madVR - high quality video renderer (GPU assisted)

[sneaker_ger]

In addition to what golagoda said:
Chroma upscaling must be used even for displaying Full HD (1920x1080) on a Full HD screen, because Blu-Ray and HDTV (and DVD) use 4:2:0 chroma subsampling, which means that the chroma resolution is lower than 1920x1080. Only the luma has the full resolution and will not be scaled if video and display resolution are identical.

[Gser]

In addition to what golagoda said:
Chroma upscaling must be used even for displaying Full HD (1920x1080) on a Full HD screen, because Blu-Ray and HDTV (and DVD) use 4:2:0 chroma subsampling, which means that the chroma resolution is lower than 1920x1080. Only the luma has the full resolution and will not be scaled if video and display resolution are identical.

What is the chroma resolution exactly?

[Budtz]

It would be nice if there was some kind of documentation explaining chroma and flushing and all the other settings that is pretty much geberish for the average user. What they do and what settings are optimal depending on the conditions of the user. Pretty big part of this thread is regarding settings. I have not bin able to find a guide. Only parts or personal and undocumented feelings about wich are right.

[TheShadowRunner]

Yes in some cases, for example there are blu-rays with 1440x1080p content and others with 1920x1080p content, so if you viewed the 1440x1080 at full screen 16:9 on a 1920x1080 monitor it would obviously need scaling to 1920x1080

hmm why would you want to stretch 4/3 contents to 16/9 in the first place. ^^;
If you watch this type of content @1080p, keeping correct AR, madVR shouldn't do any kind of scaling. (?)

Edit: k, wasn't aware bluray dealt with anamorphic wide-screen for higher rez (1440×1080) too.

[e-t172]

hmm why would you want to stretch 4/3 contents to 16/9 in the first place. ^^;
If you watch this type of content @1080p, keeping correct AR, madVR shouldn't do any kind of scaling. (?)

Most 1440x1080 content use non-square pixels and is meant to be stretched to 16:9. French HDTV (H.264) is broadcast this way.

[TheShadowRunner]

Most 1440x1080 content use non-square pixels and is meant to be stretched to 16:9. French HDTV (H.264) is broadcast this way.

I see, thanks ;)

[6233638]

6233638
iSunrise

Very goog attempt to visualize various resampling kernels:

http://svn.int64.org/viewvc/int64/resamplehq/doc/kernels.htmlThanks for the link, I think it's been posted before. It's good information, but doesn't mention aliasing at all, which I think is a very important factor for image quality.

This is why I think that ultimately, the best way to determine which scaling algorithm is "best" (or at least, which you prefer) is using real-world examples.

Synthetic tests can only tell you so much.

[Portioli]

Great Post 623398!

BTW, i dont think that we will have any problem for 1080p blurays on 4K displays.



i do not know what sony uses for RealityCreation but
there is a significant difference on Sony VPL-VW1000.

reality creation [off]

http://www.cine4home.de/tests/projektoren/Sony_VPL-VW1000/Sony_VW1000_Teaser_clip_image042.jpg


reality creation [on]

http://www.cine4home.de/tests/projektoren/Sony_VPL-VW1000/Sony_VW1000_Teaser_clip_image044.jpg


i dont think that htpc`s will not find their way.
When 4K will be the new standard
for 1080p--> 4K scaling
we will have the CPu & Gpu Power
to work on-the-fly with non-linear algorithms (e.g. NNEDI3), followed by a good sharpening algorith (e.g. LimitedSharpenFaster)

photos are from: http://www.cine4home.de/

[sneaker_ger]

What is the chroma resolution exactly?

With the common 4:2:0 subsampling there is only a single chroma pixel each four luma pixels. You have to divide both the vertical and horizontal resolution by two:
Full HD:
Luma: 1920x1080
Chroma: 960x540

[mr.duck]

What is the 0 in 4:2:0 ?

luma:chroma:??

[Nevilne]

http://en.wikipedia.org/wiki/Chroma_subsampling#Sampling_systems_and_ratios

[mindz]

Still getting this weird color change effect on .MKV's where if i choose deinterlacing ON i get different colors then if i have deinterlacing OFF! Skins and more reddish plus reds are brighter with deinterlacing on.

This happens only on MKV, DVD, AVI are ok.

Anybody got a hunch :-)?!

[iSunrise]

Still getting this weird color change effect on .MKV's where if i choose deinterlacing ON i get different colors then if i have deinterlacing OFF! Skins and more reddish plus reds are brighter with deinterlacing on.

This happens only on MKV, DVD, AVI are ok.

Anybody got a hunch :-)?!
Please provide some screenshots (you can do them in windowed mode) or a sample of your MKV, otherwise madshi (or other users) can´t verify what you are seeing.

[Andy o]

Try to disable all video processing in your graphics driver control panel. For example, AMD's CCC has a bunch of options in its video section that can mess up color like that.

[Tom Keller]

I'm getting completely random freezes using any version of madVR higher than .65 it seems. I can't reproduce the problem and these freezes happen completely at random. It happens about once every 2-3 films so it isn't a problem that occurs with high frequency. I've tried updating and reverting video drivers (ATI), using different decoders, different splitters, and different media players and I still get freezing with any of the newer madvr versions.
Me too. But these are very short freezes for me (around a couple of 100 milliseconds - at most). They are more kind of short lags. And those freezes/lags are very rare (1-2 times while playing a 45 min episode of a series). I tried different madVR settings, different decoders, different Players and different graphics driver versions without success. Only reverting to the older madVR version 0.65 seems to help.

Strangely i have different graphics hardware (=> NVIDIA 9600GT), running under XP SP3. So i assume: it's not a gpu related problem...

[mindz]

Please provide some screenshots (you can do them in windowed mode) or a sample of your MKV, otherwise madshi (or other users) can´t verify what you are seeing.

Alright, so ive made a pretty clear comparison screenshot:

This only happens with .MKV. DVD (MPEG2) and AVI for instance, dont have this issue.

1. Deinterlacing OFF
http://dl.dropbox.com/u/31835397/Deint%20OFF.png

2. Deinterlacing ON
http://dl.dropbox.com/u/31835397/Deint%20ON.png


Oh and btw, what image is the one with the right colors :)?

[madshi]

No, it doesn't. Desktop composition is always equal to screen refresh rate from before loading the first (or should I say - most recent) file, so if I manually switch to correct refresh rate beforehand, composition matches it of course. After that, it stays the same until the change to different refresh rate.
Ok, thanks, I'll check if I can reproduce that.

This time with full log-file :

Window mode : file plays OK, no dropped frames,
FSE mode : lots of "no vsync slot available"
Hmmmm... I can see that "no free vsync lost available" comes a lot. I've made a small change. Please check with the next madVR version (when it's released). If it still occurs, please send me a new log then. Thanks.

The focus bug (for dual screen PCs) has always been there for me. It doesn't happen a lot, but I can definitely remember it happening w/ every version I've used.
It just happened a minute ago and I captured the screen to show you what happens. Steps that made it happen:
[Firefox minimized at this point]
1) Start playing a mkv file on monitor 1 (windowed mode).
2) Drag the mpc-hc window to monitor 2.
3) Double click the video to go to exclusive mode on monitor 2.
4) Click the Firefox icon on the Taskbar to restore the Firefox window back to maximized mode.
[Focus bug occurs; that's when I took the screenshot]
Do these steps reliably reproduce the problem? Or does it only happen once in a hundred tries or so?

