madVR - high quality video renderer (GPU assisted)

[Kazuya]

Which bar ?
The one who makes a thin border at the right and bottom of the green boxes ?

I saw it but I was thinking it was the sharpness that makes it appear.

[leeperry]

I completely disabled all sharpening and I'm sending native res to madVR...it should remain untouched.

we already had this problem before, yesgrey is aware of it...sometimes I would just get a black bar right there.

[yesgrey]

we already had this problem before, yesgrey is aware of it...sometimes I would just get a black bar right there.
If I remember correctly this was due to not using the right chroma placement option.
Try adding "cplaceU=1, cplaceD=1" to t3dlut call.

Also, you could try using yv12toyuy2 and yuy2toyv12 instead of the Avisynth core functions.

[leeperry]

ok I'll try! but t3dlut is a CPU hog...I'd rather have ColorMatrix() do it properly in the first place. Hopefully tritical could have it worked out, but OK...I'll try to play around w/ additional settings first.

also, yv12toyuy2 requires mod4..not too convenient.

[Shark321]

EEDI/NEDI? I don't even know how it works. I don't expect it would run well on a GPU pixel shader. If you hate aliasing I'd suggest to use Lanczos4 (sharp, but has ringing) or SoftCubic50 (soft, but no ringing).

Here is some info about Edge-Directed Interpolation. Just look at the images at the bottom of the page, amazing results!

http://chiranjivi.tripod.com/EDITut.html

p.s.

I've tried using MadVR 0.11 and had some great results, but it crashed about once per hour (mostly when moving 30 seconds forward or back in the movie). Using Halis again.

[Kazuya]

I want it !!!! :eek:

[ryrynz]

Looks to be the holy grail of interpolation, can we get it?

[tetsuo55]

NEDI looks awesome!

And if i understood the math correctly, its relatively easy to calculate too, especially on shaders.

[tetsuo55]

looks like NEDI has already been improved with iNEDI.
But even iNEDI has already been improved with ICBI

http://www.comp.leeds.ac.uk/bmvc2008/proceedings/papers/43.pdf

The scaled images are near-perfect.

And its up to 10times faster when compared to NEDI

EDIT:
And even ICBI has been surpassed:

http://www.eurasip.org/Proceedings/Eusipco/Eusipco2009/contents/papers/1569192778.pdf

[madshi]

looks like NEDI has already been improved with iNEDI.
But even iNEDI has already been improved with ICBI

http://www.comp.leeds.ac.uk/bmvc2008/proceedings/papers/43.pdf

The scaled images are near-perfect.

And its up to 10times faster when compared to NEDI

EDIT:
And even ICBI has been surpassed:

http://www.eurasip.org/Proceedings/Eusipco/Eusipco2009/contents/papers/1569192778.pdf
Good find !!

[mark0077]

That looks sweet. :D Can't wait to see it in action (hopefully madshi will find time to add it :D)

[leeperry]

now if we could also see these in ffdshow :o

[tetsuo55]

This document talks about an improvement over MEDI, EMEDI

http://www.eie.polyu.edu.hk/eInfo4Students/msc_thesis/msc06_03.pdf

However it does not contain any math explaining it.

EDIT:

Order of release:
NEDI 2001
EIDI 2003
INEDI jan 2008
ICBI sept 2008
MEDI aug 2009
EMEDI unkown

it looks like MEDI scores better phycovisually, at the cost of a bit of PSNR and SSIM when compared to some of the older ones.

[Thunderbolt8]

^^just asking, I guess all this stuff is only useful when it comes to resizing, right? meaning when I play an original blu-ray 1080p file on a 1080p screen it wont be affected?

[Grmpf]

^^just asking, I guess all this stuff is only useful when it comes to resizing, right? meaning when I play an original blu-ray 1080p file on a 1080p screen it wont be affected?

No. The information on a Blu-ray is in 4:2:0 format, so only the luma information has 1920x1080 "pixel", but both chroma channels only have a halved horizontal and vertical resolution (1/4 of the "pixel"), so you need to upscal those before you can convert it to RGB for your display. Look at madshis first posts in the thread for some explaining pictures.

[Thunderbolt8]

hm and what if you have YCbCr as output and not RGB?

[Grmpf]

hm and what if you have YCbCr as output and not RGB?

What does your display require ? No not the HDMI-Input, the display itself, the panels - RGB right ? So somewhere there is the need to go from YCbCr to RGB and if you output YCbCr from your HTPC then you are still stuck with the electronic inside your display/projector to change it to RGB. Besides even without going through this, your YCbCr signal is still missing 3/4 of the color information and *thats* the point of scaling it up.

