madVR - high quality video renderer (GPU assisted)

[Niyawa]

Hey people, me again. The thing is, there's this new codec pack being developed, we call it KCP (http://haruhichan.com/forum/showthread.php?7545-KCP-%93Kawaii-Codec-Pack%94). And I'm looking for some tips regarding the scaling algorithms of madVR, since they want to implement it with different presets for different machines. I would like everyone's opinion if possible.

In this order: chroma upsampling - image upscaling - image downscaling.

Ultra-tier: Jinc 4 taps, AR enabled; Jinc 4 taps, AR enabled; Spline 4 taps. linear enabled. (for very powerful PCs)
High-tier: Mitchell-Netravali, AR enabled; Lanczos 4 taps, AR enabled; Spline 4 taps, linear enabled. (For reasonable powerful PCs)
Mid-tier: Mitchell-Netravali, AR disabled; Lanczos 3 taps, AR disabled, Spline 3 taps, linear disabled. (For average PCs)
Low-tier: Bilinear; Bilinear; Bilinear. (For reasonable not so powerful PCs)
Lowest-tier: madVR is not installed and EVR-CP is set to default renderer. (For really outdated/not powerful PCs)

Any opinions/suggestions/comments regarding it will be really helpful now. I already did some search on the matter, but I don't feel like it's enough.

[alizard]

Madshi, the problem is still present with the test build.

Here is the problem file source: link (http://thepiratebay.se/torrent/6886368/Meet_Joe_Black_1998_HD_720p_BRRip_5.1AAC_x264-ILPruny)

Maybe this will help you replicate the problem on your hardware, since I can't upload you the sample directly without the levels being magically corrected.

[Asmodian]

It would not surprise me if the E-350 is much too slow for MadVR. As a graphics device it seems to be quite a bit less powerful than Intel's HD-3000, maybe as little as half the graphics performance. It is even lower power than the HD-2000.

That plus this quote:
I made the same observation on my HD3000.

Its about fast enough for 24p/25p content, but 50p/60p will not work properly if you use expensive upscaling.
For the HD2000, it'll be slower. How much? I cannot comment on that.

I was kind of hoping that the next generation Intel iGPU might be fast enough for all upscaling, but we would need a 100% performance increase on shader performance to be on the safe side, and i don't see that happening yet. :(

Makes me think the E-350 just isn't strong enough.

note: This quote is from before Jinc resize so the expensive upscaling isn't that expensive.

[AndreaMG]

I don't really like to comment too much about future features. What I can say is that I'm not going to do any extra work to support NVidia's propriaraty 3D shutter glasses etc, if that's what you mean with "NVidia 3D Vision".

Sorry, I meant Nvidia 3DTVPlay, or whatever "software magic" is making programs like Powerdvd be able to play 3D blu-rays, are you gonna do that too? I know you do not like to talk about your future plans but you talked about 3D first when you asked us if we were interested in 3D (remember?:)) so I think my curiosity is legitimate;). Anyway I was only wandering if it would be technically possible without being Cyberlink... Thanks anyway and keep up the good work:)

[Asmodian]

Mid-tier: Mitchell-Netravali, AR disabled; Lanczos 3 taps, AR disabled, Spline 3 taps, linear disabled. (For average PCs)

I understand Catmull-Rom AR with Linear Light to be a very good downscaling method that is also pretty fast. I am not sure how it compares to Spline w/o AR or linear light for GPU performance but I think it would be higher quality.

I hope you plan to base these settings off of the GPU not the CPU?

[robl45]

so from reading here, if I set my system to 24hz refresh rate as my projector isn't working all that great on 23.976, then reclock should take care of everything and run the 23.976 movies at 24 without frame drops and stuff? when i've tried anything except almost dead 23.976, the OSD on madvr shows that i'm going to get frame drops like every minute or so. is the OSD just wrong?

[Niyawa]

I understand Catmull-Rom AR with Linear Light to be a very good downscaling method that is also pretty fast. I am not sure how it compares to Spline w/o AR or linear light for GPU performance but I think it would be higher quality.
Spline is said to have the most accurate image quality while still being fastest, and it never lagged here either. So this is one was the best option for now.

Just tested with that one, and it lagged a lot, I disabled AR and linear and stopped lagging, but Spline still seems to be a little more closer to the original.

I hope you plan to base these settings off of the GPU not the CPU?
Yes. The mid to lowest-tier were made based on my own GPU (Intel HD4500), while the other 2 based on search and others opinions.

so from reading here, if I set my system to 24hz refresh rate as my projector isn't working all that great on 23.976, then reclock should take care of everything and run the 23.976 movies at 24 without frame drops and stuff? when i've tried anything except almost dead 23.976, the OSD on madvr shows that i'm going to get frame drops like every minute or so. is the OSD just wrong?
Nope. Bugmen0t already explained that madVR won't change to 24@60p because it doesn't do well with it, but ReClock will still change the timing stats, and that's what is misleading.

[DragonQ]

Hey people, me again. The thing is, there's this new codec pack being developed, we call it KCP (http://haruhichan.com/forum/showthread.php?7545-KCP-%93Kawaii-Codec-Pack%94).
What can it/will it do that LAV can't? If nothing, then I suggest you stop wasting your time. :sly:

EDIT: Oh I see, it's like a media pack kinda thing that includes the player, codecs, renderer and subtitle renderer. Hmm.

