madVR - high quality video renderer (GPU assisted)

[robpdotcom]

There's no problem at all using CoreAVC with madVR. It's what I'm doing here, too, atm.

I wanted to do some IVTC testing with CoreAVC+madVR, but it seems there is no option to output interlaced frames from CoreAVC - the closest option is: None (Weave). Am I missing something?

[akane750]

Hi.

In case it matters, I use MPC-HC + madVR combo on a Win7 x64 OS.

I recently upgraded madVR to v0.80 (from way back - v0.73) and noticed something : when, during video playback, I hit right arrow key ("frame step forward" hotkey ONLY - mouse-click/spacebar don't cause it to happen), the playback pauses...at least it pretends to - MPC interface's Pause button becomes "selected" (indicating the playback's paused) and underneath it "Playing" status changes to "Paused", BUT the timer in the lower right corner keeps going, the playback progress slider keeps going, and audio keeps playing*. Then, if during this "fake pause", I perform any of the pause actions again, it seems to, let's call it, "actually pause"**. Afterwards, resuming playback by taking it off pause makes the player jump forward as it tries to bring video back in sync with audio. For FullHD movies this pretty much makes the player eat up all available resources and stutter for a long time (might as well save yourself the time and simply close the player window and open the movie in a new one). By rolling back several times, I traced this bug's first appearance back to v0.78 (so right now I'm using v0.77)

*there's no sound anymore, but adjusting player volume during this "fake pause" makes it come back - as if audio was muted instead of paused, although volume, prior to the adjustment, is at maximum.

**spacebar/mouse-click makes the player report all the frames between when "fake pause" started and when "actual pause" started as dropped; "framestep back", however, does not - apparently, it performs the "framestep back" action on the clip's "actual" position.


P.s. two thoughts of my own - first, I noticed "improved framestepping" in v0.78 patch notes - maybe it's related to that fix? And second, in case the wall of text above doesnt make much sense, after all, this bug "performs" similar to the "Picture Freeze" feature on TVs - picture freezes, audio keeps playing, "un-freezing" makes it skip ahead to sync with the signal again.

[shimaflarex]

Hi.

In case it matters, I use MPC-HC + madVR combo on a Win7 x64 OS.

I recently upgraded madVR to v0.80 (from way back - v0.73) and noticed something : when, during video playback, I hit right arrow key ("frame step forward" hotkey ONLY - mouse-click/spacebar don't cause it to happen), the playback pauses...at least it pretends to - MPC interface's Pause button becomes "selected" (indicating the playback's paused) and underneath it "Playing" status changes to "Paused", BUT the timer in the lower right corner keeps going, the playback progress slider keeps going, and audio keeps playing*. Then, if during this "fake pause", I perform any of the pause actions again, it seems to, let's call it, "actually pause"**. Afterwards, resuming playback by taking it off pause makes the player jump forward as it tries to bring video back in sync with audio. For FullHD movies this pretty much makes the player eat up all available resources and stutter for a long time (might as well save yourself the time and simply close the player window and open the movie in a new one). By rolling back several times, I traced this bug's first appearance back to v0.78 (so right now I'm using v0.77)

*there's no sound anymore, but adjusting player volume during this "fake pause" makes it come back - as if audio was muted instead of paused, although volume, prior to the adjustment, is at maximum.

**spacebar/mouse-click makes the player report all the frames between when "fake pause" started and when "actual pause" started as dropped; "framestep back", however, does not - apparently, it performs the "framestep back" action on the clip's "actual" position.


P.s. two thoughts of my own - first, I noticed "improved framestepping" in v0.78 patch notes - maybe it's related to that fix? And second, in case the wall of text above doesnt make much sense, after all, this bug "performs" similar to the "Picture Freeze" feature on TVs - picture freezes, audio keeps playing, "un-freezing" makes it skip ahead to sync with the signal again.

Same happens here on MPC-HC (but the hotkey to framestep is Ctrl+Right, but I'm not sure if I've changed it; Frame stepping through Play->Framestep also reproduces this issue).
Didn't test another renderer.

[madshi]

using madvr the colour setting in the ati control center doesn't work with videos
can someone fix this?
This is as intended because I don't trust the GPU algorithms to work reliably. I do plan to add some kind of color controls to madVR, but not very soon. If you want to calibrate your display properly, you can use the yCMS functionality. But for that you'd need to have your own meter.

While testing I reset my madVR settings and while watching madVR's OSD I missed two things
1. Total processing time (real time not the sum off all threads time) of each frame. Currently we have rendering, deinterlacing and present times but they can't be simply added together because they are in different threads, right?
They can't be simple added, and there's no way to calculate a proper "sum", because the processing partially overlaps.

2. Count of repeated frames - I've asked about this a long time ago but it wasn't possible with the rendering path used then. Has this changed now?
Not yet.

This reminded me of another problem that I'm not sure I've seen reported - when the playback stops at the end of the movie the player's OSD that's drawn through the special madVR API isn't working.
The madVR OSD APIs only work if playback isn't stopped. The OSD only works in paused or playing mode.

