madVR - high quality video renderer (GPU assisted)

[detmek]

Originally Posted by madshi

Is that only when using 10bit output, or also when using 8bit output?

Uhm, Pentium G3320 IGPU? Never heard of that yet! My best guess is that the driver are at fault. I've no such problem as you report with my HD4000 on Windows 8.1 x64, at least. MPDN calls D3D11 APIs in a different way. IMHO my way is better. In your specific situation my way seems to make problems, though, but I believe the GPU drivers are at fault for that.

Originally Posted by detmek

Its a Haswell Pentium processor. It uses GT1 GPU with 10EU units. HD4000 has 20EUs, I think. I may try with previous driver.

Nope, its not a driver. I even did something I wanted to do in last 10 days - I upgraded Windows 8.1 to Windows 10 TP 10074. Problem remains.

Driver version is different on Windows 10, it 10.18.15.4124. It doesn't matter if I use 10bit, 8bit or usual 7bit output, MPC-HC or MPC-BE, x86 or x64.

************
Upgrading from Win 8.1 to Win 10 was not a good idea. It created some system instability (high CPU usage). I had to refomat so I decided to try with Windows 7. Problem remains.

[aufkrawall]

Have created a tracker entry regarding the NV NNEDI3 quadrupling issue.

----------------

Regarding chroma scaling:
I think it was a good idea to remove the NNEDI3 chroma doubling options since they were pretty much redundant because of NNEDI3 chroma scaling and luma doubling/quadrupling options.
I think no one will miss it.

However, in the first post of this thread, it says:
"- no shortcuts, highest quality has priority over anything else"
which is not true anymore with that C-R chroma "shortcut".
I could probably search for days and wouldn't find a (good) example for doing chroma scaling completely via Jinc3. But who knows, I already stumbled over nasty artifacts with NNEDI3 chroma scaling. If it didn't matter how chroma is scaled, you could remove everything instead of bilinear and C-R for it.
It's not impossible that pure Jinc3 scaling can look better than C-R, while the other way round is highly unlikely.

I think MPDN has a nice compromise with those quality templates.
I really want to have a maximum quality mode without any shortcuts, like promised in the first post. It's also about the impression of not educating or teaching users what is good for them and what is not.
When using NNEDI3 image doubling and quadrupling with each 64 neurons, I really don't care if rendering is 1-2% slower due to pure Jinc3 chroma scaling. It doesn't matter.

[JPulowski]

Originally Posted by JPulowski

Hi madshi,

Thanks for including LumaSharpen. I did some testing and here is some feedback:

• Looks like currently you are using the old version of LumaSharpen, which is 1.4.1. Current version is 1.5.0, which is included with Reshade + SweetFX bundle. I have already modified it for MPC-HC use, you can find it here.
• Current madVR implementation uses Pattern 2 (9 tap gaussian, 4+1 texture fetches), although it looks nice on video games it doesn't look that good when it comes to videos. In my opinion, it would be better if you'd include the other patterns as options.
• This is more of a personal request: I am normally used to MPC-HC's shader management. For instance I do always apply LumaSharpen, no matter if there is resizing or not. With madVR even though I can use profiles it gets complicated in this case. Because sometimes I watch videos downsized in windowed mode, which LumaSharpen does not apply unless it is pre-resize. And pre-resize and post-resize(upscale) applications of LumaSharpen is quite different from each other and requires additional fiddling. It would be nice if there would be an option to apply LumaSharpen always as post-resize.

Also just to inform, after version 0.88.2 NNEDI3 upscaling/doubling artifacts came back for my NVIDIA GTX690. If I remember correct it was a Kepler-specific bug in NVIDIA drivers which was fortunately fixed with the driver version 350.12. I tried downgrading madVR but the issue is still there.

[Anime Viewer]

Originally Posted by madshi

About Jinc3 vs. Catmull-Rom: I've yet to see even a difference between Jinc3 vs *Bilinear*, when talking a replacement for chroma doubling. But if nobody can find any evidence for Jinc3 helping, then why allowing users to make a bad choice? Actually, it's not even that: Of course I could add just another setting to let users select the chroma doubling algorithm (e.g. Catmull-Rom or Jinc3), but really, I want to *reduce* the number of similar options, not increase them. The number of scaling options is already larger than I like. I will probably decrease them further at some point, too.

I agree. I see no difference between Jinc3 and Bilinear or any other scaling algorithm (other than DXVA2 which alters shading and color) when features like image doubling and enhancements like FineSharp and SuperRes are enabled. I am all for you eliminating upscaling, and other options (ex: certain dithering options, etc) that are no longer useful given new features implemented in madVR.