[Niyawa]

What can it/will it do that LAV can't? If nothing, then I suggest you stop wasting your time. :sly:
It's a pack for anime (includes LAV + madVR + xy-VSFilter + Haali). And since CCCP is not optimal for quality and performance, this pack will.

Edit: Yes. Haha. To put it bluntly, it's basically the same as my guide, but automatically. Basically for lazy people.

[aufkrawall]

It would not surprise me if the E-350 is much too slow for MadVR. As a graphics device it seems to be quite a bit less powerful than Intel's HD-3000, maybe as little as half the graphics performance. It is even lower power than the HD-2000.

Ok, I wasn't aware that it's that slow. I didn't buy it and I don't really care, I just hoped that there would be an option about madVR on them.
Too bad that they have postponed Kaveri with GCN GPU.


That plus this quote:


Makes me think the E-350 just isn't strong enough.

Regarding Intel: Since SNB HD 3000 is capable of Lanczos3 chroma scaling for most content, I'm confident that there'll be enough power with Haswell GT3 for even 1080p60 or additional Lanczos luma scaling.


note: This quote is from before Jinc resize so the expensive upscaling isn't that expensive.
4k won't be possible :D but maybe 1080p24.

[robl45]

Nope. Bugmen0t already explained that madVR won't change to 24@60p because it doesn't do well with it, but ReClock will still change the timing stats, and that's what is misleading.

thats not what I asked, I am currently running 23.976 refresh and want to switch to 24 refresh rate, no one said anything about 60hz.

[Niyawa]

thats not what I asked, I am currently running 23.976 refresh and want to switch to 24 refresh rate, no one said anything about 60hz.
Oh I see, I misread that, my apologies. madshi made a comment about it in page 817 if I remember correctly. So it works.

[leeperry]

I don't really know. When I wrote that code I think I asked what values I should use a long time ago in this thread, and the answer I got was what madVR does now. Of course nothing is set in stone. If the general consensus is to detect > 720 as HD now, I'm willing to consider that. However, the consensus was different when I started writing madVR. I'm not sure if the situation has changed. So why would it be different now than it was when I asked the last time? So I'll need more than 2 people confirming that madVR's behaviour should be changed. Otherwise I'm in danger of changing now, and then someone else will cry about all his PAL files being misdetected as HD by the latest madVR build.
I think this whole story started with HR having this manual option for 601/709 and we/I whined for automatic detection.....then the 1024*576 magic numbers came up, but that was years ago and at the time ppl were feeding HR with upscaled video from ffdshow(1024*576 made a lot of sense for upscaled PAL DVD's) and there wasn't nearly as much HD content being encoded to SD as nowadays anyway.

Now half a decade later, things have changed: most mVR users will prefer to let it upscale SD and there's now a truckload of HD content(whatever BD/HDTV or consumer camera) being encoded to SD....so providing us with a simple "SD/HD frontier" setting sure would help IMO. You would leave it on the current figures and I would prolly set it to something like "958x538" et voilà my 1/4 HD movies will use the 709 matrix/gamut :)

every added option makes the settings dialog more difficult. I don't want to end up with an ffdshow like settings dialog with 3 billion different settings. Do you know why ffdshow has a settings dialog like that? Because there was no one saying "no" when users asked for all kinds of funny options.
I don't think I would agree here as IME coders who worked on ffdshow were either porting avisynth scripts, having fun adding new capabilities(such as Haruhiko's RGB32 dithering and conversion) or simply adding settings for their own personal use(like all the crapola albain added in the sub-menus of the decoders with the ulterior motive of serving his own (shareware IIRC) MediaPortal project(which was rejected BTW (http://forum.team-mediaportal.com/threads/integration-with-ffdshow.82613/), so we get all this **** for nothing huh).

FFdshow is what happens when you have a piece of code being modified by dozens coders and some of them being newbies (http://www.viva64.com/en/b/0161/), others pro coders with time to kill and others with ulterior motives....all this without a project manager who would say "no". I don't think ffdshow coders ever cared much about end-users whining tbh, and I wouldn't expect them to considering that they were all essentially working for free.

You could always add an "advanced settings" page for all the stuff only OCD'ed users would ever care for :devil:


Shouldn't the encoders these days know that they should store the needed matrix and primaries into the bitstream?
Well, with x264 most certainly but with AVI/XVID and WMV3 this is a no-go AFAIK.

[nevcairiel]

(1024*576 made a lot of sense for upscaled PAL DVD's)

...what?
Thats not an upscaled PAL DVD, thats a normal PAL DVD after AR correction to get square pixels. Its also not a "magic number", thats a 720x576 DVD stretched to 16:9 AR.

On the topic, i say stick to the logic, no reason to break files that worked fine until now, rather teach people creating new files to do proper encodes.
The old files won't go away or fix themself, but new files may get into better shape.

