rate down to 30 on MKV, and pin the video player process to half of the cores, and see how that works? That should tell us whether it's due to the bandwidth or due to, like, CPU scheduler.

[videoh]

We're having so much fun I decided to try with GPU encoding for this use case we are discussing:

DGSource() + NVENCC + 2080 Ti + 7700K: 1:05.

I tried to choose settings that produced results perceptually approximating x264 medium.

[LigH]

Isn't the video decoding done in a separate PureVideo decoder chip which may not have developed as much as the 3D rendering and shader operations? ... The closet was the room with the least needs for space in this house.

[FranceBB]

Originally Posted by MeteorRain

That means the video renderer can't keep up because of the limited PCIe bandwidth? Very interesting.

Also the number of lanes that the CPU has with the GPU is important. For instance, Xeon CPUs like the W-3265M has 64 PCI-E lanes, while its consumer counterpart probably has less lanes. The thing is that it's not only a matter of GPU, but also a matter of which CPU you are using. As a matter of fact, having a very good GPU with a not so good CPU will limit its efficiency.

[redbtn]

Can I safely index MKV (hevc) with audio tracks and subs for now? I read somewhere earlier that i need remux MKV with only video track for better indexing, cuz another tracks can impact on accuracy.

[videoh]

Originally Posted by LigH

Isn't the video decoding done in a separate PureVideo decoder chip which may not have developed as much as the 3D rendering and shader operations? ... The closet was the room with the least needs for space in this house.

Quite right, LigH. Seems to me that the main nVidia development effort goes into gaming-relevant things. We really need super high PCI bandwidths too.

[MeteorRain]

Originally Posted by redbtn

Can I safely index MKV (hevc) with audio tracks and subs for now? I read somewhere earlier that i need remux MKV with only video track for better indexing, cuz another tracks can impact on accuracy.

Fairly safe. It used to index all other tracks and would create huge index file. Now they are all skipped. Let us know if you hit issues again.

[StvG]

@HolyWu, I built LSMASHSource from your repo using this CMake file taken from @MeteorRain repo. When tested for seeking issues with that file my build has errors but your build is fine. Something wrong with my building process? ffmpeg 4.2.1 / 4.3 used without difference.
There is patch for replacing avresample with swresample in your repo but avisynth exlibs - #pragma comment( lib, "libavresample.a" ).

[HolyWu]

Originally Posted by StvG

@HolyWu, I built LSMASHSource from your repo using this CMake file taken from @MeteorRain repo. When tested for seeking issues with that file my build has errors but your build is fine. Something wrong with my building process?

It's kinda tricky as the workaround I did to fix VC-1 seeking issue was reverting back some changes in libavcodec's VC-1 parser which made LSMASHSource unhappy. The issue has been filed for years.

[StvG]

Thanks for the info.

[FNSCAR]

@HolyWu ,I downloaded your build from release page of your github but there is no vsLSMASHSource.dll in folder.
Does it support vapoursynth?

[ChaosKing]

Originally Posted by FNSCAR

@HolyWu ,I downloaded your build from release page of your github but there is no vsLSMASHSource.dll in folder.
Does it support vapoursynth?

It works for avisynth and vapoursynth.

[FNSCAR]

Originally Posted by ChaosKing

It works for avisynth and vapoursynth.

I update vapouraynth to latest R47.2 and it works.Thank you.

[DJATOM]

R47.2 is buggy. Stick with R46 or go to latest R48 test. It contains fixes for all spotted bugs of R47.

[Atak_Snajpera]

Question: Would be possible to use multiple threads to speed up indexing process? For example by using some kind of chunk indexing?
I'm asking because NVMe SSDs with read speed of 3GiB/s+ are getting much more affordable. I have Xeon E5-2690@3.2GHz (8C/16T) and I see that i'm experiencing a bottleneck due to single threaded code.
In this example I'm indexing 4k movie (~50Mbps) from RAM disk...

1,24GiB/s is a max I can get on single core. ffmsindex is even more cpu intensive (only ~0,5GiB/s).

Currently indexing of ~60 GiB 4k movie takes ~1 minute. With proper multi-threading we could easily reduce indexing time to ~15s!