Though we allready see that this is not the case for every Hardware with Intels new HD Graphics Generation for example and most probably it's gonna change @ least for the scaling part also in the new AMD and Nvidia Generations so in the End Shader based scaling will be not a good thing anymore for everyone but revert into a bad thing for many user as GPU resources will be consumed for nothing and could be avoided.
What you say would be true only if the DXVA based scaling had the same quality as shader based scaling. That is not the case with current hardware. E.g. Intel's scaling currently only can do NV12 -> NV12 scaling, 8bit only, without dithering. I don't know what the future will bring, but you shouldn't praise DXVA scaling quality before it actually proves to produce quality worth praising.

And when there will be options for filters FFD?
I'm not sure I understand what you mean?

i've tested a few weeks different configs to avoid dropped frames, but there is no solution. tested with 1080p23 and 1080p24 material.
First of all you should check the madVR OSD to make sure that the frame drops are not caused by empty queues. If the frame drops occur even though all queues are full, then it's probably related to a refresh rate mismatch. This problem should not occur when you use Reclock, though. You should disable all VSync related options in Reclock, also make sure you reset the Reclock timing database.

@madshi
It seems that you block "Pause" message to prevent disappearing subtitles. But what about MPC-HC translations? For example with Polish or Czech language message still shows and subtitle disappear.
Currently I'm manually blocking the "Pause" message for all languages I know. If you tell me the Polish/Czech "Pause" translation, I can block them, too.

I wanted to do some IVTC testing with CoreAVC+madVR, but it seems there is no option to output interlaced frames from CoreAVC - the closest option is: None (Weave). Am I missing something?
Sadly, CoreAVC does not output any interlaced information. For Blu-Rays it's not much of a problem, just manually activate madVR's deinterlacing + IVTC and it should still work alright. In any case it won't harm to complain in the CoreAVC thread. I've tried to convince BetaBoy to output proper deinterlacing information, but it seems the CoreAVC devs don't like the idea. IMHO they're misguided in this case.

I recently upgraded madVR to v0.80 (from way back - v0.73) and noticed something : when, during video playback, I hit right arrow key ("frame step forward" hotkey ONLY - mouse-click/spacebar don't cause it to happen), the playback pauses...at least it pretends to - MPC interface's Pause button becomes "selected" (indicating the playback's paused) and underneath it "Playing" status changes to "Paused", BUT the timer in the lower right corner keeps going, the playback progress slider keeps going, and audio keeps playing*.
I'll check if I can reproduce this.

[kasper93]

Currently I'm manually blocking the "Pause" message for all languages I know. If you tell me the Polish/Czech "Pause" translation, I can block them, too.
It would be great. I can help with Polish:
Pause - Wstrzymano
Play - Odtwarzanie

For other translations we can check mpc-hc translation files for this message. If you want to add other also, I've made list Pause.txt (http://dl.dropbox.com/u/16282309/mad/Pause.txt) (UTF-8)

[madshi]

It would be great. I can help with Polish:
Pause - Wstrzymano
Play - Odtwarzanie

For other translations we can check mpc-hc translation files for this message. If you want to add other also, I've made list Pause.txt (http://dl.dropbox.com/u/16282309/mad/Pause.txt) (UTF-8)
Thanks, added all those.

[leeperry]

just manually activate madVR's deinterlacing + IVTC
humm, how do you do that? All I can see in mVR is an option that says "activate deinterlacing if upstream filter says so"...I'd love an option to force IVTC in the VR as a last resort, but I guess the option you're referring to is in the decoders that are built into mVR or something?

[madshi]

humm, how do you do that? All I can see in mVR is an option that says "activate deinterlacing if upstream filter says so"...I'd love an option to force IVTC in the VR as a last resort, but I guess the option you're referring to is in the decoders that are built into mVR or something?
You can force deinterlacing on by pressing "Ctrl+Alt+Shift+D" (first press shows the current deinterlacing state, 2nd press toggles deinterlacing on/off). You can toggle madVR between video and film mode deinterlacing by using the key combo "Ctrl+Alt+Shift+T". Film mode uses a self written IVTC algorithm, which does proper decimating. DXVA can do IVTC, but not decimating.

Edit: This works for both internal and external decoders.

[robpdotcom]

Just an idea. I don't know how far off automatic detection of telecined content is, but in the meantime I propose a work-around:

If the file name ends in ".23p", and the content is 59i, then IVTC should automatically be enabled. Or, we could add something like ".ivtc" to the file name.

One more thing. Will it ever be possible for Reclock to know that madVR is performing IVTC? Without knowing, it seems like it cannot sync playback, since it thinks the video is 29Hz.

[magnusr]

Is the Radeon 7750 good enough for madvr? Or do I need a 7770 in my HTPC?

[ryrynz]

Madshi, can we get a list of all the keyboard shortcuts listed in a text file with the next version please?

[madshi]

If the file name ends in ".23p", and the content is 59i, then IVTC should automatically be enabled.
I already added this to my sources a couple of days ago. The file name doesn't need to end in ".23p", there just needs to be a "23p" somewhere in the file name.

One more thing. Will it ever be possible for Reclock to know that madVR is performing IVTC? Without knowing, it seems like it cannot sync playback, since it thinks the video is 29Hz.
I don't know exactly how Reclock works. I thought maybe it would count the Direct3D::Present() calls somehow? If not, the only it could know that madVR is applying IVTC is by directly communicating with madVR. Last time I contacted James (Slysoft) about something like this, he wasn't interested at all.