[leeperry]

Thats not an upscaled PAL DVD, thats a normal PAL DVD after AR correction to get square pixels. Its also not a "magic number", thats a 720x576 DVD stretched to 16:9 AR.
Well, 720*576 turns into 1024*576 once you apply the proper 16/9 DAR but at the time most everyone was doing it in ffdshow AFAIK so HR was essentially seeing 1024*576. I'm really not sure that you could get the 16/9 DAR value from the original VOB files to HR....and even if you could, upscaling using Spline in ffdshow looked better than bicubic in HR in many ppl's opinions. Also, HR was giving "ghost lines (http://img520.imageshack.us/img520/4422/linesyl8.png)" due to broken code in its HLSL scaler so better feed it 1:1(the same code from Haali was in MPC-HC and Casimir666 fixed it by changing a coeff from 1.0 to 0.98 (http://mpc-hc.svn.sourceforge.net/viewvc/mpc-hc?view=revision&revision=380))

The "magic" term was being sarcastic, but I'm sure you got that ;)

On the topic, i say stick to the logic, no reason to break files that worked fine until now, rather teach people creating new files to do proper encodes.
The old files won't go away or fix themself, but new files may get into better shape.
I don't see what the big deal would be to allow the end-user to set his own SD/HD frontier setting, leave it as is by default et voilà everyone's happy :)

Oh yah, learning all the one-click script kiddies the diff between 601 and 709, good luck with that...especially when the conversion is done in 8bit, I believe it will be less harmful to keep the 709 coeffs encoded as is and revert them properly at the last stage in mVR than going 709>601>mVR.

And 29.97fps HD encoded to SD will still require the REC.709 gamut so do you also expect encoders to roll gamuts when encoding HD to SD from REC.709 to SMPTE-C? PQ will suffer quite a bit after the successive matrix and gamut 8bit conversions http://forum-images.hardware.fr/icones/smilies/whistle.gif

PS: and now mVR can detect the original dimensions of the stream being decoded by ffdshow, so I don't really see how a 1024*576 PAL DVD could ever occur IRL.

PPS: For some reason, the matrix tag works fine on folder/filenames when using LAV's splitter on individual files but it doesn't work at all when using PotP's seamless playback feature, would you like a log or could you please look into it?

:thanks:

[turbojet]

Madshi, let me know if you want me to report the problem to MPC-BE devs although I'm 0 for 4 on getting a reply from them. Unfortunately there's no changelog for these builds, just the stable build which crashes with madvr 85.x on an extended display.

Thinking more about 601/709 changing to 720x would help but it wouldn't solve my issue with the old hdtv caps I did with xvid 8-9 years ago I did at 720x or less to play on a single core cpu before x264 or colormatrix. But they are tagged 'hdtv' for these a customizable keyword would be best. I understand you not wanting to add a bunch of settings like ffdshow. Isn't there a few settings you can set with files in madvr directory now? How about 601 and 709 files containing keywords with wildcards that sets matrix as 2nd priority after bitstream?

[Hwasin]

I'm curious. Why does MadVR use so much GPU power when the video is paused?

[madshi]

I've tested the E-350 with full floating point processing: It fails hard, terrible frame drops.
So, it's too slow?
Oooh, that's interesting... :) Without full/half floating point processing, basically EVR converts to 8bit RGB, I believe. madVR refuses to do that because it's too low quality. So it seems EVR isn't really faster than madVR in this case, it just uses lower (too low) quality by default than madVR.

You could try the test build from yesterday (see a few posts up), which now uses float16 instead of float32. Maybe that will be enough to make madVR run fluidly on your E-350? What happens if you switch EVR to half floating point processing? Is that running smoothly?

In this order: chroma upsampling - image upscaling - image downscaling.

Ultra-tier: Jinc 4 taps, AR enabled; Jinc 4 taps, AR enabled; Spline 4 taps. linear enabled. (for very powerful PCs)
High-tier: Mitchell-Netravali, AR enabled; Lanczos 4 taps, AR enabled; Spline 4 taps, linear enabled. (For reasonable powerful PCs)
Mid-tier: Mitchell-Netravali, AR disabled; Lanczos 3 taps, AR disabled, Spline 3 taps, linear disabled. (For average PCs)
Low-tier: Bilinear; Bilinear; Bilinear. (For reasonable not so powerful PCs)
Lowest-tier: madVR is not installed and EVR-CP is set to default renderer. (For really outdated/not powerful PCs)

Any opinions/suggestions/comments regarding it will be really helpful now. I already did some search on the matter, but I don't feel like it's enough.
I'm not sure if Jinc4 is really noticeably better than Jinc3. Some people say yes, other people say they find Jinc3 to be sharper than Jinc4. Jinc4 might not be better, just different. So I wonder if it might make sense to use Jinc3 AR for Ultra-tier. It might leave some room to make upscaling fluid for 60p processing, or for 4K upscaling, because Jinc4 is quite a bit slower than Jinc3.

Choosing the right chroma upscaling algorithm when you can't use Jinc is a difficult topic. Different people have different opinions here. A lot also depends on the source material. If the source material has reasonably good quality and no "baked in" aliasing in the chroma channel, then using a sharp algorithm like Lanczos3 or Bicubic75 might make sense. If the source has really bad quality and/or the chroma channel has serious aliasing in it, then I'd rather recommend SoftCubic100. But since you have to choose one setting, you might want to test which looks better to you with a couple of video samples. In any case I think I would prefer Bicubic AR over Mitchell-Netravali AR.

