151

(46 replies, posted in Эксплуатация SVP)

pensioner600 wrote:

Hi guys. Sorry for the Google translation. There is 13900K+4090+DDR5+6800cl32. Monitor G7 1440p 240Hz. Before that I used DmitriRender for a very long time, and before it there was SVP for a long time, almost from the beginning of its appearance. With the new system I'm trying to switch to RIFE, since there are fewer artifacts. Can I ask you a few questions?
1. Used to MPC-BE and MPC-HC. Works well with them. Can I use them, or is MPV better?
2. Always used mpcVR. It works well too. Can I use it, or is madVR better? It's just that mpcVR converts HDR>SDR much more easily, using much less system resources than madVR.
3. For downscaling 4K>1440p I used mpcVR Jinc2m (with DmitriRender). Now at RIFE there is a lot of workload with this. The only working solution I found is After video frame size - Decrease to screen in SVP. If I understand correctly, this is downscaling BEFORE, not AFTER, and RIFE copes with this much easier. But this downscaling is of worse quality than Jinc2m. Is there any other option other than After video frame size - Decrease to screen? So that it does not affect performance.
4. GPU threads 4? Performance boost enabled?
5. What are Ensemble and V2? For some reason V2 does not work for me, and Ensemble puts almost twice as much load on the video card without obvious differences in the image. Is it better to use the regular 4.15 lite?
6. Is it possible to somehow increase the Scene change threshold? Since 15% is not enough and in very dynamic scenes jerks and freezes appear and I really don’t like it, it neutralizes all the smoothness. At least 20-25% is needed))
Thank you))

4. Не надо, хватит и двух потоков, если же ты делаешь все до 240 кадров, то тут зависит от разрешения, выше трех потоков прыгать нету смысла!
5. Енсембль вообще трогать не стоит, V2 подверсии легче перевариваются видеокартой и соответственно снижают требования для более высоких частот интерполяции!
- 5а. у меня 4070 и чтобы подгонять по производительности видеокарту, приходится делать 2 потока для 165 кадров (165Гц), и соответственно уменьшать размер кадра до старых добрых 720пипи!
- 5б. в твоем случае у тебя хватит производительности чтобы подгонять это дело с 4.9 (4.15лайт) для полноценных 240 кадров и соответственно те же добрые 720пипи!
- 5в. в 240 кадрах нету смысла ибо РИФЕ столкнется с проблемой SVP алгоритма, когда чрезмерная интерполяция начинает дублировать свои же собственные созданные кадры (зависит от разрешения)!
6. Порог смены сцен работает иначе с каждой новой версией, например 30% для 4.4 это 20% для 4.14 и так по нарастающей!

abraxas wrote:
RickyAstle98 wrote:

HAGS ON is 25% perf drop? I think it would increase performance, no?

There isn't a perf drop in every situation. I did some research earlier, it was a "big" problem especially with Windows 10 and - later on - with 30xx GPUs from Nvidia. With the 40s you can actually nearly always switch on HAGS, in particular it is a mandatory requirement for DLSS3. Also important seems to be a CPU, which is always bored anyway (i.e. the CPU has no problem doing the scheduling work).

My own (not very extensive) tests showed that there is no recognizable difference in RIFE with my 4070 and i9 12900KF (Win11 24H2, Build 26100.1). Note: I'm not really interested in pure FPS, I'm interested in whether I can see/"feel" it and how the load on the GPU develops (i.e. how much room for improvement there is). With the 4K movies, which I downscale to slightly more than 1440p, there was no difference at x2.

But - your milage may vary, I suggest doing some testing yourself. Especially if you are using VR.

Its not requirement for DLSS3, its requirement for DLSS frame generation itself, and I tested right now, I see about 6% performance increase with HAGS OFF! In games all FPS are the same, RTX 4070/12700!

aloola wrote:
unrealit wrote:

Hi!

Sorry for the stupid Question!

But , How can i get Benchmark Values?

And what is HAGS?

What do you mean?