Is the Radeon 7750 good enough for madvr? Or do I need a 7770 in my HTPC?
Of course more speed never hurts, but the 7750 should do great. In the end it all depends. If you want to upscale to 4K with deinterlacing and best scaling algorithms then maybe even the 7770 might not be fast enough, I don't know for sure. Maybe I'll also add some funny scaling algorithms at some time in the future which may require insane amounts of GPU power. Don't know right now. IMHO between the 7750 and 7770 get the 7750, because it delivers more performance/watt. Or if you want to get something faster than the 7750, maybe the 7850 could be another good choice. It's supposed to come out early March. Of course you could also wait to see what NVidia will have to offer with their new 28nm chips coming out in April.

Madshi, can we get a list of all the keyboard shortcuts listed in a text file with the next version please?
Just search the changelog for "ctrl" and you should find all the keyboard shortcuts.

[ryrynz]

I remember spotting them in there now, there is the odd forum post regarding them though, any chance of getting them on the front page perhaps?

Maybe I'll also add some funny scaling algorithms at some time in the future which may require insane amounts of GPU power.

Please do.

[aufkrawall]

madshi,
some more detail regarding the FRAPS issue:
The video doesn't play with the official FRAPS decoder (install FRAPS to get it) only.
With e.g. LAV decoder and splitter it works.

[rijnton]

Sorry if this has already been addressed in this thread, but I didn't have the time to read the whole, thread being that long.
Today I decided to set madVR as output in MPC-HC and it works great !
However I have a question with regards to Display Modes. I have a HDTV and entered 1080p23, 1080p24, 1080p25, 1080p29, 1080p30, 1080p50 and 1080p60 because these values are supported by my tv. Playing 23,976 videos my tv switches to 1080/24p.
But when I play a dvd that has 25fps, the refresh rate stays at 60Hz. I expected that it should go to 50Hz as entered in the Display Modes. Am I doing something wrong here ?
Btw, when I Ctrl+J I see Display rate 60Hz, Composition rate 23,976 and Movie 25fps.
What does "Composition rate" mean ?

[FlashGordon]

Will it ever be possible for Reclock to know that madVR is performing IVTC? Without knowing, it seems like it cannot sync playback, since it thinks the video is 29Hz.

At the moment I can't test this, but try telling Reclock that the framerate is 23.976 instead of 29.97. You'd have to go into the Reclock properties and manually change it from "Automatic" to 23.976. I think in theory this will tell you that it is synchronizing playback, but I'm not sure exactly how smooth it'll be as I haven't watched anything in this manner.

The way I understand it is, Reclock detects the framerate before IVTC as 29.97, and madvr performs IVTC after that fact... but I'm not sure how well it'll sync if you have Reclock treat it as 23.976 to begin with. Worth a try anyway

[robpdotcom]

try telling Reclock that the framerate is 23.976 instead of 29.97.

Thanks, I forgot about that option.

[madshi]

I remember spotting them in there now, there is the odd forum post regarding them though, any chance of getting them on the front page perhaps?
The changelog contains all the information that is necessary, I think. I don't want to put the keys on the first page because I plan to make the key shortcuts adjustable in a future version.

madshi,
some more detail regarding the FRAPS issue:
The video doesn't play with the official FRAPS decoder (install FRAPS to get it) only.
With e.g. LAV decoder and splitter it works.
Ok. Can you point me to the exact official FRAPS decoder download with which the problem occurs?

However I have a question with regards to Display Modes. I have a HDTV and entered 1080p23, 1080p24, 1080p25, 1080p29, 1080p30, 1080p50 and 1080p60 because these values are supported by my tv. Playing 23,976 videos my tv switches to 1080/24p.
But when I play a dvd that has 25fps, the refresh rate stays at 60Hz. I expected that it should go to 50Hz as entered in the Display Modes. Am I doing something wrong here ?
If you enter display modes in that list which your GPU does not support, Direct3D will just choose a "random" refresh rate instead. If your DVD is 25fps and you have 1080p25 in your display mode list, then madVR will select 1080p25 as the best fitting mode. But most GPUs don't support 1080p25 (unless you setup a custom resolution mode). So my best guess would be that madVR told Direct3D to use 1080p25, but your GPU does not support that mode, so Direct3D used 1080p60 instead. The solution is easy: Make sure that all modes you entered into the list are actually supported by your GPU. Probably just removing 1080p25 would solve the issue. Or alternatively create a 1080p25 custom resolution, so that your GPU supports it (NVidia should allow this, ATI probably not).

Btw, when I Ctrl+J I see Display rate 60Hz, Composition rate 23,976 and Movie 25fps.
What does "Composition rate" mean ?
That's the rate Windows Aero / Desktop Composition renders at. Ideally for smooth motion playback it should be identical to your display refresh rate.

[Andy o]

IME ReClock syncs (resamples) just fine, even if you don't tell it manually the frame rate is 23.976 and the icon is yellow. It will still resample to a derivative of 60/1.001 because it will still detect your refresh rate as 23.976 (or whatever is close).

[Qaq]

An idea about the Luma&Chroma scalers:
Kinda same thing like ABX test in foobar - by pressing some hot key you start changing lima or chroma (or combo) scalers on the fly and see what looks better for you. For statick or moving picture, with adjustable time intervals, etc.
What do you think?