Generally downscaling with linear light enabled has some benefits, but it also introduces quite a lot of heavy ringing. So I wouldn't ever use linear light without activating AR at the same time. Also, if you use linear light, it might make sense to stay away from anything more than 2 taps (Lanczos, Spline, Jinc) because 3+ taps algorithms have "wider" ringing than 2 taps algorithms. A good downscaling algorithm seems to be Bicubic AR with linear light enabled. I think that might be a better choice than Spline, due to the ringing problem when enabling linear light processing. If you disable linear light processing, then for downscaling choosing a 3 taps algorithm should be fine. Personally, I'd prefer Lanczos3 over Spline3, but that's just my personal preference. If you prefer Spline3, that's fine, too.

Madshi, the problem is still present with the test build.

Here is the problem file source: link (http://thepiratebay.se/torrent/6886368/Meet_Joe_Black_1998_HD_720p_BRRip_5.1AAC_x264-ILPruny)

Maybe this will help you replicate the problem on your hardware, since I can't upload you the sample directly without the levels being magically corrected.
Ouch. You said the levels were correct with v0.85.2, but not with v0.85.4. Could you please try v0.85.3? Then at least I know for sure which build introduced the problem.

Sorry, I meant Nvidia 3DTVPlay, or whatever "software magic" is making programs like Powerdvd be able to play 3D blu-rays, are you gonna do that too? I know you do not like to talk about your future plans but you talked about 3D first when you asked us if we were interested in 3D (remember?:)) so I think my curiosity is legitimate;). Anyway I was only wandering if it would be technically possible without being Cyberlink... Thanks anyway and keep up the good work:)
I haven't even started development, so I can't really say anything at this point. I simply don't know myself...

so from reading here, if I set my system to 24hz refresh rate as my projector isn't working all that great on 23.976, then reclock should take care of everything and run the 23.976 movies at 24 without frame drops and stuff? when i've tried anything except almost dead 23.976, the OSD on madvr shows that i'm going to get frame drops like every minute or so. is the OSD just wrong?
The frame drop estimate per minute by madVR is just that: An estimate. Don't put too much weight on that. Instead look at the real counted frame drops. And trust your eyes. And yes, Reclock should be able to make 23.976fps movies play smoothly at 24.000Hz.

I don't think I would agree here as IME coders who worked on ffdshow were either porting avisynth scripts, having fun adding new capabilities(such as Haruhiko's RGB32 dithering and conversion) or simply adding settings for their own personal use(like all the crapola albain added in the sub-menus of the decoders with the ulterior motive of serving his own (shareware IIRC) MediaPortal project(which was rejected BTW (http://forum.team-mediaportal.com/threads/integration-with-ffdshow.82613/), so we get all this **** for nothing huh).

FFdshow is what happens when you have a piece of code being modified by dozens coders and some of them being newbies (http://www.viva64.com/en/b/0161/), others pro coders with time to kill and others with ulterior motives....all this without a project manager who would say "no". I don't think ffdshow coders ever cared much about end-users whining tbh, and I wouldn't expect them to considering that they were all essentially working for free.
Well, in any case, nobody said "no", whether it was to end users or to developers. So that's why I'm saying "no" often enough to keep the madVR settings dialog from becoming too much of a mess.

You could always add an "advanced settings" page for all the stuff only OCD'ed users would ever care for
Yes, but I wonder when people finally understand that my time is limited and that it's not the right time right now to work on funky settings tweaks. You guys sometimes seem to think that you just need to bug me long enough with funny wishes and then they automatically become true. It's not that I can simply say "ok" and then magically something develops by itself. If I implement all those funny little things then you can kiss goodnight to any bigger improvements I had planned instead.

Thinking more about 601/709 changing to 720x would help but it wouldn't solve my issue with the old hdtv caps I did with xvid 8-9 years ago I did at 720x or less to play on a single core cpu before x264 or colormatrix. But they are tagged 'hdtv' for these a customizable keyword would be best. I understand you not wanting to add a bunch of settings like ffdshow. Isn't there a few settings you can set with files in madvr directory now? How about 601 and 709 files containing keywords with wildcards that sets matrix as 2nd priority after bitstream?
I don't want to add customizable keywords. That's *much* too much effort for just the handful people who would ever use that. I guess I could add a couple hard coded keywords, though, if it really means that much to you. E.g. "hdtv" or "bluray" or "blu-ray" to make madVR choose a 709 decoding matrix. But then, I wonder if that's really the right thing to do. After all, shouldn't a Blu-Ray, when downscaled to SD resolution, be converted to 601 matrix? So even if it has "bluray" or "hdtv" in the name, since it was obviously downscaled to SD resolution, I still think the proper matrix to detect is 601.

In the end the only reliable solution is to make all encoders understand that they should put the matrix and primaries into the video bitstream.

I'm curious. Why does MadVR use so much GPU power when the video is paused?
Which media player are you using? potplayer?

[hannes69]

http://madshi.net/madVR854colorTest2.rar

Does this fix the issue? @hannes69, could you please test this build, too?


No change, same behaviour like 0.85.4...