[huhn]

Originally Posted by JPulowski

Also just to inform, after version 0.88.2 NNEDI3 upscaling/doubling artifacts came back for my NVIDIA GTX690. If I remember correct it was a Kepler-specific bug in NVIDIA drivers which was fortunately fixed with the driver version 350.12. I tried downgrading madVR but the issue is still there.

this bug isn't fixed. after installing madVR you just have to use the 32 bit version of madVr first and the bug will trigger if you start the 64 bit version first the bug will not trigger.

[huhn]

Originally Posted by Anime Viewer

I agree. I see no difference between Jinc3 and Bilinear or any other scaling algorithm (other than DXVA2 which alters shading and color) when features like image doubling and enhancements like FineSharp and SuperRes are enabled. I am all for you eliminating upscaling, and other options (ex: certain dithering options, etc) that are no longer useful given new features implemented in madVR.

as far as i know DXVA is altered by settings in your GPU driver. and nvdia doesn't dither so you get a ton of banding.

[JPulowski]

Originally Posted by huhn

this bug isn't fixed. after installing madVR you just have to use the 32 bit version of madVr first and the bug will trigger if you start the 64 bit version first the bug will not trigger.

Just tested it and can confirm that it works. Thanks.

[Anime Viewer]

Originally Posted by huhn

as far as i know DXVA is altered by settings in your GPU driver. and nvdia doesn't dither so you get a ton of banding.

Its possible its gamma as opposed to color itself that may be shifting, so I'll look in my Nvidia gamma settings to see if that is indeed what is happening. I haven't seem any increased banding in my testing as of yet...

[madshi]

Originally Posted by bcec

I think it is the display (Samsung hu8550). When feeding RGB signal via HDMI, only 60Hz shows 10-bit image. However, if I change nVidia output to YCbCr, then I get 10-bit image at all refresh rates.

Ok. I would think twice about outputting YCbCr, though. It's better to output RGB. I'd suggest that you test in 60Hz 10bit vs 8bit (with dithering *enabled*), to find out if you can see any difference at all. Probably you won't see any difference, which means you can safely just stick to 8bit. Really, the 10bit mode is probably only useful for those very very few users who have a true native 10bit displays. 99.9% of today's displays are really only 8bit or less, and they dither internally.

Originally Posted by huhn

but that's the problem i get tearing with 8.
and don't worry doesn't make any sense to me too...

Oh, I see. Somebody else had the same problem in Windows 10, and got rid of it by downgrading to a non WDDM-3.0 driver, or something like that. I think it was aufkrawall, IIRC?

Originally Posted by thegilpins

Hi with version 88.x I get madvr crashes sometimes when opening or closing video files within potplayer. This only happens if I enable 'use Direct3D for presentation'. I have attached two crash files. This is easily reproducible.

You did also send me a crash report via email, right? So far I've not been able to reproduce the problem, unfortunately, but I'll try again later. This is probably a driver bug, but I can't say for sure right now.

Originally Posted by detmek

Nope, its not a driver. I even did something I wanted to do in last 10 days - I upgraded Windows 8.1 to Windows 10 TP 10074. Problem remains.

Driver version is different on Windows 10, it 10.18.15.4124. It doesn't matter if I use 10bit, 8bit or usual 7bit output, MPC-HC or MPC-BE, x86 or x64.

************
Upgrading from Win 8.1 to Win 10 was not a good idea. It created some system instability (high CPU usage). I had to refomat so I decided to try with Windows 7. Problem remains.

Ouch. I'm sorry that you spent so much time with different OSs just to find out it didn't help at all. Unfortunately I don't know what to do. I can't reproduce the problem you have on my PC. And it seems nobody else has the same problem which you have. I've no clue why.

If it would be very easy for you, maybe you could create a short digicam video, showing the problem you're seeing? But don't bother if it would be hard for you to do.

Originally Posted by aufkrawall

Have created a tracker entry regarding the NV NNEDI3 quadrupling issue.

Thanks.

Originally Posted by aufkrawall

However, in the first post of this thread, it says:
"- no shortcuts, highest quality has priority over anything else"
which is not true anymore with that C-R chroma "shortcut".
I could probably search for days and wouldn't find a (good) example for doing chroma scaling completely via Jinc3. But who knows, I already stumbled over nasty artifacts with NNEDI3 chroma scaling. If it didn't matter how chroma is scaled, you could remove everything instead of bilinear and C-R for it.
It's not impossible that pure Jinc3 scaling can look better than C-R, while the other way round is highly unlikely.

I think MPDN has a nice compromise with those quality templates.
I really want to have a maximum quality mode without any shortcuts, like promised in the first post. It's also about the impression of not educating or teaching users what is good for them and what is not.
When using NNEDI3 image doubling and quadrupling with each 64 neurons, I really don't care if rendering is 1-2% slower due to pure Jinc3 chroma scaling. It doesn't matter.