[pankov]

They can't be simple added, and there's no way to calculate a proper "sum", because the processing partially overlaps.
I see. I just wanted to have a simple way of telling if the GPU is too slow for the given content or there are other factors causing frame drops. Can you do something like "total processing time: X.XX ms" and just ignore the partial overlap .... if it's small enough?
The madVR OSD APIs only work if playback isn't stopped. The OSD only works in paused or playing mode.
Is this planned to change or it's not possible at all and the player should change the way it draws its UI when the playback is stopped?
This will lead to blinking of the screen but I guess it's better than not drawing some UI elements.

[madshi]

An idea about the Luma&Chroma scalers:
Kinda same thing like ABX test in foobar - by pressing some hot key you start changing lima or chroma (or combo) scalers on the fly and see what looks better for you. For statick or moving picture, with adjustable time intervals, etc.
What do you think?
Even better would be a split screen preview (e.g. left side scaler A, right side scaler B). These kind of features would be nice, but currently I've so many more important things to do...

I see. I just wanted to have a simple way of telling if the GPU is too slow for the given content or there are other factors causing frame drops. Can you do something like "total processing time: X.XX ms" and just ignore the partial overlap .... if it's small enough?
No, it just doesn't make any sense. A good way of checking your GPU performance would be the madVR benchmark version. You should find the download with a forum search.

Is this planned to change or it's not possible at all and the player should change the way it draws its UI when the playback is stopped?
It's not planned to change, it would cost a lot of time to change. Blight is planning to pause playback instead of stopping it, to work around the problem.

[pankov]

Thanks for the info and all your great work

[Qaq]

Even better would be a split screen preview (e.g. left side scaler A, right side scaler B). These kind of features would be nice, but currently I've so many more important things to do...
No rush, I just want to disсuss it if anyone interested.

[kerman]

2. Count of repeated frames - I've asked about this a long time ago but it wasn't possible with the rendering path used then. Has this changed now?

Not yet.

Im confused here, No count on repeated frames? just a real time estimation. But so, whats the difference -technically speaking- between repeated, dropped and delayed frames? I mean, the cause or implications behind each one. Is the lack of processing or displaying, what determines repeated, dropped or delayed?

I thought refreshing rate and fps were totally self-reliant processes, and their syncing was ony conditioned by the match of their respective clocks.

[madshi]

Im confused here, No count on repeated frames? just a real time estimation. But so, whats the difference -technically speaking- between repeated, dropped and delayed frames? I mean, the cause or implications behind each one. Is the lack of processing or displaying, what determines repeated, dropped or delayed?

I thought refreshing rate and fps were totally self-reliant processes, and their syncing was ony conditioned by the match of their respective clocks.
Oh man, explaining all this stuff in detail would take more time than I have at my hands. So I'll just mention 2 examples:

(1) If your GPU display mode 1080p25, but the movie framerate is 24p, the video decoder will output 24 frames per second, but the GPU sends 25 frames per second to the display. So one of the 24 frames coming from the decoder must be displayed twice ("repeated"). Without the repeating the video would play too fast. madVR does not currently count/report repeated frames.

(2) If your GPU display mode is 1080p24, but the movie framerate is 25p, the video decoder will output 25 frames per second, but the GPU sends only 24 frames per second to the display. So madVR has to drop one of those 25 frames, because otherwise playback would be too slow.

If the video playback speed is too high or too low, audio sync would be lost, because audio always plays with the "correct" speed. Except if you're using Reclock and if you're telling Reclock to change the audio speed. But that's a complicated topic on its own.

[njfoses]

New official nvidia drivers released. http://www.geforce.com/News/articles/nvidia-geforce-295-73-whql-drivers-released

[rijnton]

Probably just removing 1080p25 would solve the issue. Or alternatively create a 1080p25 custom resolution, so that your GPU supports it (NVidia should allow this, ATI probably not).


That's the rate Windows Aero / Desktop Composition renders at. Ideally for smooth motion playback it should be identical to your display refresh rate.

Thanks madshi. Removing 1080p25 did it.
Composition rate is also at 50Hz now.

[aufkrawall]

Ok. Can you point me to the exact official FRAPS decoder download with which the problem occurs?

FRAPS VFW decoder is included in the free basic FRAPS version:
http://www.fraps.com/setup.exe
It happens both with MPC HC and LAV avi splitters.

Thanks. :)

[Razoola]

New official nvidia drivers released. http://www.geforce.com/News/articles/nvidia-geforce-295-73-whql-drivers-released

Pity they still have the same issue of thinking the HDMI cable I have attached to my gfx card is DVI.

[robpdotcom]

IME ReClock syncs (resamples) just fine, even if you don't tell it manually the frame rate is 23.976 and the icon is yellow. It will still resample to a derivative of 60/1.001 because it will still detect your refresh rate as 23.976 (or whatever is close).


I'm pretty sure the yellow icon means it isn't changing the speed.

@madshi:

Do you have a preference for where we upload samples? I used to use megaupload or filesonic, but they have been shut down or no longer allow file sharing.

[madshi]

IIRC, mediafire is pretty good.

[pururin]

IIRC, mediafire is pretty good.

Indeed.

[6233638]

Even better would be a split screen preview (e.g. left side scaler A, right side scaler B). These kind of features would be nice, but currently I've so many more important things to do...I actually find this kind of comparison to be fairly useless unless you're using a very small display. Having a key combination that toggled the settings would make differences much easier to spot, as your eyes are fixed in one place.