Thanks smile

use the script here.
https://github.com/AmusementClub/vs-mlrt/discussions/19

HAGS = Hardware Accelerated GPU Scheduling
you could google it for more info.

In RIFE with HAGS on performance might drop up to 25%.

HAGS ON is 25% perf drop? I think it would increase performance, no?

unreality wrote:

Unfortunately it doesn't work either. I would be satisfied if I could lower the 4K (3080x2160p) resolution a bit. but how do I do that? I set SVP resize to 19201080, but it doesn't do that?

Frame resize -192001080

dawkinscm wrote:
RickyAstle98 wrote:

Please reatach my reedited comment again, dont reply to it, will be deleted smile

I deleted my reply so i don't have anything to reattach. Don't worry about it smile

*DELETED*

dawkinscm wrote:
RickyAstle98 wrote:
dawkinscm wrote:

Is that for build_optimization or running at 72fps.

My RTX 4070 doing perfect smooth 4.4-4.9 models, 4.10+ models not smooth, with or without build optimization or running at 72fps or dropped frames...
Update : opt level 1 helps 4.10 to 4.14 models for smoothness...

That's interesting but I don't like the fast action artefacts the models below 4.15 produce. Running at 72fps there are some slow pan issues on a couple of movies but for now I will stick with v4.15 lite/72fps with maybe the occasional dip back down to 60fps if I see any major issues for a particular movie.

Please reatach my reedited comment again, dont reply to it, will be deleted smile

dawkinscm wrote:
RickyAstle98 wrote:
dawkinscm wrote:

Update:
I think my initial impressions were correct. For certain artefacts, Rife v4.15 lite 72fps works almost as well as Rife 4.15 60fps. For others they are about the same with both running at 60fps. So Rife v4.15 lite 72fps is not quite as smooth at 60fps but it minimises certain artefacts while being  smooth enough to not be a problem.

In my case problem not with artefacts, frames are mismatched or delayed a bit!

Is that for build_optimization or running at 72fps.

My RTX 4070 : 4.4 to 4.9 models perfectly smooth at any factors at any fps (no build optimization or smth)
Models 4.10 to 4.16 not smooth at all (looks like every Xth frame delayed or mismatched like stutters every 1 second)
Update : opt level 1 helps recover smoothness for 4.10 to 4.14 models!

dawkinscm wrote:
RickyAstle98 wrote:
dawkinscm wrote:

Yeah I just tried it and if it is making a difference, it's at the margins and maybe pushing 4.15 lite closer to 4.15. But I will need to test with 4.15 to be sure. Also I'm not used to waiting this long for engine build anymore lol.

Hehe!

Update:
I think my initial impressions were correct. For certain artefacts, Rife v4.15 lite 72fps works almost as well as Rife 4.15 60fps. For others they are about the same with both running at 60fps. So Rife v4.15 lite 72fps is not quite as smooth at 60fps but it minimises certain artefacts while being  smooth enough to not be a problem.

In my case problem not with artefacts, frames are mismatched or delayed a bit!

dawkinscm wrote:
RickyAstle98 wrote:

But what TRT doing in my case? No drops no stutters but not smooth (any after 4.10), even with new optimization level, ah yes wait >
Builder optimization level 5 (default 3), adds another 3 minutes to build engine cache lock and only 3% performance!

Yeah I just tried it and if it is making a difference, it's at the margins and maybe pushing 4.15 lite closer to 4.15. But I will need to test with 4.15 to be sure. Also I'm not used to waiting this long for engine build anymore lol.

Hehe!

dawkinscm wrote:
RickyAstle98 wrote:

But what TRT doing in my case? No drops no stutters but not smooth (any after 4.10), even with new optimization level, ah yes wait >
Builder optimization level 5 (default 3), adds another 3 minutes to build engine cache lock and only 3% performance!

Yeah I just tried it and if it is making a difference, it's at the margins and maybe pushing 4.15 lite closer to 4.15. But I will need to test with 4.15 to be sure. Also I'm not used to waiting this long for engine build anymore lol.