I could also use double float (64bit) for all madVR calculations and store all temp results in 32bit float. That would in theory increase the precision and exactness of madVR's rendering. But it wouldn't bring a visible improvement, and would slow down everything a lot.

The first page says "highest quality has priority over anything else", and that is still my goal. However, the point here is that I don't think that there's a visible difference between Catmull-Rom and Jinc for chroma doubling. The situation is different for chroma upscaling. There the difference is much bigger than for chroma doubling.

You won't get highest quality by using the most expensive algorithm for every part of the processing pipeline. You will get highest quality by spending your GPU for algorithms which make a difference. Instead of worrying about chroma doubling, have a look at FineSharp and SuperRes.

Look, I could have hidden what I'm doing. I could just have said nothing and hidden the Catmull-Rom information from the OSD. Nobody would ever have known. There are many such things hidden in various places. Not only in madVR, but also in MPDN. E.g. NEDI falls back to Lanczos *without* AR. Or SuperRes for chroma upsampling uses bilinear upscaling to upscale the error. These are fixed parts of the algorithms and no global setting in MPDN or madVR influences that. I guess you might call this "shortcut", but I think it only qualifies as a shortcut in a negative way if it ends up being visible. If I used the most expensive algorithms everywhere, madVR would be 5 times slower than it is now, and you wouldn't see the tiniest difference in quality. It's all about spending the GPU power wisely, so you have some left for algorithms which do make a visible difference.

But I'm willing to bow to the majority. What does everybody else say? Should I "allow" Jinc3 to be used for chroma doubling? Instead of blocking it and using Catmull-Rom instead?

Originally Posted by JPulowski

Also just to inform, after version 0.88.2 NNEDI3 upscaling/doubling artifacts came back for my NVIDIA GTX690. If I remember correct it was a Kepler-specific bug in NVIDIA drivers which was fortunately fixed with the driver version 350.12. I tried downgrading madVR but the issue is still there.

Nothing I can do about it.

BTW, I double checked and my version of LumaSharpen is identical to that from version 1.5.0, except that my code also has an optional experimental limiter in it. And I'm using the "best" pattern which is available in either 1.4.1 or 1.5.0, which is pattern 3.

-------

I've done some digging and hacking and raping Direct3D, I think I have the refresh rate problems sorted out. Give this build a try:

http://madshi.net/madVR885refreshRateFix.zip

Should work for D3D9 and D3D11 now, and also for 64bit. Once again, I've spent a lot of time to fix Microsoft's stupidity. They actually *intentionally* use 59.940Hz (23.976Hz) even if you ask for 60.000Hz (24.000Hz). There's no way to convince Direct3D to use 60.000 (24.000), except by going behind their back and twisting and binding their fingers, like I've done. I wonder if there's any other software besides madVR which can switch FSE mode to 24.000/60.000? I kinda doubt it...

[pirlouy]

Dude, I hope I wont ruin your day/night, but it does not work for me. It still switches to 59 when entering FullScreen.
Let's hope it's only me. xD

[MS-DOS]

Originally Posted by madshi

What does everybody else say? Should I "allow" Jinc3 to be used for chroma doubling? Instead of blocking it and using Catmull-Rom instead?

Someone suggested a trade quality option, it could be a good way to please placeboquality maniacs without sacrificing the speed for everyone else.
I'm fine with CR. Not sure if its much faster than, say, Lanc3AR though.

[Fullmetal Encoder]

Originally Posted by madshi

But I'm willing to bow to the majority. What does everybody else say? Should I "allow" Jinc3 to be used for chroma doubling?

I think you should use your best judgment on such matters. But I appreciate that you let us know what you are doing so that so that in some of these cases we can compare the different algorithms used on different content. I think that's important for educating ourselves on the real world value of what we're using.

[huhn]

Originally Posted by madshi

But I'm willing to bow to the majority. What does everybody else say? Should I "allow" Jinc3 to be used for chroma doubling? Instead of blocking it and using Catmull-Rom instead?

i know content where nnedi3 chroma upsampling does have a very huge effect... pixel art computer games. but i don't think that matters.

leave it as it is i wouldn't waste power on such a thing.

[madshi]

Originally Posted by pirlouy

Dude, I hope I wont ruin your day/night, but it does not work for me. It still switches to 59 when entering FullScreen.
Let's hope it's only me. xD

Did it ever work for you, in D3D9, I mean?

Originally Posted by huhn

i know content where nnedi3 chroma upsampling does have a very huge effect... pixel art computer games. but i don't think that matters.