What I do when comparing scalers is take screenshots, and paste them as new layers in a Photoshop document. That way I can easily toggle between different scaling options, and view them at higher magnification for close scrutiny. Photoshop also has some features which make seeing the difference between images easier as well.

If there was some way to automate creation of several screenshots of the current image and composite them into a layered TIF file, that would be much more useful.

What I would really like though, if it's possible, would be to view chroma (and luma, I suppose) on its own. That would make selecting a chroma algorithm much easier. (though at this point, I feel like I've gone through enough test cases that I have a good idea of what the options are now)

An idea about the Luma&Chroma scalers:
Kinda same thing like ABX test in foobar - by pressing some hot key you start changing lima or chroma (or combo) scalers on the fly and see what looks better for you. For statick or moving picture, with adjustable time intervals, etc.
What do you think?This might make things easier for you:

Algorithms which should not be used for Luma upscaling: Nearest Neighbour—Severe aliasing.
Bilinear—High levels of aliasing.
Catmull-Rom—High levels of aliasing and ringing.
Bicubic 50—Virtually identical to Catmull-Rom
Bicubic 60—Marginally less aliasing and more ringing.
SoftCubic 100—Far too soft to be used for just about anything, though it will actually hide ringing in the source if it's really bad.
Lanczos 4—Significant ringing for a slight improvement in aliasing over Lanczos 3.
Lanczos 8—Hilarious amount of ringing, though it does handle aliasing well.
Spline 4—Virtually no change from Spline 3, except for adding a lot of ringing.I would not recommend using linear light upscaling at all, but if you really want to use it, SoftCubic 100 is the only one where it doesn't show a dramatic reduction in image quality. The problem with is is "negative ringing" where objects get thick black outlines around them, which are often highly aliased. It's a shame, as the appearance of linear light scaling definitely is better otherwise.

This is just one example, and it's taken from game footage because that makes the differences easier to spot. I've also changed the levels in Photoshop to make things easier to see.

I recommend that you open these in separate browser tabs (middle click, or ctrl+click) and swap between them to compare. (ctrl+tab to go forwards, ctrl+shift+tab to go back) They're posted in order of sharpness.

http://img718.imageshack.us/img718/2146/nearestneighbour.th.jpg
Nearest Neighbour (http://img718.imageshack.us/img718/2146/nearestneighbour.jpg), to give you an idea of the source material. Note that there is no black outline around any of this.

So, out of the algorithms that are left, what do we have?

Spline 3 & Lanczos 3. These are the most complex and require the most GPU power. They are the sharpest algorithms, and as a result, have quite a lot of ringing. The difference between the two really comes down to Lanczos 3 having less aliasing, but more ringing than Spline 3.

Bicubic 75 is still a sharp algorithm that has even less ringing than either and still handles aliasing reasonably well.

http://img820.imageshack.us/img820/7003/lanczos3.th.jpg
Lanczos 3 (http://img820.imageshack.us/img820/7003/lanczos3.jpg)
http://img546.imageshack.us/img546/6248/spline3.th.jpg
Spline 3 (http://img546.imageshack.us/img546/6248/spline3.jpg)
http://img545.imageshack.us/img545/4503/bicubic75.th.jpg
Bicubic 75 (http://img545.imageshack.us/img545/4503/bicubic75.jpg)

The next step down in sharpness, is Mitchell-Netravali and SoftCubic 50. They're fairly similar in terms of overall sharpness, though Mitchell-Netravali is sharper on fine details.

Mitchell-Netravali is a great balance between sharpness and ringing, but unfortunately, it shows quite a lot of aliasing. (look at the diagonal line at the left of the image)
SoftCubic 50 has somewhat more ringing, but handles aliasing much better.

http://img441.imageshack.us/img441/3513/mitchellnetravali.th.jpg
Mitchell-Netravali (http://img441.imageshack.us/img441/3513/mitchellnetravali.jpg)
http://img850.imageshack.us/img850/6282/softcubic50.th.jpg
SoftCubic50 (http://img850.imageshack.us/img850/6282/softcubic50.jpg)

And below that, as far as sharpness is concerned, are higher levels of SoftCubic. Personally I don't see much use to SoftCubic 60, but would use 70/80. SoftCubic 80 has the least amount of ringing of any of the scaling algorithms available in madVR, and very low aliasing. As the name implies, it can look quite soft, but I find this tends to hide MPEG compression artefacts in the source without removing detail as a noise reduction algorithm would. 70 is slightly sharper at the expense of some ringing being visible in rare circumstances.

http://img837.imageshack.us/img837/1561/softcubic80.th.jpg
SoftCubic 80 (http://img837.imageshack.us/img837/1561/softcubic80.jpg)

It looks very bad at this high magnification, but when actually watching DVD content, I don't really notice the sharpness, just the lack of artefacts.


Out of all the algorithms, I would probably recommend choosing between Bicubic 75, Mitchell-Netravali, or SoftCubic 80, depending on your preference for sharpness and your tolerance for aliasing/ringing.



Chroma is a much easier decision in my opinion. The choice there is between Bicubic60/75, or Mitchell-Netravali. Anything else is either too sharp and exhibits ringing, or too soft and desaturates parts of the image.

Right now, I am using Mitchell-Netravali. I had avoided this in the past due to aliasing problems being visible with Luma upscaling, but I have yet to find anything where this has been a problem, and MN exhibits the least amount of ringing with chroma scaling.