Perhaps useful for madshi to know that it's Windows 8 before reporting a bug?
Yes, that would have been useful... ;)

Here's a log for a crash I'm having with multiple monitors and a DVD-Image source. Occurs when video is playing or paused, when I move MPC-HC to another (any)display(I have 4).
Unfortunately a log usually doesn't help for crashes. You'll have to wait for a future version with better crash analyzation features.

I've also noticed a potential bug in the GPU queues. With 24CPU/8GPU, if the Decoder Queue drops significantly 23-24/24 -> 11-24/24 the Upload and Render Queues have a tendency to drop a bit as well 7-8/8 -> 5-8/8 or lower causing delayed/dropped frames, even though and the GPU isn't stressed and has plenty of data to keep the queues full. [...]

Anything you can do to increase the resiliency of the GPU queues? Theoretically the Upload queue should always be full as long as the Decoder queue is higher, but that isn't happening for some reason.
If the decoder queue is losing ground, the decoder must work harder to fill the queue again, which pushes the CPU to its max. This may eventually also limit the upload speed (which needs some CPU assistance). But I think the bigger picture might be something different: You seem to think that the CPU queue losing ground results in the other queues to go south, too, but maybe the explanation could also be something different: It's possible that the reason why your CPU queue goes south is also the reason why the other queues go south. Maybe your PC is busy with "something" during that time and doesn't really allow doing much else, which would result in all queues losing frames. Of course this is just a thought, and it could be wrong.

I'm thinking that re-tweaking thread priorities so the uploads queue has a higher priority than the decoder queue may help?
I have my doubts about that, but here's a test build:

http://madshi.net/madVRuploadPriority.rar

No, unfortunately I continually drop frames regardless of the queues in .79. There are so many permutations (each option increases the test case exponentially) that I have spent hours on this, but I have tried running in seperate threads, multiple flush options, different scaling algorithms, feeding different signals from different filters, etc...Since your post I focused on having 8 frames in advance (or backbuffers with that option disabled) and no matter what I try I simply drop frames on a second-by-second basis (where .78 allows me to run SoftCubic/4-Tap Spline in fullscreen exclusive without any frame-drop.)
Well, I asked you something very specific, I didn't ask you to test all permutations. Basically, restore your v0.78 options, then set both CPU and GPU queues to 8. Leave all other settings the same. If that doesn't help then I don't know what will. Most other people seem to have no slowdown with v0.79 compared to v0.78, so I don't really know why you do. It's very weird...

The render and present queues for me are a bit of an issue, I have mine starting off pretty high then after 10 seconds or so it drops right down. I'm nearly always running around 3-4/12 or less on my 550 Ti i7 system regardless of the decoder, player, CPU availability and GPU clock settings (I've disabled PowerMizer with no effect.) Anyone else have a particularly low render and present queues running an Nvidia card? I'm hoping 0.80 helps a little with the odd frame drop I'm getting, guess we'll see!
Which queue starts dropping first? Which queues exactly stay full, and which go down? I need exact information. Make sure you have antialiasing and anisotropic filtering disabled. Also try using less demanding scaling algorithms. But really, your 550 should be able to handle everything you throw at it just fine.

And is there any point in maxing out the CPU queue size? any ill effect such as increased CPU load?
Just try what works best for you. CPU load increases only during the queue fill time. Once the queue is full, CPU load should not be any higher than with a smaller queue. One "ill effect" of having a large CPU queue is higher RAM consumption.

Don't know if it's new. madVR recovers when switching between multiple desktops, but not when enabling/disabling desktops. The whole screen flickers when you do this, so I assume dwm.exe or perhaps the video driver is being restarted. Resuming from sleep causes the same issue.
Ok, thanks, I've put this on my to do list, but it's not a very high priority at the moment.

I have problem with madvr and rgb32 output with ffdshow
image is cut in tow parts (up there is image, down there is black screen)
http://img15.hostingpics.net/thumbs/mini_482396Sanstitre.jpg (http://www.hostingpics.net/viewer.php?id=482396Sanstitre.jpg)

with nv12 ou yv12 output in ffdshow ther is no problem
http://img15.hostingpics.net/thumbs/mini_197490Sanstitre.jpg (http://www.hostingpics.net/viewer.php?id=197490Sanstitre.jpg)

I have madvr0.74 but even problem with all version of madvr.

Same problem with all version of ffdshow (I use the last version but I try with many versions)

Edit: It works with RGB24. The problem occurs with only RGB32
Please retry with the next madVR version (when it's released), I've changed something which may help.

P.S. If you block OSD message any time you encounter word "Play" or "Pause" in OSD text, here translation of these words into Russian: "Play"="Воспроизведение", "Pause"="Пауза".
I've added that, but I'm not sure if it will work because I don't know if MSVC++ supports unicode strings in the editor. Just give it a try with the next version when it's released.

I think that under known problems/limitations you should make it known that you cannot screencap yet.
Ok, added it to the known limitations.

I found something in the deinterlacing of madVR that i do not understand. Whenever i enable deinterlacing on a progressive .mkv, the color changes. Reds are less red, and colors overall look better. How is this possible?
I've found a sample which shows the problem on my PC, I'll try to fix this for the next build.

Before posting the log I did a clean install of the latest Nvidia-Drivers 285.58. After reading your suggestion I once again uninstalled and reinstalled the Nvidia-Drivers.
But nothing changed. Sometimes madVR is working, sometimes I have a black screen.

My GPU is a Nvidia ION (first generation) and I tried running on 24Hz and 60Hz.
That's pretty sad. I think my sister has an ION laptop, maybe she let's me borrow it for a while.

I am using LAV filters + madVR + ReClock. I get 0 dropped frames, 0 delayed frames and no presentation glitches. madVR OSD reports 1 frame repeat/drop every 1.05 days or so. Even though the stats look so perfect, I feel the playback is not very smooth, especially when I play 23.976 fps and 25 fps videos. I have two displays: a monitor and an HDTV. Both have a native refresh rate of 60 Hz. I have forced a custom resolution of 72 Hz on my monitor and 48 Hz on my TV, for 23.976 fps videos and also for 25 fps videos (by speeding them down using ReClock). I don't understand why I am not getting supersmooth playback.
First of all disable Reclock's VSync correction, then reset the VSync timer database. Maybe that already fixes the issue?

Can anyone point me to the dynamic range setting (that is being referred) in CCC? I can't find it anywhere.
Not all GPUs/drivers have it. I had it with a HD6850, but not with a HD4770.

Just curious, I know the Video settings in ATI's CCC have an effect when madVR is deinterlacing, but is this still the case if madVR doesn't receive the interlace flag from the decoder?
The Video settings only have an effect on madVR if DXVA deinterlacing is activated. How DXVA deinterlacing was activated (through decoder interlaced flags or manually) doesn't matter.

[madshi]

Maybe somebody can confirm if madVR gamma processing ( Im using the 2.0 setting, not 2.2 ), does process the whole image or only the picture without the the black belts in 2.35 aspect ratio movies?
Gamma processing is performed before scaling. All black bars which are part of the encoded video image do get gamma processed. But all black bars which are a consequence of image scaling are not gamma processed.