[madshi]

No change, same behaviour like 0.85.4...
Ok. alizard reported he has correct colors when using DXVA scaling with v0.85.2. Could you please check/confirm? If you have the same experience, please also check v0.85.3. This is only about *DXVA scaling*. No tests are needed with DXVA scaling turned off.

[secvensor]

You mean the mouse cursor? Do no media player hotkeys work in that situation at all? Or does just this one hotkey stop working? This specific sound control hotkey is a media player defined hotkey, correct? Or is it from some external software?


MPC-HC all ver. mouse scrooll and keyboard key's don't work in Exclusive mode (when cursore on exclusive timeline)
http://i54.fastpic.ru/big/2012/1221/dc/8f10ff6e79291eee8bd735f11d3970dc.jpeg

[leeperry]

Yes, but I wonder when people finally understand that my time is limited and that it's not the right time right now to work on funky settings tweaks. You guys sometimes seem to think that you just need to bug me long enough with funny wishes and then they automatically become true. It's not that I can simply say "ok" and then magically something develops by itself. If I implement all those funny little things then you can kiss goodnight to any bigger improvements I had planned instead.
Yep, I fully realize that and I wasn't even the one who whined for this in the first place....my only problem is that tags don't work when using the seamless playback feature of PotP so I can't fix my most pesky files. I've checked and Reclock doesn't see the filenames either(the "Creator source filter:" field in the mVR logs being empty) so without a custom SD/HD threshold setting I'm SOL on all those files..I'll try to ask the PotP coders to look into it but that might be complicated.

[6233638]

Hey people, me again. The thing is, there's this new codec pack being developed, we call it KCP (http://haruhichan.com/forum/showthread.php?7545-KCP-%93Kawaii-Codec-Pack%94). And I'm looking for some tips regarding the scaling algorithms of madVR, since they want to implement it with different presets for different machines. I would like everyone's opinion if possible.

In this order: chroma upsampling - image upscaling - image downscaling.

Ultra-tier: Jinc 4 taps, AR enabled; Jinc 4 taps, AR enabled; Spline 4 taps. linear enabled. (for very powerful PCs)
High-tier: Mitchell-Netravali, AR enabled; Lanczos 4 taps, AR enabled; Spline 4 taps, linear enabled. (For reasonable powerful PCs)
Mid-tier: Mitchell-Netravali, AR disabled; Lanczos 3 taps, AR disabled, Spline 3 taps, linear disabled. (For average PCs)
Low-tier: Bilinear; Bilinear; Bilinear. (For reasonable not so powerful PCs)
Lowest-tier: madVR is not installed and EVR-CP is set to default renderer. (For really outdated/not powerful PCs)

Any opinions/suggestions/comments regarding it will be really helpful now. I already did some search on the matter, but I don't feel like it's enough.These are bad defaults, and what is "best" typically depends on the type of content being watched, and user preference.

If your source material is acceptable with sharper upscaling algorithms, you should use Jinc 3 AR for Chroma & Luma upscaling.
Catmull-Rom anti-ring is the only option to use with linear light downscaling. If you can't run that, don't use linear light scaling.
Any other filters used with linear light scaling introduce too many artefacts.


Below that, I would suggest Bicubic 75 anti-ring Chroma, Mitchell-Netravali is too soft.
Lanczos 3 AR or Spline 3 AR if you prefer sharper upscaling filters.
I don't really have an opinion on what's the best downscaling filter to use if you can't run the anti-ringing filter with linear light scaling. You definitely want to avoid 4-tap filters though.


Below that, Bicubic 75 without the anti-ringing filter for chroma, or Mitchell-Netravali if your preference is lower ringing at the expense of softer/desaturated chroma.
I would then suggest SoftCubic 80, or Mitchell-Netravali without the anti-ringing filter for luma upscaling, depending on the quality of the source.
Again, not sure what I would suggest for downscaling, probably just Catmull-Rom without the anti-ringing or linear light options.

[DragonQ]

It's the anti-ringing that kills performance, at least on my setups. I can just about manage Bicbuic 75 AR chroma & Jinc3 luma, but I can't enable AR on both chroma and luma on any of my GPUs (nVidia GTS250, nVidia NVS 5200m, Intel HD 4000). The only one I haven't tested is my GT430...that might fare better but it's still a low-end card so I don't expect much.

[Weirdo]

I have madVR installed both manually (for MPC-HC & DVBViewer) in a custom directory and as part of JRiver Media Center. It seems madVR's settings cannot be separated between the two. For example, I want automatic exclusive mode enabled for MC and disabled for the others. Is there a way to do this?

[nevcairiel]

The settings are stored in the registry (among other things), so they are shared.

[hannes69]

Ok. alizard reported he has correct colors when using DXVA scaling with v0.85.2. Could you please check/confirm? If you have the same experience, please also check v0.85.3.

0.85.2 and 0.85.3 give the same results like 0.85.4, so sorry no confirmation.

[Weirdo]

The settings are stored in the registry (among other things), so they are shared.
I see, thanks. If it's not too much of a hassle, an option to store settings separately (in a file?) would be great.

[Niyawa]

Based on the recommendations of madshi and the (famous, haha) 6233638, I changed it accordingly:

In this order: chroma upsampling - image upscaling - image downscaling.