Simple test with fastest model >
4.4 opt level 1 > engine build time ~32 seconds (-2% performance)
4.4 opt level 2 > engine build time ~70 seconds (-1% performance)
4.4 opt level 3 > engine build time ~120 seconds *
4.4 opt level 5 > engine build time ~300 seconds (+3% performance)

dawkinscm wrote:
RickyAstle98 wrote:

With TRT 10.0 my RTX 4070 struggles to 24>96 hd source (4.16 lite realtime), but easy 24>120 with TRT 9.2 hehe!
Maybe you know solution, why any models after 4.10 not smooth enough? But difference between 4.12 and 4.4 a whole pit, but SMOOTHNESS drop?
Maybe its frame pacing problem, just listen, 4.4v2 24>72 looks very smooth, but has a lot artefacts (no dropped frames), starting 4.10 everything looks like every 3th frame are mismatched (still no dropped frames)!

As I said, trt 10.0 is "one to avoid" because it cuts performance by around 30% while possibly doubling GPU usage.


aloola wrote:

has anyone tried the option builder_optimization_level=5?
The resulting engine may have better performance compared to an engine built with a lower optimization level.

They key phrase there is "may" have better performance. Yes I've tried this a couple of times. It's supposed to try to find more potential heuristics from what I understand but I didn't notice any real difference. But I can't remember if I have tried with my increased workspace so I might try again.

But what TRT doing in my case? No drops no stutters but not smooth (any after 4.10), even with new optimization level, ah yes wait >
Builder optimization level 5 (default 3), adds another 3 minutes to build engine cache lock and only 3% performance!

dawkinscm wrote:

Two bits of news. A while ago I said that increasing the frame rate can help with certain fast movement artefacts. Well initial testing suggests that since v4.15 lite is actually "lite",  I'm now able to run SVP at 72fps and it is helping with certain specific types of fast movement artefacts. It's dropping exactly 12 frames a second which makes sense but the helpful effect is not being lost. I still need to do more testing but this looks promising.

Also, Tensorrt v10.0.0 is available. I couldn't believe how bad the benchmarks look but the notes confirm that this is definitely one to avoid.

With TRT 10.0 my RTX 4070 struggles to 24>96 hd source (4.16 lite realtime), but easy 24>120 with TRT 9.2 hehe!
Maybe you know solution, why any models after 4.10 not smooth enough? But difference between 4.12 and 4.4 a whole pit, but SMOOTHNESS drop?
Maybe its frame pacing problem, just listen, 4.4v2 24>72 looks very smooth, but has a lot artefacts (no dropped frames), starting 4.10 everything looks like every 3th frame are mismatched (still no dropped frames)!

SHTH34D wrote:
RickyAstle98 wrote:
oriento wrote:

where is it available ?

https://github.com/AmusementClub/vs-mlr … nal-models

Uhhh my virus software is saying this straight up has a trojan in it...

Who cares? I download TensorRT models from there, and my software saying no virus!

oriento wrote:
dawkinscm wrote:

Rife v4.15 is out. It is a clear improvement over previous models for reducing double images in fast movement. I need to do some testing but I think it uses a little more GPU than previous models because even with v2 my GPU never drops below 100%. Packet drops sometimes happen when starting or when FF/REW but settles down quickly with no further drops during playback. No obvious regressions from previous tests.

where is it available ?

https://github.com/AmusementClub/vs-mlr … nal-models

framo wrote:
flowreen91 wrote:

If done correctly, restart SVP and then you should be able to see selectable AI models with (v2) in their name like this:
https://gyazo.com/9d6d1360852e8e92f834425e69803cb0

im using ncnn/vulkan and theres no drop down menu for me, am i doing sthg wrong?
i downloaded the v4.14 model and named it "rife_v.4.14f" , put the name in the rife_trt_model section in application settings. how can i make sure it is using the correct model?