It would be nice if in a later version the queues could be set lower in hopes that HD plays on an X300M with MadVR.
What is an X300M? How much dedicated GPU RAM does it have? Which GPU queue size does madVR choose? And which is the max that works well for you?

Also more lanczos taps options would be nice, like 6 and would there be any advantage to more than 8 taps?
IMHO 4 taps is the max I would use. Using more just adds too much ringing, without any real benefit. I've only added 8 taps on request, but I don't recommend using it.

I was just testing sending an RGB32 signal from ffdShow (instead of YV12 or NV12) and for some reason only the top half of the screen is rendered. What could be causing this?
Please try again with the next build, hopefully fixed.

They are hard to capture, they only appear when paning. It's like the picture splits in the top, I've seen it before but don't remember how to remove it.
Sounds like tearing to me. This is what exclusive mode is made for. Try going fullscreen, so madVR can enter fullscreen exclusive mode. Does that make the problem go away?

I would like to suggest a ( small ) new feature concerning the auto refresh rate switching
The display mode changer is nothing but a hack in its current form. A new version will come sooner or later with better features and usability. For now you'll have to live with the way things are, until my priority list tells me that it's time to redesign the display mode changer. Too many other things to do at the moment, which are more important.

With the release of madVR 0.80 or soon after, can we get an updated version of madVR which supports benching the internal decoders and possibly now dxva deinterlacing as well? It's now been around 6 months since the initial version of madVR [benchmark] was released, so it's about time to get a new version which supports all the new madVR functionality since then.
Yes yes, but I've sooooo many other things to do, as well. Updating the benchmark will take some development time, so I can't promise it will be very soon. It is on my to do list, but so many other things are, as well.

Hi ... If I set MadVr to change resolution for example. 1080p24 and leaves the exclusive mode show CTRL+J 60Hz. My monitor switches to 24Hz but MadVr shows 60Hz in windowed mode and movie stutters. Is this normal? and yes, the movie is 23,976 :-)
Looks like a problem with the Aero composition rate. This has been reported before. I'm not sure if there's anything I can do to fix it, but I'll look into it. The best solution for now probably is to use exclusive mode.

[B]madshi
Hi. Can you look why this sample (http://www.mediafire.com/?7do9qzdvbuae34l) crash MadVR but works correctly with Haali's Video Renderer
Thanks, I can reproduce the crash, will fix it in the next build.

MadVR does not change to the right displaymode on 30fps material. I tried watching anime that have 24 or 23.976 fps in 72Hz switch and MadVR switched right. But on 30 fps it does not switch to the 120 Hz I set, it switches to 100 Hz that I made for 25 or 50 fps videos. Is that an error?
Does that 120Hz display mode really exist? madVR can only switch to modes that exist. Can you switch to 120Hz in the GPU control panel? If so, please send me a log which shows the switch to 100Hz when it should have switched to 120Hz. Thanks.

Edit 2: I just allowed MadVR to switch to 1080p72 only, nothing else. But it now switches to 72 Hz on every video. May it be 24 fps (for what it was intended to) 30 fps, 50 fps or 60 fps. I tried adding 1080p120 after 1080p72 but it didn't work, it still just switched to 72 Hz. If I set 1080p120 first and then 1080p72, it just stays at 120 Hz and never ever switches. Please fix this problem :/ I just now realized how bugged this is.
Please send me a log for the "1080p120 first and then 1080p72" situation, too. My display doesn't support 120Hz, so I can't test this here.

I have setup consisting of a thinkpad T500 with a panasonic plasma TV hooked up as secondary display (dvi->hdmi), When a movie starts up fullscreen on the tv and this is switched to exclusive mode, it just displays the first frame and then you hear the audio running. If i then right-click on mpc-hc it brings up the main menu, and the video displays and sort of 'catches up'..
Also if i move the cursor to the thinkpad screen (primary) and activate another window the video will freeze again on the tv. This has been so with the versions from the last month or so. makes it a little hard to use the pc for other things while something is playing...
Can you please try to locate with which version exactly this problem started? A log file would also be helpful (please try to keep it as short as possible).

The 0.79 version of madvr caused massive dropped frames untill i increased the CPU and GPU queue sizes manually??
How were they set by madVR? And which value did you increase them to?

I'm getting completely random freezes using any version of madVR higher than .65 it seems. I can't reproduce the problem and these freezes happen completely at random. It happens about once every 2-3 films so it isn't a problem that occurs with high frequency.
This could be some kind of hardware problem. Maybe your GPU or CPU or mainboard are getting too hot?

select madvr output i got this http://i.imgur.com/FYpM2.jpg
select EVR custom pres i got http://i.imgur.com/L4MIk.png

is it just me or the EVR custome pres gives more smoother, less blocky, better image?
You can't compare by using different frames! The only way to do a fair comparison is to use *exactly* the same frame. With Anime content it's even harder: The same frame is encoded multiple times. You gotta make sure you even get the same encoded frame, because otherwise you might compare an I-frame (usually higher bitrate) to a B-frame.

DXVA has some algorithms for removal of noise and compression artifacts. They may help with *really* bad sources. In that situation EVR/VMR may look "better" than madVR because madVR doesn't do any such thing. For reasonably good sources applying any kind of noise/artifact removal usually reduces visible details.

Similarly, I do not recommend SoftCubic at all for chroma. I do not understand why so many people use it for chroma—perhaps because it's one of the defaults? SoftCubic blurs the image too much, bleeding colour out of the edges and desaturating things.

I have done several rounds of testing on the scaling algorithms, and always end up back at Bicubic 75 for chroma. It seems to be the best balance between ringing, and preserving the original image from the source. (i.e. not desaturating it, or blurring things too much)

For upscaling, I have consistently settled on SoftCubic70 without linear light scaling, though this seems to be entirely based on personal preference.

It's not that I am against having a sharp image, but any other setting (other than higher levels of SoftCubic) results in intrusive aliasing when playing back DVDs, which breaks the immersion for me.
It's kinda funny. You choose SoftCubic for Luma upscaling, saying that you like it best for it's lack of ringing and aliasing. But yet you ignore the same reasons when talking about Chroma upscaling. Every argument you use against SoftCubic for Chroma upsampling also applies to Luma upscaling. And every argument you use for SoftCubic for Luma upscaling also applies to Chroma upsampling.

You can look at frequency burst test patterns until your nose turns red, but all it proves is that resamplers with negative lobes fare better at reproducing frequency bursts. Real life content only *very* rarely has alternating lines with totally different content. That's especially true for Chroma.

Does anyone else miss bicubic 1.0 from EVR?

It's the only advantage I see of evr over madvr currently. It's still my preferred resizer of clean hd video because of great detail with little ringing and below average aliasing. With madvr the ringing and aliasing would probably be reduced based on comparisons of the other resizers shared in both renderers.

I primarily watch dvd's but for me evr with any of the resizers is not good enough for upscaling sd to hd, too much ringing and aliasing, while madvr does an excellent job of minimizing the artifacts albeit less detail.
Can you make screenshots that compare madVR's sharpest Bicubic to madVR's Lanczos3 and to EVR's Bicubic 1.0? If EVR's Bicubic 1.0 really has merit, I'll consider adding it. But I want to see some screenshot proof first.