So far, I have only found one image where using SoftCubic may actually be beneficial for chroma scaling, and I believe that this is a problem in the source material, rather than the scaling algorithm used:

http://img51.imageshack.us/img51/2146/nearestneighbour.th.jpg
Nearest Neighbour. (http://img51.imageshack.us/img51/2146/nearestneighbour.jpg) This shows why I believe it's the source material at fault, rather than the Chroma algorithm used.
http://img542.imageshack.us/img542/4503/bicubic75.th.jpg
Bicubic 75 (http://img542.imageshack.us/img542/4503/bicubic75.jpg), tends to be most faithful to the source material where where saturation is concerned, but can exhibit ringing with some material. (shows as colour getting lighter towards the edges of objects) In some cases 60 seems to be better, in others, 75.
http://img851.imageshack.us/img851/3513/mitchellnetravali.th.jpg
Mitchell-Netravali (http://img851.imageshack.us/img851/3513/mitchellnetravali.jpg) Somewhat less saturated than Bicubic 75 on fine details, but almost no ringing.
http://img192.imageshack.us/img192/2050/softcubic100.th.jpg
SoftCubic 100. (http://img192.imageshack.us/img192/2050/softcubic100.jpg) Eliminates the banding, but completely desaturates the image.

[magnusr]

Of course more speed never hurts, but the 7750 should do great. In the end it all depends. If you want to upscale to 4K with deinterlacing and best scaling algorithms then maybe even the 7770 might not be fast enough, I don't know for sure. Maybe I'll also add some funny scaling algorithms at some time in the future which may require insane amounts of GPU power. Don't know right now. IMHO between the 7750 and 7770 get the 7750, because it delivers more performance/watt. Or if you want to get something faster than the 7750, maybe the 7850 could be another good choice. It's supposed to come out early March. Of course you could also wait to see what NVidia will have to offer with their new 28nm chips coming out in April.

Thanks for the answer. I also wonder will a 6570 or 6670 be enough aswell? Seeing they are like half the price of a 7750.

If the 6570 or 6670 is enough will the memory mather if its DDR3 or GDDR5?

[noee]

fwiw, my main HTPC now has a 6570 fanless and it handles everything I throw at it, with probably the most "intense" being the Samsung demo and a few VC1 blurays. Recently with the LAV DXVA2_native option, I've tested with EVR and it works well too, rarely breaks a sweat. DXVA2_Copyback works well too with madVR up to the bitrate of the Samsung demo. I prefer, however, software decoding with madVR rendering. I playback everything on a secondary monitor (HTPC) with jRiver Media Center.

If you've got a sample of something you're wondering about, I can test it quite easily with the various decode options (h/w, s/w).

[magnusr]

The important thing for me would be that it plays 1080p24Hz mkvs (my bluray backups). In MPC-HC using lav and madvr.

Here is from the shop in Norway:

Card NOK USD

6570 641 128
6670G 775 155
7750 1096 219
7770 1260 252

[JarrettH]

Thanks for another great comparison 6233638 :D:D

I settled on Mitchell many many months ago. IMO, HD resolution movies hold up well using Catmull, but I often forget to switch algorithms.

[Andy o]

I'm pretty sure the yellow icon means it isn't changing the speed.

It's still resampling audio though, so the video/audio mismatch is taken care of.

[ryrynz]

but I often forget to switch algorithms.

Which is why I'd love to see the ability to auto switch algorithms based on content or even better which ffdshow preset is active :D

As far as basic switching goes, Madshi could we get the ability to switch between chosen algorithms with a hotkey along with a OSD message regarding which one is active at time of switching?

[6233638]

Thanks for another great comparison 6233638 :D:D

I settled on Mitchell many many months ago. IMO, HD resolution movies hold up well using Catmull, but I often forget to switch algorithms.Glad you liked it.

Mitchell-Netravali was my original choice for Luma scaling with madVR, but I found aliasing to be problematic with too many DVDs and had to abandon it in favour of SoftCubic 50, which has gradually increased to 80 over time.

I'd recommend either Mitchell-Netravali or Bicubic 75 over Catmull-Rom. If I recall correctly, Mitchell-Netravali has roughly the same amount of aliasing, but far less ringing, and Bicubic has a little more ringing, but far less aliasing. Any of the algorithms on my "do not use" list are there because better choices exist, depending on your preferred balance between sharpness, ringing and aliasing - at least in my opinion.

I debated whether or not to put Lanczos 3 on the "do not use" list, but some people do seem to prefer it over Spline, and either seems equally objectionable to me, with far too much ringing. (Spline wins some cases, Lanczos wins others, though I'm leaning towards Spline now, when I was leaning towards Lanczos previously)

[nevcairiel]

Sounds like its about time madshi implements his non-ringing Lanczos4 scaler in madVR. :)

Speaking of, can madVR somehow leverage the GPUs built in scaler? I've been told Intels GPUs have a nice Lanczos4 scaler with ringing reduction, and i would like to compare it to madVRs other scalers. :)
The only way i know of to access it is using vanilla EVR, but for proper comparison switching renderers might introduce other factors..

[madshi]

Which is why I'd love to see the ability to auto switch algorithms based on content or even better which ffdshow preset is active :D
All in due time.