Huh?

dawkinscm wrote:
RickyAstle98 wrote:

Lets say about RTX per equivalent model performance >
RTX 2070 > v4.4
RTX 3070 > v4.9
RTX 4070 > v4.11

     RTX 4080 > 4.14 (lite) smile

I have almost same perf with lite and standard so yeah youre goddamn right!
RTX 5090 > v4.20

dawkinscm wrote:
aloola wrote:
dawkinscm wrote:

Isn't this a separate issue to the TensorRT bug you refer to?  I don't think the use of dynamic shapes reduces the memory footprint enough to help @oriento who is literally running out of memory. Using performance mode reduces the amount of memory being used. Lower number Rife models also use less memory. Maybe for a 20 series card a user should do both by using the SVP default Rife 4.9 which works almost as well as later Rife models but is less taxing on the GPU.

it's the same
performance boost off = dynamic shape
performance boost on = static shape

you can test it yourself with performance boost off + model v2 4.7+

Oh right, good to know. After all this time still learning new things. Thanks smile

Lets say about RTX per equivalent model performance >
RTX 2070 > v4.4
RTX 3070 > v4.9
RTX 4070 > v4.11

Kadet89 wrote:
anders.nilsson wrote:

Is there some comparing of RIFE and Standard svp engine? Cant find. Thanks

https://www.youtube.com/watch?v=ZhDc8p_GsZ8&t=208s

Bad comparison, but okay! Just for sure, generally RIFE are 6-8 times slower than SVP algo, much accurate, expensive!
At movies RIFE (starts 4.4) provides more smoothness and 70% playback time less artefacts (about 88% for 4.14 models)!
Example 2 : SVP algorithm cant do 6x interpolation accurately enough from low frame rate sources (24>144) because every 4/5th frame will blended together or duplicated (content dependant)!

In my case my chinese 165Hz GSync monitor doesnt breaks VRR d3d11va hardware decoder, and cant be out of range because 24fps content synced to 48Hz auto, less/x3Hz hehe

smile

flowreen91 wrote:
Blackfyre wrote:

It's banding and crushed blacks that are the issue.

Maybe Chainik is right.
Blackfyre can u please try to see if u can reproduce that exact scenario with:
RIFE disabled (normal interpolation) + SVP + MPV ? example: https://gyazo.com/8cedda0a3fe610f50810178b9e54da7b
How about
SVP (interpolation disabled) + MPV ? example: https://gyazo.com/21658ebad006e2bcf284a73e168be62e
How about
SVP fully closed + direct MPV launch ? example: https://gyazo.com/049c2643d14b0087b9c6354f7da73815
Would be nice to know that it's MPV fault cause of random decoders configurations and leave it at that xD

RickyAstle98 wrote:

Also I dont see need for RIFE+SVP content, because Lossless Scaling program updated to DXGI frame generation that only doubles framerate, VRR breaks, but its working, I do SVP to 82FPS and then enable LSFG and now I have 164FPS

RickyAstle98 I just tried the Lossless Scaling LSFG frame generation combined with MPV capped at 120 fps and i'm seeing major pixels wiggling issues.
It's like it tries to ununiformly speed up patches of screen, giving you an uneven movement like waves.
It looks like an abomination on this test scene: https://drive.google.com/drive/folders/ … drive_link
My suggestion would be to use the SVP resize parameter to resize video to a lower resolution until you hit the spot where your system can handle RIFE at "164 fps without dropping frames, issues, flickering, bla bla bla bla bla..."
That way SVP will fully control the scene movement at all times.

If you use Lossless Scaling app for the upscaling, i compared performance between using Lossless Scaling app with Nvidia Image Scaling ON and MPV with Nvidia Image Scaling shader and using the app was much slower.
You can add yourself the Nvidia Image Scaling shader by following the instructions here and compare the performance differences: https://drive.google.com/drive/folders/ … sp=sharing

I dont have pixel issues, 165Hz GSync monitor, my system with RTX4070 cant handle HD@165 starting 4.9 stable, so then I decide use LSFG with DXGI swap, better FPS before LSFG, better pixels/movement interpolation,
I said I tick RIFE to odd number 82 then LSFG to 164, all models do 164FPS without drops, even at FHD! Anime test videos doesnt count, besause I use LSFG only for movies and games! Generally as RIFE booster!
RTX 4070 can handle 82FPS at any RIFE models up to FHD but then LSFG that and I see any content at my refresh rate, which RIFE cannot, so what? Anime test videos not for LSFG procedures, who cares?!