Me too. But these are very short freezes for me (around a couple of 100 milliseconds - at most). They are more kind of short lags. And those freezes/lags are very rare (1-2 times while playing a 45 min episode of a series). I tried different madVR settings, different decoders, different Players and different graphics driver versions without success. Only reverting to the older madVR version 0.65 seems to help.
Which queues drop (and which don't) exactly when these freezes happen?

[Damien147]

With madvr in exclusive mode you just can't minimize/maximize/close on the second monitor.

[mindz]

Another weird one i had encountered earlier on aswell. This time the problem is with DVD MPEG2 deinterlacing. I made two screenshots, one where deinterlacing is ON, and the other deinterlacing OFF. Now with the deinterlacing ON the MPEG2 content is actually not being deinterlaced.

Screenshots:

Deint ON:

You can see the jaggies everywhere, the subtitles and for example on the teeth of the surgeon, but a trained eye can see it everywhere i guess.

http://dl.dropbox.com/u/31835397/Deint%20ON%20MPEG.png

Deint OFF:

This seems about right, right :)?

http://dl.dropbox.com/u/31835397/Deint%20OFF%20MPEG.png

[cyberbeing]

if the decoder queue is losing ground, the decoder must work harder to fill the queue again, which pushes the cpu to its max. This may eventually also limit the upload speed (which needs some cpu assistance).
This is exactly what I suspect the problem is, since the problem only seems to occur in worse-case scenarios when my CPU load briefly gets maxed out to 96-100%, and results in extremely slow decoding speed for up to a half-second or so. When a high CPU queue drops >8 frames or so, the other queues begin to drop as well.

It's possible that the reason why your cpu queue goes south is also the reason why the other queues go south. Maybe your pc is busy with "something" during that time and doesn't really allow doing much else, which would result in all queues losing frames. Of course this is just a thought, and it could be wrong.

Doubtful considering this WinXP SP3 installation never has anything running in the background, and I'm not seeing any DPC latency spikes above ~100µs during playback. Remember that I'm only seeing the other queues begin to drop once the madVR decode queue drops very significantly. The something this PC is busy doing is decoding video which my CPU (AMD X2 dual-core) can barely handle. As long as my GPU load is within a nominal range, the Upload and Render queues don't budge from 100% full, except when this happens. The question would appear to be, 'What is getting CPU starved?'. madVR? GPU driver? Windows Kernel?

I still wouldn't dismiss the possibility that the problem is triggered only by the madVR Decode queue dropping and has nothing to do with the CPU getting starved, but since the two are interrelated, I'm unsure how to test that in a controlled environment.


I have my doubts about that, but here's a test build
I'll test that build later today.

[FlashGordon]

This could be some kind of hardware problem. Maybe your GPU or CPU or mainboard are getting too hot?

That's the first thought I had, but I have pretty good cooling. It just seems odd to me that this only happens with the newer madvr versions.

Is there anything I can do to single out the issue when these freezes happen?

[HitomiKun]

I may have discovered why madVR switches to the wrong Hz rate. Something involving ffdshow, but Im not sure about that right now, Im still getting these wrong Hz rates. Yeah and I also would like to know how to save logfiles in mpc-hc. I searched the interface and also searched on google about saving logfiles, but without success. Anyone tell me please?

Edit: Oh I guess I have to build an debug version of mpc-hc? Well I'm on it right now, not sure if that's the solution.

Edit 2: Nope, didn't help. Does not change anything in the interface. :/

Btw.
Deint ON:

You can see the jaggies everywhere, the subtitles and for example on the teeth of the surgeon, but a trained eye can see it everywhere i guess.

https://dl-web.dropbox.com/get/Photos/Deint%20ON%20MPEG.png?w=f4c0bdd8

Deint OFF:

This seems about right, right :)?

https://dl-web.dropbox.com/get/Photos/Deint%20OFF%20MPEG.png?w=3d9e6bb0

@mindz -> You have to move your pictures into your Public folder, rightclick the pictures and select "copy public link" under the DropBox menu. The links you posted here are private links that ONLY you can see. (If you deleted your Public folder as you started using DropBox (like I did when I was new to DropBox) then you can restore it on your DropBox page. Or you just create a new folder named "Public" in your DropBox folder.)

[Andy o]

Not a MPC-HC log, but a madVR log. Go to your madVR folder, and there should be a file named "activate debug mode.bat" click on that, and next time you play a file with madVR it will save a log on your desktop. It can get quite big, so make sure to just play enough time for your issue to show. Also a very good idea to zip it before you upload it.

What TV/display are you using, that can do 72, 100 and 120Hz, by the way?

[HitomiKun]

What TV/display are you using, that can do 72, 100 and 120Hz, by the way?
I bought this Monitor 2 weeks ago after searching for months. I use the Asus VG236H. It supports all Hz rates from 50~122 Hz when it's connected via Dual-Link DVI. It's actually a 3D Monitor but I bought the H version that does not include NVIDIA 3D glasses because I needed a monitor that isn't just 60 Hz. (HE version includes 3d glasses so it's more expensive)
Specs for my monitor here: Asus VG236H (http://www.asus.com/Display/LCD_Monitors/VG236H/)
A good review here: Asus VG236 Review (http://www.digitalversus.com/lcd-monitor/asus-vg236h-p8757/test.html)

Thanks for your help, I'm going to get these logs now.

[cyberbeing]

I'll test that build later today.
Well the good news is it does seem help stop the Upload Queue from dropping until the Decode Queue drops below the max value. The bad news is that while it helps, it doesn't completely eliminate the delayed frames usually produced in these sections. Now it seems the Render queue may be getting slightly starved, since it sometimes drops one frame below the Upload Queue in that build. In both builds there are oddly delayed frames @120Hz when the Decode Queue drops slightly (high CPU load), but the Upload/Render/Present queues remain apparently full.

Could you produce another couple builds?

One with something like:
VSync | Time Critical
Render | Highest
Upload | Above Normal
Decode | Below Normal

Another with something like:
VSync | Time Critical
Render | Time Critical
Upload | Highest
Decode | Lowest

[HitomiKun]

Alright after some time testing around I found the issue behind all of these wrong Hz rates. I have ffdshow raw video filter enabled for 8 bit videos to use deband filters. I recently tested stuff with 10 bit videos and I disabled ffdshow raw video filter for those videos because it makes my 10 bit videos lag. MadVR then chooses the right display modes on any FPS. But when I watch 8 bit videos with ffdshow raw video filter enabled, Hz rate selection gets messy.

So here is my NVIDIA Control Center:
http://img141.imageshack.us/img141/9140/available20refresh20rat.png (http://dl.dropbox.com/u/48947610/MadVR%20Bug%20Report/Available%20Refresh%20Rates.png)
The custom resolution on top is the one that allows me to use 72 Hz.

3 Log files: http://dl.dropbox.com/u/48947610/MadVR%20Bug%20Report/madVR%20logs.zip

I tested:
- 1080p72, 1080p100, 1080p120 with my 30 fps video (download link below) -> It switched to 100 Hz
- 1080p72, 1080p120 with my 30 fps video -> It switched to 72 Hz
- 1080p120, 1080p72 with an 23.976 fps Anime (no download link because the file I tested it on had 190 mb) -> Nothing happened.