Ultra-tier: Jinc 3 taps, AR enabled; Jinc 3 taps, AR enabled; Lanczos 3 taps. AR enabled. (for very powerful PCs)
High-tier: Bicubic 75, AR enabled; Lanczos 3 taps, AR enabled; Lanczos 3 taps, linear disabled. (For reasonable powerful PCs)
Mid-tier: Mitchell-Netravali, AR disabled; Lanczos 3 taps, AR disabled, Spline 3 taps, linear disabled. (For average PCs)
Low-tier: Bilinear; Bilinear; Spline 3 taps, linear disabled. (For reasonable not so powerful PCs)
Lowest-tier: madVR is not installed and EVR-CP is set to default renderer. (For really outdated/not powerful PCs)

I believe Catmull-Rom is a little too soft in comparison to Lanczos for downscaling, and my priority is to make the image closest possible to the original resolution. For this same reason I've put Spline 3 taps for low-tier, since it's the closest to Lanczos while not being so power consuming. I've taken 6233638 advice on Jinc and Bicubic for chroma upscaling, they work well with Ultra and High but their rendering time for mid to lowest preset is too much though. If you guys has anything else to add that I forgot/misread/not putting in account, please do so.

[ajp_anton]

Spline 3 taps for low-tier, since it's the closest to Lanczos while not being so power consuming.I thought all n-tap filters were the same speed for any given n.

edit: Also, about the "lowest tier".
While I love madVR's quality, I think it should have a little more aggressive "trade quality for performance" options. On my desktop quality is of course high priority, but as I travel much and need long battery life, I not only have to trade away madVR's quality but also its features by going with another renderer.
Just a suggestion to expand madVR a little outside of its original purpose, so don't take it as a demand or anything =).

[6233638]

I believe Catmull-Rom is a little too soft in comparison to Lanczos for downscaling, and my priority is to make the image closest possible to the original resolution.Unless something has changed since I last did my testing, Catmull-Rom anti-ring scaling in linear light is by far the most accurate downscaling algorithm.

Sharper algorithms tend to introduce a lot more ringing, and with linear light scaling, particularly nasty dark ringing around the edges of high contrast areas.

Just like adding more taps does not necessarily make an upscaling filter better (actually, it usually makes most things worse) switching to another algorithm when downscaling in linear light has consequences that may not be apparent at a quick glance.

Some of the impact is lost here due to the reduced colour palette of a .gif, but it shows some of the problems you get when downsampling with Lanczos in linear light rather than Catmull-Rom, both with the anti-ringing filter enabled.

http://www.abload.de/thumb/catrometjna.gif (http://www.abload.de/img/catrometjna.gif)
(fireworks from Tron Legacy Blu-ray)

[Niyawa]

I thought all n-tap filters were the same speed for any given n.
Lanczos is a little more hard on my hardware than Spline. Not too much, but it does make a difference.

Unless something has changed since I last did my testing, Catmull-Rom anti-ring scaling in linear light is by far the most accurate downscaling algorithm.

Sharper algorithms tend to introduce a lot more ringing, and with linear light scaling, particularly nasty dark ringing around the edges of high contrast areas.

Just like adding more taps does not necessarily make an upscaling filter better (actually, it usually makes most things worse) switching to another algorithm when downscaling in linear light has consequences that may not be apparent at a quick glance.
I see... thanks for the input, I really usually trust more my eyes for those situations, so I would probably have never realized that.

As suggested by Mark0077 and with 6233638 suggestion, I applied some changes to make more clear how the presets will work.

Chroma Upsampling
Ultra-tier: Jinc 3 taps. AR enabled, Linear disabled.
High-tier: Bicubic 75. AR enabled, Linear disabled.
Mid-tier: Bicubic 75. AR disabled, Linear disabled.
Low-tier: Mitchell-Netravali. AR disabled, Linear disabled.
Lowest-tier: Bilinear.

Image Upscaling
Ultra-tier: Jinc 3 taps. AR enabled, Linear disabled.
High-tier: Lanczos 3 taps. AR enabled, Linear disabled.
Mid-tier: Lanczos 3 taps. AR enabled, Linear disabled.
Low-tier: Lanczos 3 taps. AR disabled, Linear disabled.
Lowest-tier: Bilinear.

Image Downscaling
Ultra-tier: Catmull-Rom. AR enabled, Linear enabled.
High-tier: Catmull-Rom. AR enabled, Linear enabled.
Mid-tier: Catmull-Rom. AR enabled, Linear disabled.
Low-tier: Catmull-Rom. AR disabled, Linear disabled.
Lowest-tier: Bilinear.

What do you guys think? Looks pretty sweet.

[nevcairiel]

Arent Bicubic and M-N both just cubic splines, which should be in the same performance class?

[alizard]

Madshi, 85.3 introduced the problem, but 85.4 fixed the problem for some files except for the file I provided a sample for. That file will only play with correct levels with 85.2. And if you're wondering, I did test that file with 85.3 also, no luck.

[Niyawa]

Arent Bicubic and M-N both just cubic splines, which should be in the same performance class?
They basically are, but some of the videos I played here, M-N did a better job with them, while Bicubic 75 did a few drops (almost nothing, but it did). Didn't test with 50 or 60 yet, but I don't believe it will make a huge difference.