As far as basic switching goes, Madshi could we get the ability to switch between chosen algorithms with a hotkey along with a OSD message regarding which one is active at time of switching?
It's not as easy as it might seem, because several of those algorithms also have options (e.g. Bicubic 50 / 75 etc). Well, I could cycle through every possible combination of algorithms + options, but I'm not sure how useful it would be, really. If you want to compare 2 specific algorithms, you would probably want to switch between them with a key press, instead of having to cycle through 20 or so algorithms in order to get back to the one you want to compare to.

Sounds like its about time madshi implements his non-ringing Lanczos4 scaler in madVR. :)
Yeah, I would really like to do that, cause it's an area which I find quite interesting. Sadly, I still think there are more important things I need to do first.

Speaking of, can madVR somehow leverage the GPUs built in scaler? I've been told Intels GPUs have a nice Lanczos4 scaler with ringing reduction, and i would like to compare it to madVRs other scalers. :)
Leveraging the GPU's built in scaler is on my to do list. It would work somewhat similar to how I implemented DXVA deinterlacing. The Intel scaler should be interesting, in a quick comparison is seemed slightly less capable than my non-ringing algorithm draft, but better than standard Lanczos. There's one catch, though: I've been told the Intel scaler currently only supports NV12 -> NV12 scaling, so it's limited to 8bit 4:2:0 input and output, and it doesn't use dithering internally. Anyway, it's on my to do list, but there are so many other things there... :scared: I wish I had more time for madVR, but I do need to work on my commercial projects once in a while, too, to earn my living...

[madshi]

Sounds like its about time madshi implements his non-ringing Lanczos4 scaler in madVR. :)
P.S: I'm not 100% sure yet, but it's possible I might need OpenCL for this, with Direct3D9 <-> OpenCL interop. AMD does not support D3D9 interop, at least not yet, so for now I don't really have a good compute capability for AMD GPUs. Which may make implementing non-ringing algorithms difficult/slow. I still hope that AMD will add D3D9 interop at some point in time, since OpenCL 1.2 now has an official extension for that.

[CruNcher]

I actually find this kind of comparison to be fairly useless unless you're using a very small display. Having a key combination that toggled the settings would make differences much easier to spot, as your eyes are fixed in one place.

What I do when comparing scalers is take screenshots, and paste them as new layers in a Photoshop document. That way I can easily toggle between different scaling options, and view them at higher magnification for close scrutiny. Photoshop also has some features which make seeing the difference between images easier as well.

If there was some way to automate creation of several screenshots of the current image and composite them into a layered TIF file, that would be much more useful.

What I would really like though, if it's possible, would be to view chroma (and luma, I suppose) on its own. That would make selecting a chroma algorithm much easier. (though at this point, I feel like I've gone through enough test cases that I have a good idea of what the options are now)

This might make things easier for you:

Algorithms which should not be used for Luma upscaling: Nearest Neighbour—Severe aliasing.
Bilinear—High levels of aliasing.
Catmull-Rom—High levels of aliasing and ringing.
Bicubic 50—Virtually identical to Catmull-Rom
Bicubic 60—Marginally less aliasing and more ringing.
SoftCubic 100—Far too soft to be used for just about anything, though it will actually hide ringing in the source if it's really bad.
Lanczos 4—Significant ringing for a slight improvement in aliasing over Lanczos 3.
Lanczos 8—Hilarious amount of ringing, though it does handle aliasing well.
Spline 4—Virtually no change from Spline 3, except for adding a lot of ringing.I would not recommend using linear light upscaling at all, but if you really want to use it, SoftCubic 100 is the only one where it doesn't show a dramatic reduction in image quality. The problem with is is "negative ringing" where objects get thick black outlines around them, which are often highly aliased. It's a shame, as the appearance of linear light scaling definitely is better otherwise.

This is just one example, and it's taken from game footage because that makes the differences easier to spot. I've also changed the levels in Photoshop to make things easier to see.

I recommend that you open these in separate browser tabs (middle click, or ctrl+click) and swap between them to compare. (ctrl+tab to go forwards, ctrl+shift+tab to go back) They're posted in order of sharpness.

http://img718.imageshack.us/img718/2146/nearestneighbour.th.jpg
Nearest Neighbour (http://img718.imageshack.us/img718/2146/nearestneighbour.jpg), to give you an idea of the source material. Note that there is no black outline around any of this.

So, out of the algorithms that are left, what do we have?

Spline 3 & Lanczos 3. These are the most complex and require the most GPU power. They are the sharpest algorithms, and as a result, have quite a lot of ringing. The difference between the two really comes down to Lanczos 3 having less aliasing, but more ringing than Spline 3.

Bicubic 75 is still a sharp algorithm that has even less ringing than either and still handles aliasing reasonably well.

http://img820.imageshack.us/img820/7003/lanczos3.th.jpg
Lanczos 3 (http://img820.imageshack.us/img820/7003/lanczos3.jpg)
http://img546.imageshack.us/img546/6248/spline3.th.jpg
Spline 3 (http://img546.imageshack.us/img546/6248/spline3.jpg)
http://img545.imageshack.us/img545/4503/bicubic75.th.jpg
Bicubic 75 (http://img545.imageshack.us/img545/4503/bicubic75.jpg)

The next step down in sharpness, is Mitchell-Netravali and SoftCubic 50. They're fairly similar in terms of overall sharpness, though Mitchell-Netravali is sharper on fine details.