Tested video: 30 fps sample (http://dl.dropbox.com/u/48947610/MadVR%20Bug%20Report/Intro.mkv) (11.5 mb)

My ffdshow raw video filter settings:
http://img210.imageshack.us/img210/1142/ffdshow20raw20video20fi.png (http://dl.dropbox.com/u/48947610/MadVR%20Bug%20Report/ffdshow%20raw%20video%20filter%201.png) http://img535.imageshack.us/img535/1142/ffdshow20raw20video20fi.png (http://dl.dropbox.com/u/48947610/MadVR%20Bug%20Report/ffdshow%20raw%20video%20filter%202.png) http://img442.imageshack.us/img442/1142/ffdshow20raw20video20fi.png (http://dl.dropbox.com/u/48947610/MadVR%20Bug%20Report/ffdshow%20raw%20video%20filter%203.png)
The rest is on default.

So for now I'd suggest blocking/disabling ffdshow raw video filter would fix these wrong Hz rate issues. I still need ffdshow raw video filter for 8 bit videos so it would be great if this gets fixed.

[6233638]

It's kinda funny. You choose SoftCubic for Luma upscaling, saying that you like it best for it's lack of ringing and aliasing. But yet you ignore the same reasons when talking about Chroma upscaling. Every argument you use against SoftCubic for Chroma upsampling also applies to Luma upscaling. And every argument you use for SoftCubic for Luma upscaling also applies to Chroma upsampling.My problem with using SoftCubic for chroma is that it causes colour bleeding and desaturation at any level. This has got me wanting to do another test on chroma scaling though. Even SoftCubic 50 is still softer than Bicubic 75 which I am currently using, so I don't expect it to end up getting recommended.

The difference seems to be that chroma is being scaled up considerably more than luma is, and so you need different algorithms to have a good match with chroma and luma information. (to minimise colour bleeding and/or desaturation)

I'm definitely interested in having another look at the real-world differences are between Bicubic 75 and Lanczos 3 with Chroma. With Luma, it definitely seems like Lanczos is the better choice from my recent testing. (similar ringing, but better aliasing and sharpness)

You can look at frequency burst test patterns until your nose turns red, but all it proves is that resamplers with negative lobes fare better at reproducing frequency bursts. Real life content only *very* rarely has alternating lines with totally different content. That's especially true for Chroma.I agree, this is why the majority of my testing is with real-world content. They can still be useful at times though.


Update: I have done some preliminary testing. My conclusion is still the same on SoftCubic: not at all suitable for chroma scaling.

Lanczos 3 has too much ringing to be used with chroma, despite its good aliasing properties. (not evident in the comparison below)

Surprisingly, while I don't like it for luma at all, Spline 3 is a possibility, but my recommendation is still Bicubic 75 for now, until I have been able to carry out further testing.

Can you guess which of these are SoftCubic 50/100? (highlight below to see what each image is)

http://img402.imageshack.us/img402/2581/bc75.png
http://img832.imageshack.us/img832/3937/98152063.png
http://img805.imageshack.us/img805/7985/26771010.png
http://img507.imageshack.us/img507/1082/sc50.png
http://img51.imageshack.us/img51/3977/sc100.png

Bicubic 75, Lanczos 3, Spline 3, SoftCubic 50, SoftCubic 100


Unlike luma scaling algorithms where it is the viewer's preference for a smoother/sharper image, I don't believe that is true with chroma at all. There are algorithms which are simply not suitable for chroma scaling in my opinion.

[cyberbeing]

Catmull-Rom is the lowest you can go for chroma sharpness without introducing desaturation on highly detailed sources such as CrowdRun. It doesn't ring, it's reasonably sharp, and it's arguably well suited for 2x chroma scaling as the most numerically accurate bicubic filter. If I wanted softer chroma I'd go with Mitchell-Netravali at the expense of slight desaturation. If I wanted sharper chroma I'd go with Spline36 at the expense of ringing.

For Luma my preference has always been Spline36. Lanczos has too much ringing. Bicubic -0.75 isn't sharp enough. The only time I ever notice aliasing when using Spline36, is if the source itself had aliasing.


For day to day use:
Spline36 | Luma
Catmull-Rom | Chroma

Highly detailed sources with bothersome chroma ringing/blocking:
Spline36 | Luma
Mitchell-Netravali | Chroma

Eye cancer sources w/ ringing, blocking, haloing, and/or aliasing:
SoftCubic50 | Luma
Mitchell-Netravali | Chroma

Unlike luma scaling algorithms where it is the viewer's preference for a smoother/sharper image, I don't believe that is true with chroma at all.
There are algorithms which are simply not suitable for chroma scaling in my opinion.
While I agree with you that SoftCubic is unsuitable for chroma scaling, that is nothing more than my personal preference since I highly value chroma detail.

Madshi's personal preference is soft chroma with zero aliasing. SoftCubic which completely obliterates chroma aliasing fits his preference perfectly. Losing detail is a necessary evil.

For a vaguely similar yet different reason, I find SoftCubic unsuitable for luma... unless I wanted to destroy source flaws and eliminate aliasing. Losing detail is — wait, what detail? Reducing eye cancer is for the greater good.

You preferring SoftCubic for luma is just as bad as madshi preferring SoftCubic for chroma from my point of view. You both use SoftCubic to accomplish similar things, except you favor chroma detail/sharpness and madshi favors luma detail/sharpness. Subjective views of others cannot be judged as right or wrong.

[mindz]

@mindz -> You have to move your pictures into your Public folder, rightclick the pictures and select "copy public link" under the DropBox menu. The links you posted here are private links that ONLY you can see. (If you deleted your Public folder as you started using DropBox (like I did when I was new to DropBox) then you can restore it on your DropBox page. Or you just create a new folder named "Public" in your DropBox folder.)

Thanks! Ive changed it, now it should work!

[turbojet]

madshi: It's actually an ATI X2300M ati link (http://www.amd.com/us/products/notebook/graphics/ati-mobility-hd-2000/x2300/Pages/x2300-specs.aspx) with 128 MB of RAM. Checking the auto queue is very difficult because after switching to auto the madvr osd shows for maybe half of a second until all goes black and computer freezes but 14 cpu and 8 gpu is what I think it reads. Is there any other way to check it?

According to madvr with SD 40-50 MB is used. with 720p its 100-110 but with my desktop gpuz and process explorer show much more ram usage but neither of those programs can read the gpu load/ram usage of the X2300M which makes troubleshooting a bit more difficult. But even with nearest neighbor about half the frames are dropped on 720p with 4/4 queues. Madvr has the only vsync that eliminates tearing on that laptop, mpc-hc makes little difference and reclock is effective but tears occasionally when the camera pans.

That's okay about more than 8 taps of lanczos I probably wouldn't be able to use it anyhow. Lanczos8 is still my preferred resizer for upscaling. When scaling SD there's always a bit of ringing in certain scenes and the added ringing from lanczos8 doesn't bother me nearly as much as the aliasing from other resizers. I'm warming to bicubic75 though mainly because it's still pretty sharp and is light enough to display 720p/1080p without dropping frames with an nvidia 9500GT.

