madVR - high quality video renderer (GPU assisted)

[sheppaul]

According to this post, intel has dedicated hardware scalers and provides the best scaling similar to lanczos and ar, while amd and nvidia are using bi-linear and shaders.

http://www.avsforum.com/forum/26-hom...l#post23847902

Hmm, interesting. BTW, I can confirm that dxva2 scaling on AMD is not certainly bilinear.
A sample to verify: http://forum.doom9.org/showthread.ph...60#post1705360

Bilinear


DXVA2

[huhn]

Originally Posted by sheppaul

Hmm, interesting. BTW, I can confirm that dxva2 scaling on AMD is not certainly bilinear.
A sample to verify: http://forum.doom9.org/showthread.ph...60#post1705360

on my AMD 270 DXVA scaling is clearly bilinear.

[sheppaul]

Originally Posted by huhn

on my AMD 270 DXVA scaling is clearly bilinear.

This is pretty strange. Change a video renderer to EVR default or try windows media player which is using EVR by default.

I'm using "14.12 AMD Catalyst Omega driver" with HD7850.

[huhn]

Originally Posted by sheppaul

This is pretty strange. Change a video renderer to EVR default or try windows media player which is using EVR by default.

I'm using "14.12 AMD Catalyst Omega driver" with HD7850.

EVR is not using bilinear.

so there are 2 possibilities:

madVR has a bug and bilinear is used when DXVA scaling is used. or EVR is not using DXVA scaling.

edit: looks like it is a bug in madVR. I did a small test look here: http://abload.de/img/evrsplita9ubm.png
this is EVR with split demo mode.
would be nice if someone can check madVR again with this sample and see if it is bilinear for you too. after that I will create a bug entry in the bug tracker.

[sheppaul]

Originally Posted by huhn

EVR is not using bilinear.

so there are 2 possibilities:

madVR has a bug and bilinear is used when DXVA scaling is used. or EVR is not using DXVA scaling.

Are you sure? EVR default seems to use dxva2 scaling while EVR C/P uses bilinear by default. I have a same result with potplayer supporting dxva2 scaling now. Would you open the sample with widnows media player and let me know what it looks like? WMP uses a vanilla EVR and it looks like this in my screen which is obviously not a bilinear and it's the same as dxva2 scaling.

ps. I don't use any post-processing filter in catalyst control center.

[huhn]

can you please show what your results with madVr if madVR uses biliniear for you too it is very clear a bug in madVR.

both EVR in MPC-HC and windows media player don't use bilinear for me but madVR with dxva does.

[michkrol]

Originally Posted by huhn

edit: looks like it is a bug in madVR. I did a small test look here: http://abload.de/img/evrsplita9ubm.png
this is EVR with split demo mode.
would be nice if someone can check madVR again with this sample and see if it is bilinear for you too. after that I will create a bug entry in the bug tracker.

To my understanding, what your image shows is:
on the left DXVA scaling with video enhancements off,
on the right DXVA scaling with video enhancements on.

The "bug" is probably as simple as madVR disabling video enhancements on your system, but not for some other people and that's the reason you get different results - different driver's settings.

[huhn]

Originally Posted by michkrol

To my understanding, what your image shows is:
on the left DXVA scaling with video enhancements off,
on the right DXVA scaling with video enhancements on.

The "bug" is probably as simple as madVR disabling video enhancements on your system, but not for some other people and that's the reason you get different results - different driver's settings.

the test was made to show that AMD DXVA2 with "enhancements off" is bilinear.

so you didn't get bilinear with DXVA in madVR using a AMD card? this would mean it is a bug my side which i'm totally fine with.

[sheppaul]

Originally Posted by huhn

can you please show what your results with madVr if madVR uses biliniear for you too it is very clear a bug in madVR.

both EVR in MPC-HC and windows media player don't use bilinear for me but madVR with dxva does.

I've already posted the result with madVR in the two screenshots above. It's pretty strange I have a same result with MPC-HC/BC/WMP.

The "bug" is probably as simple as madVR disabling video enhancements on your system, but not for some other people and that's the reason you get different results - different driver's settings.

I have all the video enhancement settings disabled. I don't think madVR did something different here. The only difference I can see is I'm using Windows 8.1.

so you didn't get bilinear with DXVA in madVR using a AMD card? this would mean it is a bug my side which i'm totally fine with.

If then this is not a problem of madVR as I can get the same result with other players. Pretty strange.

[tjcinnamon]

Originally Posted by leonccyiu

Since Chrome and Firefox are based on open source code projects, is it possible that in future, madvr can be implemented as an option as a video renderer since most video is now delivered using html5 and rendered by the browser now.

Right now, almost all web video is scaled using billinear and does not take advantage to dxva scaling, for those with high res monitors, such as even when watching 4k youtube on the 5k imac, it makes the picture very soft and blurred.

This would allow web video such as vimeo, youtube, and any page with html5 video to be rendered with madvr for much better quality.

An obstacle I see though is madvr is 32bit and chrome is moving onto 64bit

YES YES YES!!!! THIS ^^^

I have UFC fight pass and the "HD" stream is "720p" (which I'd be okay with) if there weren't so much aliasing (jaggies). It's crazy. I started playing around with the HTML to see if I could stream it through to MPC-HC (and madvr).