Mitchell-Netravali is a great balance between sharpness and ringing, but unfortunately, it shows quite a lot of aliasing. (look at the diagonal line at the left of the image)
SoftCubic 50 has somewhat more ringing, but handles aliasing much better.

http://img441.imageshack.us/img441/3513/mitchellnetravali.th.jpg
Mitchell-Netravali (http://img441.imageshack.us/img441/3513/mitchellnetravali.jpg)
http://img850.imageshack.us/img850/6282/softcubic50.th.jpg
SoftCubic50 (http://img850.imageshack.us/img850/6282/softcubic50.jpg)

And below that, as far as sharpness is concerned, are higher levels of SoftCubic. Personally I don't see much use to SoftCubic 60, but would use 70/80. SoftCubic 80 has the least amount of ringing of any of the scaling algorithms available in madVR, and very low aliasing. As the name implies, it can look quite soft, but I find this tends to hide MPEG compression artefacts in the source without removing detail as a noise reduction algorithm would. 70 is slightly sharper at the expense of some ringing being visible in rare circumstances.

http://img837.imageshack.us/img837/1561/softcubic80.th.jpg
SoftCubic 80 (http://img837.imageshack.us/img837/1561/softcubic80.jpg)

It looks very bad at this high magnification, but when actually watching DVD content, I don't really notice the sharpness, just the lack of artefacts.


Out of all the algorithms, I would probably recommend choosing between Bicubic 75, Mitchell-Netravali, or SoftCubic 80, depending on your preference for sharpness and your tolerance for aliasing/ringing.



Chroma is a much easier decision in my opinion. The choice there is between Bicubic60/75, or Mitchell-Netravali. Anything else is either too sharp and exhibits ringing, or too soft and desaturates parts of the image.

Right now, I am using Mitchell-Netravali. I had avoided this in the past due to aliasing problems being visible with Luma upscaling, but I have yet to find anything where this has been a problem, and MN exhibits the least amount of ringing with chroma scaling.

So far, I have only found one image where using SoftCubic may actually be beneficial for chroma scaling, and I believe that this is a problem in the source material, rather than the scaling algorithm used:

http://img51.imageshack.us/img51/2146/nearestneighbour.th.jpg
Nearest Neighbour. (http://img51.imageshack.us/img51/2146/nearestneighbour.jpg) This shows why I believe it's the source material at fault, rather than the Chroma algorithm used.
http://img542.imageshack.us/img542/4503/bicubic75.th.jpg
Bicubic 75 (http://img542.imageshack.us/img542/4503/bicubic75.jpg), tends to be most faithful to the source material where where saturation is concerned, but can exhibit ringing with some material. (shows as colour getting lighter towards the edges of objects) In some cases 60 seems to be better, in others, 75.
http://img851.imageshack.us/img851/3513/mitchellnetravali.th.jpg
Mitchell-Netravali (http://img851.imageshack.us/img851/3513/mitchellnetravali.jpg) Somewhat less saturated than Bicubic 75 on fine details, but almost no ringing.
http://img192.imageshack.us/img192/2050/softcubic100.th.jpg
SoftCubic 100. (http://img192.imageshack.us/img192/2050/softcubic100.jpg) Eliminates the banding, but completely desaturates the image.

On Intel HD Graphics you get all this for free ;)

[Qaq]

6233638, thanks a lot for your post. I need some time to prepare my answer. I'll explain all my thoughts later.

[madshi]

@CruNcher, do you really have to do a full quote of a monster sized post if you just comment with one line?? Besides, you do not get "all this" for free with Intel HD. "All this" would mean a choice of various different chroma upsampling algorithms. Intel HD only offers one chroma upsampling algorithm, and I've not tested its quality.

[VincAlastor]

First of all you should check the madVR OSD to make sure that the frame drops are not caused by empty queues. If the frame drops occur even though all queues are full, then it's probably related to a refresh rate mismatch. This problem should not occur when you use Reclock, though. You should disable all VSync related options in Reclock, also make sure you reset the Reclock timing database.

thank you. the only thing what help to avoid dropped frames was to create a custom profile in nvidia resolution settings. now madvr shows (ctrl+j) display fps 23.97602-23.97609 and not 23.97077. thats enough accurate for 24000/1001 content.
i am wondering why there are not more user wich have problems caused dropped frames. its a general problem for nvidia and intel gpu's, like the anandtech article explained.

[madshi]

There is no temporal difference between the fields. If I run it though something like QTGMC I get 50p but with duplicate frames. If I run it though TIVTC I get good looking results. It is based on this I concluded that it was 2:2 material.
I've just analyzed this and it is field blended and thus can't be IVTC'ed properly. For your interest, here are 3 consecutive fields from the 50i stream:

field 1 (http://madshi.net/xfiles1.png), field 2 (http://madshi.net/xfiles2.png), field 3 (http://madshi.net/xfiles3.png)

You can see that fields 1 and 3 are ok, but field 2 is blended.

[Ryokurin]

The Windows Visa/7 audio engine does not introduce distortion. In fact it should be bit-exact if the input sample rate matches the configured sample rate.

Month old I know, but he may had been on to something.

http://support.microsoft.com/kb/2653312

It's a hotfix for distorted audio file sounds after you convert the sample rate in Windows 7 or Windows Server 2008 R2