[leeperry]

hm and what if you have YCbCr as output and not RGB?on the ATi, the YCbCr conversion is done AFTER the RGB conversion from the video renderer RGB32 mixer...it might be done in 10bit(it's never been proven either way), but it's still encoded in 8bit TMDS in the end anyway, and then your display will make a craptacular conversion to RGB internally...so it's lossy as hell :devil:

[Egh]

NEDI looks awful, of course. Directional artefacts would create nightmare for video resizing, imo.

As for FCBI/ICBI, that is much more interesting. Though I wouldn't jump for joy in regards to ICBI implementation just yet.
Yeah, times there are for unoptimized Matlab algorithms, however they do give a hint:


A C++ implementation of the ICBI methods requires about 1.2
sec. for 4x upscaling.


and the system they refer to:


on a Dell XPS M1210 laptop
with an Intel Core2 Duo T7200 2.0 GHz CPU.

My opinion: full ICBI implementation would be computationally impossible, especially taking into account that they had 128x128 images. Imagine 640x360 video being rescaled 3x times to fit into 1920px wide screen. I reckon that would take at least a second per frame on modern dual-core systems if done on CPU (i.e. optimised code and >3.0GHz freq). Would modern GPU achieve some 50x speed up compared to that? (as that is what is required to have a smooth playback).

Judging from computational times though, FCBI may be much more realistic.

[chuuey]

i still can't understand why 0.11 works so bad for me, while 0.9 works great, to watch a movie in 0.11 on 24hz i have to defocus the player to another window for playback to start ;) and if it starts it's choppy while on 60hz it's quite stable after first few frame drops...

[leeperry]

i still can't understand why 0.11 works so bad for me
what's your full system config again?

[honai]

full ICBI implementation would be computationally impossible

I recall that just a few years ago it was also "computationally impossible" to encode 1080p material in real-time on commodity hardware.

;)

[nlnl]

NEDI looks awful, of course. Directional artefacts would create nightmare for video resizing, imo.

As for FCBI/ICBI, that is much more interesting. Though I wouldn't jump for joy in regards to ICBI implementation just yet.
Yeah, times there are for unoptimized Matlab algorithms, however they do give a hint:



and the system they refer to:



My opinion: full ICBI implementation would be computationally impossible, especially taking into account that they had 128x128 images. Imagine 640x360 video being rescaled 3x times to fit into 1920px wide screen. I reckon that would take at least a second per frame on modern dual-core systems if done on CPU (i.e. optimised code and >3.0GHz freq). Would modern GPU achieve some 50x speed up compared to that? (as that is what is required to have a smooth playback).

Judging from computational times though, FCBI may be much more realistic.

ICBI download page
http://www.andreagiachetti.it/icbi/
CUDA based GPU accelerated implementation is also available upon request

So can Nvidia cards do realtime ICBI/FCBI upscaling:)?

[chuuey]

what's your full system config again?

mac mini 2ghz, geforce 9400m 256mb windows xp with sp3 :)

[leeperry]

ouh, IGP's can be nasty sometimes...shared memory and stuff

[mark0077]

ICBI download page
http://www.andreagiachetti.it/icbi/
CUDA based GPU accelerated implementation is also available upon request

So can Nvidia cards do realtime ICBI/FCBI upscaling:)?

I hope so, I wonder how this would compare to the resizers we are using today, ie lanczos, spline etc.

If ICBI works in realtime, I wonder would MEDI work :D It "looks" alot better in the MEDI documentation that was posted, than ICBI.
http://www.eurasip.org/Proceedings/Eusipco/Eusipco2009/contents/papers/1569192778.pdf

:D

[diffid]

I'm probably going to get burnt here for asking and not having read the last 88 pages of the thread but can MadVR be used to output still image sequences? For example YV12 mpeg2's .mts input to 16/32bit full float linear exr's ready for compositing apps like Nuke or After Effects?

[chuuey]

ouh, IGP's can be nasty sometimes...shared memory and stuff

i'm sure, but this one handles madvr 0.09 and cuda enabled without any problems in 24hz

[Egh]

I recall that just a few years ago it was also "computationally impossible" to encode 1080p material in real-time on commodity hardware.
;)

On CPU it is still impossible. Especially when thinking "highest quality" :)

Besides, in several years we can have better algorithms, probably. So if real time resizing cannot be achieved in the next couple of years, which judging by the figures provided in the article, it is pretty much given, thus it is pointless to implement it.

[DigitalLF]

i found something in the nvidia control panal that makes me think.... "Maximum pre-rendered frames" and the default value is "3"

i think i did read awhile ago about madvr uses "8" can madvr override the nvidia value or do i need to change it to 8?

[Kazuya]

No news about the next release ?
I really miss the subpicture pin ! :(

[Weirdo]

madVR keeps crashing with my Nvidia 9600XT (1GB), on Vista and 7. No problem on XP. Could this be due to low system RAM (also 1GB)?