Multiupload link (http://www.multiupload.com/WPGVH2464Y) for the madvr and evr pics. Most of the sharpness I'm seeing in evr is from sharpen complex, identified as -ps in the filenames, down means 1920x1080 BD downscaled to 1680x1050. up means 720x480 DVD upscaled to 1680x1050. I still notice a little more detail with evr-bicubic100 when sharpen complex is disabled. Is there anything that can be done to make pixel shaders like sharpen complex more effective in madvr?

[madshi]

I'm not a programmer and I really don't know how the whole thing works but that was easy to fix so I made patch (http://dl.dropbox.com/u/16282309/MPC-HC/madVRAllocatorPresenter.cpp.patch) for MPC-HC and now everything works good :)
Thanks, looks good, I've committed the change to MPC-HC SVN (rev 3894).

[disto]

Hi.Is it possible to add gamma correction slider in madvr ?

[zmaster]

madshi,
You're probably already tired of this question, but... Do not planned to make madVR x64? :)

p/s: Thanks for the madVR!

[vivan]

zmaster, what for?
http://forum.doom9.org/showthread.php?p=1486240#post1486240

[6233638]

Catmull-Rom is the lowest you can go for chroma sharpness without introducing desaturation on highly detailed sources such as CrowdRun. It doesn't ring, it's reasonably sharp, and it's arguably well suited for 2x chroma scaling as the most numerically accurate bicubic filter.I'll have to test some more, but Catmull-Rom has serious aliasing issues with luma which makes me wary of using it for chroma. (I really need a good test for chroma aliasing)

If I wanted softer chroma I'd go with Mitchell-Netravali at the expense of slight desaturation.I wouldn't call it slight.

Nearest Neighbour, Bicubic 75, Mitchell-Netravali:
http://i.imgur.com/99Kwf.gif

Perceptually, Bicubic keeps the same saturation/luminance as Nearest Neighbour, whereas Mitchell-Netravali has a noticeable drop.

The same thing happens with Catmull-Rom to a lesser degree.

Bicubic 75/Lanczos 3/Spline 3, all produce similar results to each other as far as brightness/saturation are concerned, the main difference there seems to be with ringing. (which doesn't manifest itself the same way that it does with luma) I'm now leaning towards Spline 3 rather than Bicubic 75 for chroma, unless I find a source where it does bad things.

I wish Photoshop could export animated PNGs.

If I wanted sharper chroma I'd go with Spline36 at the expense of ringing.So far, ringing does not seem to have been an issue with chroma from my testing, at least not compared to Bicubic 75 which has been my preference for some time now

While I agree with you that SoftCubic is unsuitable for chroma scaling, that is nothing more than my personal preference since I highly value chroma detail.

Madshi's personal preference is soft chroma with zero aliasing. SoftCubic which completely obliterates chroma aliasing fits his preference perfectly. Losing detail is a necessary evil.Personally I find the desaturation from "bad" chroma algorithms to be a fairly significant issue, and I have yet to run into chroma aliasing problems with Bicubic 75. (or spline so far) Does anyone have some good chroma test material? It's always good to have more test cases.

For a vaguely similar yet different reason, I find SoftCubic unsuitable for luma... unless I wanted to destroy source flaws and eliminate aliasing. Losing detail is — wait, what detail?Can you please show me some examples of SoftCubic upscaling losing detail? I will agree that it produces a soft image and can reduce contrast somewhat (at least at the higher levels) but I've yet to find an image where detail has been lost. I should point out that I am primarily upscaling DVDs, not 720p content. Other than game footage, where are people getting 720p material? All Blu-rays I have are 1080i/p.

Eye cancer sources w/ ringing, blocking, haloing, and/or aliasing
...
Reducing eye cancer is for the greater good.Perhaps it's just hitting a bit too close to home, but I find these comments to be in very poor taste.

[madshi]

I have a problem with madVR 0.79 that doesn't happen with 0.78.
As you probably know I use MPC-HC fullscreen refresh rate changer and it has always worked.

But with 0.79, when I play a 24fps movie my LCD starts as 60Hz, then I go fullscreen and the LCD changes to 24Hz but it shows the MPC seek bar kinda out of place, I hear the audio of the movie but no image. I pause and unpause, then I can see the movie, but 2 to 3 seconds later after the movie is already playing, the LCD changes back to 60Hz even though I didn't exit the fullscreen. It should stay at 24Hz.

I go back to 0.78 and all is fine as it has always been.
Should be fixed in the next build.

[madshi]

DXVA deinterlacing bug: When switching displays (applies especially to DVDs) MPC-HC crashes on account of DXVA. I think this is linked to madVR's DXVA auto deinterlacement detection.
This will be fixed in the next build.

[zmaster]

zmaster, what for?
http://forum.doom9.org/showthread.php?p=1486240#post1486240vivan,
Hola!
Personally, idk than madVR x64 could be better than madVR. But madVRx64 can work with My Pleer x64 -> fast decoders + AVS x64 input -> fast video processing filters.

[madshi]

madVR reports:
-creating Direct3D device failed (80004001)

This happen every time MPC is minimized and it need to play the next video

Start a video->minimize the player-> jump to next video by "playlist" or "play next in folder"-> black screen with madVR reports: -creating Direct3D device failed (80004001).
This will also be fixed in the next build.

[cyberbeing]

I wouldn't call it slight.

Nearest Neighbour, Bicubic 75, Mitchell-Netravali:
http://i.imgur.com/99Kwf.gif

Perceptually, Bicubic keeps the same saturation/luminance as Nearest Neighbour, whereas Mitchell-Netravali has a noticeable drop.

The same thing happens with Catmull-Rom to a lesser degree.

That gif pretty much confirms my own tests with CrowdRun. To put my statement in context, Mitchell-Netravali saturation and blurring are slight compared to SoftCubic100, and is the only Chroma scaler I would consider using which has noticeable desaturation. I basically use it on low-quality sources as a poor man's chroma denoise, since anything softer than Mitchell-Netravali I find blurs way to much.

As for Catmull-Rom which I use 90% of the time, the difference between it and Spline36 chroma are basically unnoticeable on real footage. That fact that Catmull-Rom is the sharpest resizer which doesn't produce halos/rings is a bigger benefit since otherwise the halos produced by Spline36 for Luma would increase. Use of Spline36 Luma also slightly restores saturation and reduces the chroma aliasing of Catmull-Rom. The two complement each other well, if aiming for the sharpest result with least ringing as I am.

Bicubic 75/Lanczos 3/Spline 3, all produce similar results to each other as far as brightness/saturation are concerned, the main difference there seems to be with ringing. (which doesn't manifest itself the same way that it does with luma) I'm now leaning towards Spline 3 rather than Bicubic 75 for chroma, unless I find a source where it does bad things. So far, ringing does not seem to have been an issue with chroma from my testing, at least not compared to Bicubic 75 which has been my preference for some time now.
As I mentioned above, chroma ringing/halos only becomes a concern when you are using a luma resizer which produces moderate ringing/halo itself. Since you use SoftCubic for luma, you can get away with pin-sharp chroma at the expense of ringing, since SoftCubic will not increase and actually partially hide the ringing after resize. I wouldn't be surprised if Spline36 chroma complemented it well.