Does such a product exist to push a videostream to MPC-HC? Perhaps that would be a way to leverage mad-vr to improve most of the highly compressed web-content

[huhn]

Originally Posted by tjcinnamon

YES YES YES!!!! THIS ^^^

I have UFC fight pass and the "HD" stream is "720p" (which I'd be okay with) if there weren't so much aliasing (jaggies). It's crazy. I started playing around with the HTML to see if I could stream it through to MPC-HC (and madvr).

Does such a product exist to push a videostream to MPC-HC? Perhaps that would be a way to leverage mad-vr to improve most of the highly compressed web-content

livestreamer can do this.

of course doesn't work with all pages.

[detmek]

Originally Posted by sheppaul

Hmm, interesting. BTW, I can confirm that dxva2 scaling on AMD is not certainly bilinear.
A sample to verify: http://forum.doom9.org/showthread.ph...60#post1705360

I have done a test with your file on Intel Pentium G3220. For some reason DXVA scalling won't activate in madVR if output from decoder is RGB and it defaults to Bilinear. I had to disable RGB output in LAV to test DXVA scalling. Even with that I had a problem - green color lines, right and at the bottom. And EVR has same problem with disabled RGB output. Here is what I got:
EVR: DXVA: LanczosAR: Bilinear: EVR-disabled RGB:

I am not sure what this means

[sheppaul]

Look at the screen closely and you'll know that there is a pixel shift to left upper in EVR (DXVA scaling). It looks like a genuine bug of EVR and it seems to cause the problem with the dxva scaling of madVR. There is a same shift in EVR and DXVA2 resizer. That's the main reason why I consider both are same.

BTW, the screenshots are from iGPU? EVR and DXVA screenshots look different. I'm not sure it's because both are using different resizers or additional processing of madVR affects the result. Would you try the sample with potplayer? I get the exactly same result with EVR and DXVA2 scaling with HD7850 at least.

[ikakun]

On my GPU, EVR-CP(set to bilinear) & madvr DXVA2 scaling looks the same. Plain EVR isn't.

BTW, using NNEDI3 on that test sample, doubling @ 16 neurons + quadrupling @ 16 neurons + jinc3ar image scaling give me better result than increasing the neurons further with or without quadrupling. And here I thought higher NNEDI3 neurons is better. Or is it safe to say that it depends on what source video you're playing? Like anime & live(or should I say real life?) video for example.

[ryrynz]

Originally Posted by ikakun

Or is it safe to say that it depends on what source video you're playing? Like anime & live(or should I say real life?) video for example.

Yes, that's it exactly.

[detmek]

Originally Posted by sheppaul

Look at the screen closely and you'll know that there is a pixel shift to left upper in EVR (DXVA scaling). It looks like a genuine bug of EVR and it seems to cause the problem with the dxva scaling of madVR. There is a same shift in EVR and DXVA2 resizer. That's the main reason why I consider both are same.

BTW, the screenshots are from iGPU? EVR and DXVA screenshots look different. I'm not sure it's because both are using different resizers or additional processing of madVR affects the result. Would you try the sample with potplayer? I get the exactly same result with EVR and DXVA2 scaling with HD7850 at least.

Yes, I noticed the shift myself but there is a difference between EVR and madVR DXVA scalling as EVR is a bit blurry. Also, on RGB output DXVA scalling won't work for madVR and EVR. Both renders default to Bilinear.

It's iGPU, I don't have discrete GPU.

Here are screenshots with PotP. I used default settings, only changing renders.
EVR: madVR DXVA: madVR Bilinear:

EVR does use DXVA scalling as I got same bug and same difference in sharpness. Maybe a driver problem.

[sheppaul]

on RGB output DXVA scalling won't work for madVR and EVR. Both renders default to Bilinear.

DXVA2 Scaler seems not working with RGB output. It is understandable as DXVA2 decoder also doesn't work with RGB output.

[madshi]

Originally Posted by sheppaul

DXVA2 Scaler seems not working with RGB output. It is understandable as DXVA2 decoder also doesn't work with RGB output.

That is correct for AMD, but not for Intel. I noticed this problem with AMD about a week ago. It seems I'll have to go back to scaling NV12 -> NV12, when using DXVA. A pity, really.

[detmek]

madshi, then why DXVA scaler won't work with RGB on my Intel GT1 GPU (Pentium G3220)?

[Asmodian]

Originally Posted by madshi

DXVA2 decoding and deinterlacing outputs NV12 surfaces. Unfortunately pixel shaders can't use them directly. So madVR has to convert the NV12 surfaces somehow to make them pixel shader compatible. There are 3 different ways madVR can use to do that. Solution (1) is copyback (downloading the NV12 data to CPU RAM, then re-uploading it to GPU in a different format). Solution (2) is conversion/processing via OpenCL, introduced in v0.87.0. And solution (3) is a copy operation on the GPU, which can be lossless with AMD GPUs, but produces a slightly blurred chroma channel with NVidia and Intel GPUs. There's a trade-quality-for-performance option to switch between (1) and (3). And there's a new option in the "rendering -> general settings" to enable/disable (2).

Does this mean that AMD users should enable the trade quality for performance options "don't use 'copyback' for DXVA deinterlacing (Intel, NVidia)" and "don't use 'copyback' for DXVA decoding (Intel, NVidia)"? And the reason it doesn't mention AMD is that no quality is lost?