[DragonQ]

Lanczos is a little more hard on my hardware than Spline. Not too much, but it does make a difference.


I see... thanks for the input, I really usually trust more my eyes for those situations, so I would probably have never realized that.

As suggested by Mark0077 and with 6233638 suggestion, I applied some changes to make more clear how the presets will work.

Chroma Upsampling
Ultra-tier: Jinc 3 taps. AR enabled, Linear disabled.
High-tier: Bicubic 75. AR enabled, Linear disabled.
Mid-tier: Bicubic 75. AR disabled, Linear disabled.
Low-tier: Mitchell-Netravali. AR disabled, Linear disabled.
Lowest-tier: Bilinear.

Image Upscaling
Ultra-tier: Jinc 3 taps. AR enabled, Linear disabled.
High-tier: Lanczos 3 taps. AR enabled, Linear disabled.
Mid-tier: Lanczos 3 taps. AR enabled, Linear disabled.
Low-tier: Lanczos 3 taps. AR disabled, Linear disabled.
Lowest-tier: Bilinear.

Image Downscaling
Ultra-tier: Catmull-Rom. AR enabled, Linear enabled.
High-tier: Catmull-Rom. AR enabled, Linear enabled.
Mid-tier: Catmull-Rom. AR enabled, Linear disabled.
Low-tier: Catmull-Rom. AR disabled, Linear disabled.
Lowest-tier: Bilinear.

What do you guys think? Looks pretty sweet.
Mid-Tier for me! :D

[6233638]

Arent Bicubic and M-N both just cubic splines, which should be in the same performance class?Yes, they should have the same performance.

If you are using the anti-ringing filter (which costs performance) my preference would be Bicubic 75.

If you are not using the anti-ringing filter, both Bicubic 75 and Mitchell-Netravali should perform the same.
Which you use should be based on your preference/tolerance for ringing & aliasing (Bicubic) or softness & desaturation. (Mitchell-Netravali)

I have found very few content that suffers from ringing or aliasing with Bicubic 75, but the desaturation from Mitchell-Netravali shows up with reasonable frequency for me.

[Niyawa]

I have found very few content that suffers from ringing or aliasing with Bicubic 75, but the desaturation from Mitchell-Netravali shows up with reasonable frequency for me.
You personal recommendation would be Bicubic then? Hm, let me do a few more tests and see if I didn't missed anything.

[madshi]

MPC-HC all ver. mouse scrooll and keyboard key's don't work in Exclusive mode (when cursore on exclusive timeline)
Ah, I see. However, on my PC the up & down keys still work fine even if the cursor is on the seekbar. I can't seem to reproduce the problem here.

0.85.2 and 0.85.3 give the same results like 0.85.4, so sorry no confirmation.
Ok, thanks.

I see, thanks. If it's not too much of a hassle, an option to store settings separately (in a file?) would be great.
You could block the registry key where madVR stores its settings by using Windows access rights restrictions. That would effectively stop madVR from using the registry for the settings.

While I love madVR's quality, I think it should have a little more aggressive "trade quality for performance" options.
The "use half frame rate for DXVA processing" is already quite agressive.

Madshi, 85.3 introduced the problem, but 85.4 fixed the problem for some files except for the file I provided a sample for. That file will only play with correct levels with 85.2. And if your wondering, I did test that file with 85.3 also, no luck.
I see, thanks.

Lanczos is a little more hard on my hardware than Spline. Not too much, but it does make a difference.
That must have been random measurement spikes. Technically, all 2-tap algorithms (SoftCubic, Bicubic, Mitchell-Netravali, Catmull-Rom) in madVR use *exactly* the same shader code. Also all 3-tap algorithms (Lanczos3 and Spline3) use exactly the same shader code. The only thing that differs is a lookup table with weights. So performance should be 100% identical between e.g. Lanczos3 and Spline3. Performance depends only on the number of taps, nothing else. Oh, except for Jinc, of course, which is totally different from everything else.

Chroma Upsampling
Ultra-tier: Jinc 3 taps. AR enabled, Linear disabled.
High-tier: Bicubic 75. AR enabled, Linear disabled.
Mid-tier: Bicubic 75. AR disabled, Linear disabled.
Low-tier: Mitchell-Netravali. AR disabled, Linear disabled.
Lowest-tier: Bilinear.
Bicubic and Mitchell-Netravali have exactly the same performance. So the difference between Mid-tier and Low-tier doesn't make much sense to me.

Generally, Bicubic 75 already has quite strong ringing. I think Bicubic 75 is a good chroma upsampling algorithm if you have the AR filter enabled. But without AR it might make sense to ease up a little on the sharpness and use maybe Catmull-Rom instead. So my suggestion would be to use Catmull-Rom for Mid- and Low-tier, as a compromise between ringing, aliasing and desaturation. Not sure if 6233638 will agree with me, though... :p Rest can stay as it is.

Image Upscaling
Ultra-tier: Jinc 3 taps. AR enabled, Linear disabled.
High-tier: Lanczos 3 taps. AR enabled, Linear disabled.
Mid-tier: Lanczos 3 taps. AR enabled, Linear disabled.
Low-tier: Lanczos 3 taps. AR disabled, Linear disabled.
Lowest-tier: Bilinear.
Looks good to me.