Can you please show me some examples of SoftCubic upscaling losing detail? I will agree that it produces a soft image and can reduce contrast somewhat (at least at the higher levels) but I've yet to find an image where detail has been lost. I should point out that I am primarily upscaling DVDs, not 720p content.
The main issues are indeed a blurry image and extreme loss of contrast which I can't stand. When an image is blurred so much that you are unable to easily see fine detail, I consider that as lost detail. Even though the detail may technically still be there if you sharpened the image back up on your TV afterwards, you shouldn't really be relying on TV post-processing to fix your image.

Other than game footage, where are people getting 720p material? All Blu-rays I have are 1080i/p.
TV recordings, Blu-ray's re-encoded to 720p, and even things like iTunes downloads. It's rare for even 1080p video of Blu-rays to have much resolution above 720p. TV broadcast at 1080i can be even worse.

Perhaps it's just hitting a bit too close to home, but I find these comments to be in very poor taste.
No offense was intended, but I feel it made it made my personal option on use of SoftCubic for luma very clear. If I use SoftCubic luma on a high quality source material it becomes subjectively unwatchable to my eyes. Yet if I have a horrible low quality source which is near unwatchable, SoftCubic can actually help by hiding some of the issue and make the video somewhat watchable.

[pirlouy]

Personally, idk than madVR x64 could be better than madVR. But madVRx64 can work with My Pleer x64 -> fast decoders + AVS x64 input -> fast video processing filters.
Your message in incomprehensible (at least for me). Please write it with basics words, make sentences, and no abbreviations please...

What I mean, x64 madVR is really not a priority right now (from what Madshi said), and it's not your mysterious message which will change its priority.

[iSunrise]

Can you guess which of these are SoftCubic 50/100? (highlight below to see what each image is)
SoftCubic100 should be the last picture and SoftCubic50 the one before that. It´s pretty easy to guess these since SoftCubic100 is heavily desaturating and bluring things, especially when used on luma. Also, contrast seems to suffer because of that. This especially applies to sources with DVD or lower resolution.

Bicubic75 easily wins all of the tests I did, be it chroma or luma. Hopefully madshi adds Bicubic100, I would really like to do some tests on that.

Since madshi also seems to like SoftCubic on chroma (like I do) an algorithm that would do SoftCubic(70-100) and also color-correct (counter the desaturation) would improve SoftCubic even more. Not sure of that is easily doable, though.

[zmaster]

x64 madVR is really not a priority right now (from what Madshi said)It he said in March. Now December.
I have information (of the other forums), that madshi is will not make x64-version, until not improve x86-version.
That is why I asked a question. If the question had been raised a couple of days/hours ago - sorry, that I do not know about it.

[SoBizarre]

No, it doesn't. Desktop composition is always equal to screen refresh rate from before loading the first (or should I say - most recent) file, so if I manually switch to correct refresh rate beforehand, composition matches it of course. After that, it stays the same until the change to different refresh rate.

Ok, thanks, I'll check if I can reproduce that.


One more note. If I was to change to correct refresh rate beforehand - manually, I'd have to do it by clicking through Windows' or graphic card's GUI. Using any utility like NirCdm will cause wrong composition rate. The same goes for ReClock's method of automatic refresh rate changing.

[mr.duck]

64 bit madVR is a complete waste of time and effort for no benefit.

Unless I am missing something, why use 64 bit? Like pirlouy said, you are not exactly being very clear.

[6233638]

That fact that Catmull-Rom is the sharpest resizer which doesn't produce halos/rings is a bigger benefit since otherwise the halos produced by Spline36 for Luma would increase.Catmull-Rom does this at the cost of having high levels of aliasing—almost as bad as bilinear scaling.

I've gone back and forth on this, testing different luma algorithms with different material, and unless someone can find a material where chroma aliasing is a problem, I'd say that you could get away with Bicubic 60 as the "softest" chroma algorithm (for reference, it seems like Catmull-Rom and Bicubic 50 are the same)

While it seemed like Spline 3 was a possibility (lower aliasing and higher sharpness than Bicubic) I did find some material where there was a noticeable decrease in quality caused by ringing.

As always, I would have to spend more time testing (ultimately, it takes a while of also just watching content and noticing problems) but Bicubic 60 may be the best compromise for chroma—either that or Bicubic 75 which I had settled on previously. Any softer than that and you run into desaturation problems, and any sharper than Bicubic 75, you run into ringing problems.
Use of Spline36 Luma also slightly restores saturation and reduces the chroma aliasing of Catmull-Rom. The two complement each other well, if aiming for the sharpest result with least ringing as I am.I can't say I agree with that at all. Using a ringing form of luma scaling with an overly soft/desaturated chroma algorithm does not "compliment" each other and "fix" the desaturation/dimming.

As I mentioned above, chroma ringing/halos only becomes a concern when you are using a luma resizer which produces moderate ringing/halo itself. Since you use SoftCubic for luma, you can get away with pin-sharp chroma at the expense of ringing, since SoftCubic will not increase and actually partially hide the ringing after resize. I wouldn't be surprised if Spline36 chroma complemented it well.I do most of my testing for chroma with nearest neighbour selected for luma so that my choice of luma scaling has no influence on it. Other than high levels of SoftCubic, the luma algorithm has a minimal effect on this at best anyway, mostly it just makes problems easier to spot when using Nearest Neighbour.

Ringing absolutely has an effect on chroma, regardless of the luma algorithm used, but it does not show up in the same way that it does with luma. The easiest way I can describe it is that it makes the edges of things look "bevelled" and you start seeing "hotspots" on finer details.

It's difficult to show, especially as a .gif, but hopefully this will illustrate it somewhat. SoftCubic was used for luma scaling to show that your choice of chroma scaling is still important.
http://i.imgur.com/qWnsP.gif

The main issues are indeed a blurry image and extreme loss of contrast which I can't stand. When an image is blurred so much that you are unable to easily see fine detail, I consider that as lost detail.Again, do you have any examples of this?

I'm curious to see what content you are using that has a lot of high frequency detail that is lost when using SoftCubic. Most high frequency detail in DVD-quality video is ringing/mosquito noise which are compression artefacts/mastering problems.

As I've said before though, you have to pick your poison. You either use an algorithm that is bad for aliasing (Mitchell-Netravali, Catmull-Rom) ringing (Bicubic, Lanczos, Spline) or sharpness. (SoftCubic)

For me, the priority of things I want to avoid are aliasing, then ringing and, I can tolerate a reduction in high frequency contrast as it's largely irrelevant to DVD-quality film content.

Luma scaling is up to the viewer, I still strongly believe that there are some algorithms (like all of SoftCubic, Mitchell-Netravali and Lanczos) that simply produce incorrect results when used for chroma scaling, regardless of the source.
Even though the detail may technically still be there if you sharpened the image back up on your TV afterwards, you shouldn't really be relying on TV post-processing to fix your image.I do not. (not sure where that came from?)

TV recordings, Blu-ray's re-encoded to 720p, and even things like iTunes downloads.Ah, everything here is broadcast in 1080i (and I was under the impression that other than a handful of sports channels, that was the case everywhere) I see no reason to compress Blu-ray to 720p other than piracy, and I was under the impression that iTunes content was DRMed and would not work with madVR.

It's rare for even 1080p video of Blu-rays to have much resolution above 720p. TV broadcast at 1080i can be even worse.But the source is 1080i/p and does not need rescaled, which was my point.