FWIW, NNEDI3 is still supported for chroma upsampling, just not for chroma doubling. Could you test that pixel art games for quality differences between 1) v0.87.21 NNEDI3 chroma doubling 2) v0.87.21 Jinc3 AR chroma doubling 3) Catmull-Rom AR chroma doubling?

[huhn]

Originally Posted by madshi

FWIW, NNEDI3 is still supported for chroma upsampling, just not for chroma doubling. Could you test that pixel art games for quality differences between 1) v0.87.21 NNEDI3 chroma doubling 2) v0.87.21 Jinc3 AR chroma doubling 3) Catmull-Rom AR chroma doubling?

sorry should be more clear. even quadrupling has a huge effect compared to jinc or other resizer.
so i'm talking about the even more demanding setting like the usually total over kill called NNEDI3 chroma doubling.
it's a 320x200 pixel art game stored in RGB not a real source.

[AngelGraves13]

Seeking is problematic unless pre-rendered frames are set to Application.

[aufkrawall]

Originally Posted by madshi

I could also use double float (64bit) for all madVR calculations and store all temp results in 32bit float. That would in theory increase the precision and exactness of madVR's rendering. But it wouldn't bring a visible improvement, and would slow down everything a lot.

I'm not exactly sure if this is a good comparison...

Originally Posted by madshi

The first page says "highest quality has priority over anything else", and that is still my goal. However, the point here is that I don't think that there's a visible difference between Catmull-Rom and Jinc for chroma doubling. The situation is different for chroma upscaling. There the difference is much bigger than for chroma doubling.

Good point.
Btw: How'd I use chroma doubling with 0.87? Testing with 4:2:2 content? If yes, this is way too much effort.

Originally Posted by madshi

You will get highest quality by spending your GPU for algorithms which make a difference.

Yes, but afaik, C-R is hardly better in any situation compared to Jinc3.

Originally Posted by madshi

Instead of worrying about chroma doubling, have a look at FineSharp and SuperRes.

SuperRes for luma is tempting.
However, I'd really like to have a deblocking filter because sharpening can make the blocking way too obtrusive.

I know this may seem like a fuss about nothing, but I still think this optimization can make much sense for slow GPUs, but not for fast ones.
Maxwell was a huge boost for HTPC and 16nm GPUs are probably not very far either.

[madshi]

Originally Posted by huhn

sorry should be more clear. even quadrupling has a huge effect compared to jinc or other resizer.
so i'm talking about the even more demanding setting like the usually total over kill called NNEDI3 chroma doubling.
it's a 320x200 pixel art game stored in RGB not a real source.

Luma quadrupling yes. But chroma doubling/quadrupling? I'd like to see a comparison screenshot.

Does the test build fix the refresh rate problem for you?

Originally Posted by AngelGraves13

With D3D11 Presentation enabled, seeking in MP4 files causes the video to skip in a loop. Disabling it fixes it.

I doubt that it has to do with MP4 files.

[AngelGraves13]

Originally Posted by madshi

Luma quadrupling yes. But chroma doubling/quadrupling? I'd like to see a comparison screenshot.

Does the test build fix the refresh rate problem for you?


I doubt that it has to do with MP4 files.

Fixed by setting Pre-rendered frames to Application instead of 1 in NVIDIA Control Panel.

[mark0077]

madshi, this is a question, not specifically related to the newer versions of madVR so absolutely no priority in a reply to this (FYI using the latest 0.88.5 version)

I just got a GTX 980 and have it overclocked as much as I can push it, so I have been experimenting with NNEDI 32 neurons for image doubling / quadrupling. Quadrupling seems to be working fine at 32 neurons on very low resolution files. I'm now testing of a movie which is 1280 x 720, so to get to 1920 x 1080 I can see the following is being used in Ctrl-J stats

chroma > Jinc3 AR
luma > Nnedi32 < Catmull-Rom AR
chroma > Catmull-Rom AR

This seems to be alot more demanding than the quadrupling of the lower resolution files, which is fine, so in gpu-z I can see GPU Load hitting ~90% but my question is.

I don't see any dropped frames, delayed frames or presentation glitches, however the queues are as follows

decoder queue 15-16 / 16
upload queue 7-8 / 8
render queue 1-2 / 8
present queue 0-2 / 8

My question is, although I know my gpu is literally at its absolute limits, how come some of the queues seem full, while others are nearly empty all the time, yet my gpu isn't under enough pressure to drop frames. Absolutely very low priority question, just thought it was worth mentioning as I'm not sure if the render and present queues not being full in this scenario may be something madVR is controlling and could potentially be improved to have them more full seeing as frames aren't being dropped.