Image Downscaling
Ultra-tier: Catmull-Rom. AR enabled, Linear enabled.
High-tier: Catmull-Rom. AR enabled, Linear enabled.
Mid-tier: Catmull-Rom. AR enabled, Linear disabled.
Low-tier: Catmull-Rom. AR disabled, Linear disabled.
Lowest-tier: Bilinear.
Looks good to me, too.

Actually, I'm planning to change madVR's defaults to Chroma Catmull-Rom, Image-Up Lanczos3 and Image-Down Catmull-Rom in the next madVR build. All without AR and without linear light, so that even "low-tier" can run smoothly. Also I'm planning to enable the trade quality for performance option "use half framerate for DXVA deinterlacing" by default. Ouch. You may want to consider enabling that option, too, for low- and lowest-tier, just to make sure even 1080i60 content runs fine.

[nevcairiel]

Half-framerate deinterlacing by default? I mean, sure, a lot of low-end GPUs don't have the performance, but isn't this a bit .. extreme?

[DragonQ]

Also I'm planning to enable the trade quality for performance option "use half framerate for DXVA deinterlacing" by default. Ouch.
Um, what?

[ThurstonX]

MPC-HC all ver. mouse scrooll and keyboard key's don't work in Exclusive mode (when cursore on exclusive timeline)
http://i54.fastpic.ru/big/2012/1221/dc/8f10ff6e79291eee8bd735f11d3970dc.jpeg

I've had this problem at least twice with MPC-HC. I couldn't get the audio delay to work. A reinstall fixed it.

[madshi]

Half-framerate deinterlacing by default? I mean, sure, a lot of low-end GPUs don't have the performance, but isn't this a bit .. extreme?
Um, what?
Yes, it is a bit extreme. I'm not sure if I should really do that, but then, I want new users to have "smooth" (ok, half framerate is only half smooth) playback when they first try madVR. First impression is important. If new users get a slide-show on the first try, they might just give up right away and never come back. And low-end GPUs could very well have trouble doing 1080i60 with madVR's default settings smoothly. E.g. my HTPC (NVidia 9400 mainboard) really struggles with 1080i60 playback, but is totally smooth with half framerate deinterlacing. Furthermore, for movie content half framerate is not any worse than full framerate. There's only a difference for video content, which isn't *that* wide-spread these days, anymore.

Any opinions on this? Am I going crazy by planning to enable the "use half framerate deinterlacing" option by default?

[DragonQ]

Honestly, I kinda think what is the point in MadVR if you get half the temporal resolution? Surely the whole point is high quality playback? I can't see how trading a small improvement in spatial resolution for a massive loss of temporal resolution is an improvement. If someone is using EVR and then tries MadVR and sees that the frame rate is halved, why would they keep using it? Granted some people wouldn't notice and some people would change the settings but we're talking about defaults here, which automatically means we're talking about non-power users.

The only way I can think this is a good idea is if there was an installer that allowed the user to choose from a small selection of presets (with the best preset, determined by detection of the user's GPU, selected by default). In that case, the lowest preset could include this.

[AndreaMG]

Any opinions on this?
I think MadVR has such a good reputation for being highest quality that making it default is like going the opposite direction. I find a good idea the possibility to choose various configurations while installing the program though...

[SamuriHL]

This is why you need an installer. :P So someone can pick what type of hardware they have "Ultra, high, medium, POS LOL" and madVR will adjust the defaults accordingly. Yes, I know, not til 1.0. Then again, I'd argue you shouldn't be worrying about low end machines until then. ;)

[AndreaMG]

It could be even better an installer with automatic hardware detection (if memory serves me right SVP already does that)

[madshi]

Well, my thinking was:

You only get half the temporal resolution with natively interlaced video content, which is rather rare these days. Today most content is really 24p based. The only common exceptions are sports and music concerts. So is it worth it deinterlacing every movie to 60p instead of 30p, although only 24 of those frames are unique? I think most users use madVR for movie watching, so doesn't it make sense to setup the default options so that even low end GPUs can do movie playback smoothly even for 60i content? Again: There's no real loss using half framerate deinterlacing for movie watching.

Furthermore: Some GPUs can't even do 1080i60 playback smoothly with EVR in full framerate. So madVR would have the edge in such a situation, with 1080i60 movie content.

But anyway, it seems nobody likes my idea, so I'll just forget about it.

[DragonQ]

I don't agree that interlaced content is rare. Almost all TV is interlaced. Only films, drama series and some comedy series are typically progressive.

Interlaced content is the main reason I much prefer to record stuff than download it or watch it on a catch-up service: the latter two methods are almost always deinterlaced losing half the temporal resolution.

I also doubt many GPUs fail to playback 1080i/30 content properly in EVR. Maybe Brazos but surely not much else - even the old Intel HD Graphics IGP can do it.

[AndreaMG]

For me smothness is a #1 priority. Actually many people, like me, use MadVR to enjoy concerts and music videos which are mostly 1080i50 or 60. GPUS are also getting stronger year by year and we want them to be used (and abused:)) not only by videogames.:)