Bicubic75 easily wins all of the tests I did, be it chroma or luma. Hopefully madshi adds Bicubic100, I would really like to do some tests on that.After going back to a more compressed source rather than DVD, I am thinking of rescinding my recommendation of Lanczos 3 luma scaling. (as a choice between it and SoftCubic 50) It does very poorly with fine, high contrast edges, so I would probably go with one of the Bicubic variants. (probably Bicubic 75, but I couldn't say without further tests)

As I've said before though, my personal choice is somewhere between SoftCubic 50–70 for luma scaling, as most video I watch that needs upscaled is of DVD or lower resolution where the sharper algorithms just show too much ringing/aliasing. Perhaps they become more acceptable with 720p sources.
Since madshi also seems to like SoftCubic on chroma (like I do) an algorithm that would do SoftCubic(70-100) and also color-correct (counter the desaturation) would improve SoftCubic even more. Not sure of that is easily doable, though.I don't think it would be an easy fix. The desaturation/dimming of colour seems to simply be because it's being softened too much. I haven't tested it, but I am confident in saying that large areas of the same colour will not be desaturated in the same way that finer details are. (basically the smaller the detail, the more it seems to desaturate)

[cyberbeing]

Catmull-Rom does this at the cost of having high levels of aliasing—almost as bad as bilinear scaling.
What aliasing? You seem to think this is a problem, but it's really not in my opinion. You'll virtually never see chroma aliasing in real content unless you use nearest neighbor or your source itself is flawed.

While it seemed like Spline 3 was a possibility (lower aliasing and higher sharpness than Bicubic) I did find some material where there was a noticeable decrease in quality caused by ringing.

As always, I would have to spend more time testing (ultimately, it takes a while of also just watching content and noticing problems) but Bicubic 60 may be the best compromise for chroma—either that or Bicubic 75 which I had settled on previously. Any softer than that and you run into desaturation problems, and any sharper than Bicubic 75, you run into ringing problems.
That's the main reason I don't use Spline36/Spline36 day-to-day, since I as well identified ringing problems using Spline36 for chroma.

The 'best compromise' is entirely subjective.

Everything sharper than Catmull-rom will produce noticeable ringing in the chroma. You may need to zoom way in to see it, but it's there, even with Bicubic -0.60. Everything softer than Catmull-rom has noticeable desaturation. Catmull-rom offers the most mathematically accurate 2x chroma Bicubic scaling. Minor residual aliasing from the sub-sampled chroma is a non-issue, especially on higher resolution content, since it contributes to perceived sharpness. To me this makes Catmull-rom the best compromise for chroma, since I find anything which rings more as unacceptable.

If you're not bothered by the level of chroma ringing which Bicubic -0.75 produces, there is no reason to question your choice. Bicubic -0.75 very well may be the best compromise for you.

I can't say I agree with that at all. Using a ringing form of luma scaling with an overly soft/desaturated chroma algorithm does not "compliment" each other and "fix" the desaturation/dimming.
Catmull-rom chorma is not overly soft or noticeably desaturated from my testing, so I can't say I agree with your assessment either.

Using Spline36 luma on chroma which contains ringing will produce stronger ringing. Remember that madVR luma scaling is for resizing combined luma/chroma not just luma.

In any case, they do complement each other:
Increased contrast complements any reduction in chroma saturation.
Increased sharpness complements any chroma blurring
Reduction in aliasing complements any chorma aliasing

Ringing absolutely has an effect on chroma, regardless of the luma algorithm used
Correct. The difference is using SoftCubic luma will not enhance the chroma ringing like Spline36 luma would.

Again, do you have any examples of this?
It happens with every video so it shouldn't need an example. It would seem you just aren't bothered it.

I'm curious to see what content you are using that has a lot of high frequency detail that is lost when using SoftCubic.
In the past few years, almost exclusively Blu-rays and HDTV. The only time I ever watch something on DVD is if it's only available on DVD.

Most high frequency detail in DVD-quality video is ringing/mosquito noise which are compression artefacts/mastering problems.

And this may be the heart of the issue from your point of view.

A desire to hide ringing, noise, artifacts, and/or mastering problems in your source is a good reason to use SoftCubic luma. This is the entire reason I stressed that my choices based around high-quality source material which doesn't have flaws which need to be hidden.

For me, the priority of things I want to avoid are aliasing, then ringing and, I can tolerate a reduction in high frequency contrast as it's largely irrelevant to DVD-quality film content.

This is where we differ. For luma I care most about sharpness, then ringing, and aliasing. For chroma I care most about ringing, then sharpness, then aliasing. I desire all high frequency detail retained with optimal per-pixel sharpness including things like film grain. If something has flaws, I want to see it (and potentially fix it with filtering/de-interlacing/re-encoding to x264).

I do not. (not sure where that came from?)
It came up because it should be near-impossible to see fine detail after using SoftCubic luma scaling because of the strong reduction in contrast and softened edges. Using SoftCubic chroma has a near identical effect on making fine chroma detail near-impossible to see because of reductions in saturation and softened edges.

The question then becomes, did you actually want to see what was hidden, or was it just artifacts? If it was actual real temporal detail and not artifacts on a higher quality source, would you then be bothered by not seeing it? You may just be the type of person who prefers a artifact-free, softer & denoised image, rather than a sharper one with more detail but minor artifacts.

[Abradoks]

64 bit madVR is a complete waste of time and effort for no benefit.

Unless I am missing something, why use 64 bit? Like pirlouy said, you are not exactly being very clear.
Personally I don't care about 64-bit madVR, but that discussion is getting annoying.
1) Well-written software needs no effort to be ported to 64-bit architecture, it only needs (cross-)compilation. The only exception is hand-written assembly, which isn't hard to port either.
2) Existence of 64-bit OS is enough reason to have your software running there in native mode. zmaster clearly stated, that he wants madVR to work with 64-bit players and decoders, which may be faster than 32-bit. And he's not the only one who asks for it.
BTW, it's general problem with closed-source software, which makes switching to the new architectures much harder, than it should be.

[iSunrise]

I don't think it would be an easy fix. The desaturation/dimming of colour seems to simply be because it's being softened too much. I haven't tested it, but I am confident in saying that large areas of the same colour will not be desaturated in the same way that finer details are. (basically the smaller the detail, the more it seems to desaturate)
Typical "real-life" footage or movies don´t have large areas of the same color, though, so this is not valid in most of the cases. Anime, disney or very old game footage may have, but if there´s no desaturation, why bother, all is fine then. The desaturation happens, because there is some kind of pixel or pixel area averaging going on, and that average value gets applied to the neighbouring pixels, which alters the colors. The finer the detail, the more averaging is going to be applied. You stated it yourself, "large areas of the same color" = no details = no desaturation.

Video games nowadays look more and more like movies, so apart from the above cases, the desaturation is what bothers me the most WRT the current SoftCubic algorithm. Reducing the contrast is not something I would think is beneficial. On future OLEDs, we should see the desaturation even better, because it´s already apparent on LCDs.

I guess that madshi would have to implement some kind of saturation boost into the SoftCubic algorithm or pre-process the image with more saturation, before SoftCubic gets applied. If we had some kind of manual control over saturation, we could also do it ourselves, though. If I understood madshi right, we are going to have some manual control over the final image, if we want to.