dawkinscm wrote:

Interesting. I guess we were all wrong lol But I did ask for a clarification and all I got was answers from everyone except @Chainik hence my assumption. But at least now we know what's actually happening. It is still all good smile

I dont think its worth relying on Chainik, as the owner of the SVP4 license over 4 years, I couldnt even get through to them through official communication lines, they are probably afraid to answer specific questions or do not care,
despite the fact that there is also a Chainik in charge...
SVP does nothing with colors? Wrong! MPV does something with colors? Yes! Both of them do something, depends on codec, colormatrix, too many variables that connected inbetween both!
Generally no issues with SDR content, only HDR issues over 4 years!
Also I dont see need for RIFE+SVP content, because Lossless Scaling program updated to DXGI frame generation that only doubles framerate, VRR breaks, but its working, I do SVP to 82FPS and then enable LSFG and now I have 164FPS,
I can also see this new frames while playback and stop frames, even NVIDIA overlay says its 164FPS instead of 82 through resized window >
So generally I can watch 720/900p content at any models at 164FPS without dropping frames, issues, flickering, bla bla bla bla bla...

dawkinscm wrote:

Two things.
First is @Chainik saying that I no longer need to use copy? The manual still says use "copy-back".
Second, I thought @Blackfyre said that there was no difference between d3d11va and copy and I didn't see those differences in pictures. To me that made sense because it the mechanism should have nothing to do with color. But I see there is a massive difference in the two pictures which I must have somehow missed. But a possible explanation for the difference would be that it is being affected by SVP which brings me back to my first point.
So still confused.

Maybe its SVP issue? Because SVP do some things with HDR contents, converting, recoloring, etc etc etc...
Maybe SVP break HDR content? I really dont see a difference, I dont watch HDR content because my panel is fakeHDR sad

Blackfyre wrote:
dawkinscm wrote:

If you are using d3d11va then SVP doesn't work properly because it doesn't have full access to all frames.

Never had such an issue with RIFE + SVP + MPV

dawkinscm wrote:

Also hwdec has very little to do with colours which is why you see no differences. But it might have some impact on banding depending on what card you are using.

Pictures look exaggerated, captured with ExpertRAW on my phone. But you can understand what I mean by banding, and the blacks are not perfect blacks, they become very very dark grey, but not perfect black levels.

hwdec=d3d11va

https://i.imgur.com/5PmmjAB.jpeg

hwdec=d3d11va-copy

https://i.imgur.com/TwPlkb7.jpeg

As to this:

RickyAstle98 wrote:

Anyway if you have GSync monitor with physical module, this issue isnt happen, because renderer now doesnt request backup memory through VRR option!

It's an LG C2 42" which I use as a desktop monitor. It has both GSync and Freesync Premium.

Either way, I am still looking for a solution. Maybe someone else can chime in if they know of a workaround or different rendering method perhaps.

Do you force video-output-levels to full? Because SVP uses limited color levels by default, and can ignore full levels from NVCP params, my case I dont see any difference in colors/banding...

Blackfyre wrote:
RickyAstle98 wrote:

hwdec=d3d11va *prevents VRR
drm-vrr-enabled=yes *prevents VRR or zero effect
d3d11-sync-interval=1 *prevents VRR or zero effect
Only these parameters prevents VRR tech, tested 4 times!

removed all 3, VRR works, but same problem as vulkan now so black levels are crushed & banding.

removed drm-vrr-enabled=yes and d3d11-sync-interval=1 as they're useless it seems.

Testing only with:

hwdec=d3d11va *Prevents VRR but black levels are good, and no banding.

Changed to d3d11va-copy *VRR works but same problem as above crushed blacks & banding

nvdec, cuda, etc all same thing, do not work properly.

hwdec=d3d11va provides the best black levels and no banding, but breaks VRR.

Anyway if you have GSync monitor with physical module, this issue isnt happen, because renderer now doesnt request backup memory through VRR option!