[flanger216]

I'm also having crashes... madVR used to work great with my Radeon 3650, but since upgrading to a Geforce GT220, I get an instantaneous crash upon loading any file in MPC-HC.

[Weirdo]

This could also be Nvidia driver/Cuda-related. On XP, sometimes madVR will not work with Cuda on, others it will. Hopefully CoreAVC 2.0 will solve some of these problems. The previous Nvidia driver (191.07) seems more stable to me.

[Weirdo]

Cuda seems to be the annoyance. I can now only get it to work randomly, say 1 out of 10 times with madVR on XP, Vista or 7. Can't figure out when and why it works.
With CUDA disabled, there are no freezes/crashes and madVR works fine.

[Seb.26]

I can now only get it to work randomly, say 1 out of 10 times with madVR on XP, Vista or 7.
With the same media ??? :confused:

[JarrettH]

Do I just set ffdshow to output YV12 to work correctly with this? I still leave haali splitter installed for mkv files?

[adam777]

Do I just set ffdshow to output YV12 to work correctly with this? I still leave haali splitter installed for mkv files?

Basically, yes.
madVR will simply not connect if the decoder isn't outputing YV12.
The splitter is doing its thing before the decoder, so you still need it.

[JarrettH]

Ohhh ok I just reinstalled the splitter :devil:

does madvr know whether the video is SD or HD? like in haali it would use different colourspace coefficients if the video were >1024 pixels

thanks:cool:

[adam777]

I ain't much of an expert for sure, but while playing the movie you can click on the madVR filter to see its options.
Personally, I see that madVR generated two different 3dlut files for me (HD - PC and SD - PC), so I'll have to assume it's doing its job correctly.

[yesgrey]

does madvr know whether the video is SD or HD? like in haali it would use different colourspace coefficients if the video were >1024 pixels
Yes..

[Weirdo]

Cuda seems to be the annoyance. I can now only get it to work randomly, say 1 out of 10 times with madVR on XP, Vista or 7. Can't figure out when and why it works.
With CUDA disabled, there are no freezes/crashes and madVR works fine.With the same media ??? :confused:Yes...

[namaiki]

I'm just a bit curious.. Could someone with a >256MB graphics card download the latest GPU-Z(0.3.8) and check GPU memory usage when running MadVR @ 1920x1080?

My Video Card: GeForce 9600m GT 256MB DDR3
OS: W7 64-bit
Running latest CCCP + MadVR
Video: various H,264, but resolution 1280x720

GPU memory usage when not using CoreAVC's CUDA thing
1280x720: 110MB (running only Vista, usage is ~60MB (for Aero, I guess))
1920x1080: 235MB

With Cuda Acceleration from CoreAVC
1280x720: 168MB
1920x1080: 235MB??? (probably not enough VRAM..)

At 1280x720, I can use CUDA with no lag, but If I stretch the video to higher than that, there will be major lagging (for video).

GPU load is ~36% at 1280x720 and ~68% 1920x1080 (both with CUDA enabled).

Memory controller load is max 18% for both resolutions.

Video engine is ~20%

by the way, I don't usually use coreavc.. I'm just curious. :V

[markanini]

I have 9800gt low power edition and i will check tomorrow now i'm off to bed.

[ManiacDC]

Does anyone know why I can't select YV12 as output in FFDShow? It's grayed out (both options under Planar are grayed out actually).

Win7 x64
ffdshow tryouts 3119

Thanks

[mark0077]

They will be blanked out I think if you choose an option below to output interlaced flag or maybe another flag. Should only take a few seconds for you to figure out which option disables them.

[ManiacDC]

They will be blanked out I think if you choose an option below to output interlaced flag or maybe another flag. Should only take a few seconds for you to figure out which option disables them.

ah-ha, found it, thanks!

[Weirdo]

Correct me if I'm wrong, madVR has tearing problems at 24Hz? Searched a bit, and seems to be the case. Tried it on a 1080p display and it is unwatchable. No problems on 50/60Hz.
Tried switching to 48Hz with Powerstrip (not natively supported by the TV), less tearing but still not good.
ReClock time-stretching to 25fps (50Hz) solves the issue but at the cost of sound quality, the critical centre channel suffering the most.

[leeperry]

timestretching is terrible in Reclock, avoid at all costs..yes many ppl did report tearing at 24Hz.

humm a few days ago, I thought it'd be great to disable "PCI-E Spread Spectrum" in my P5k Premium BIOS...eeeek, it made Reclock completely nuts.

[tmbt]

Hi guys,
i got a problem using madVR. When i use it at full screen the movie is not fluid at all. This only happen in Xp 64bit. The same pc using windows 7 32bit does not give me any errors. Someone could point me to the solution ??
Thx