madVR - high quality video renderer (GPU assisted)

[nevcairiel]

Originally Posted by TheElix

Somebody should bring it to their attention, dammit!

Its not like its a small "bug", their overall design by choosing OpenCL is severly limited, its doubtful that they'll design a new interface just like that. You can do the same with DXVA2, including deinterlacing, even without a renderer - except that its damn slow for ATI/AMD cards because of the limited GPU->CPU transfer speed.

I don't know if its a driver or a hardware problem, but i'm quite sure it affects all GPU->CPU transfers. AMD/ATI isn't all that famous for their HPC/GPGPU capabilities.
If they want to fix something, they should work on the transfer speed issue, and make using DXVA2-renderless support a viable alternative.

[pankov]

Originally Posted by madshi

How often do you change background wallpapers during video playback?

I believe that he is not changing them manually but letting Windows 7 do it automatically (like I did before). There is such an option in the Desktop background configuration window.
Sadly it interferes with madVR even in Exclusive mode so I disabled it ... it was a nice feature otherwise.

madshi,
in the last few days I've noticed some random dropping of frames without any of the queues dropping low. They moved from 7-8/8 to 6/8 but not lower and despite this I got 5-9 dropped frames maybe in series of 3 for 2-3 seconds.
Sadly the problem is random and not reproducible

I haven't change any settings in past few months and I'm pretty sure this problem didn't exist in v0.73.

I'll try to confirm if it's only v0.74 that has it but does anybody else experiencing the same and have any idea what could be wrong?

[betaking]

To madshi
If madvr.ax where the folder does not contain libmfxsw32.dll avcodec-53.mvr avutil-51.mvr these three files, then do not enable internal decoder! or put this seeting to Registry not to bin file! thanks!

[madshi]

Originally Posted by TheElix

Somebody should bring it to their attention, dammit!

I tried.

Originally Posted by nevcairiel

Its not like its a small "bug", their overall design by choosing OpenCL is severly limited, its doubtful that they'll design a new interface just like that. You can do the same with DXVA2, including deinterlacing, even without a renderer - except that its damn slow for ATI/AMD cards because of the limited GPU->CPU transfer speed.

I don't know if its a driver or a hardware problem, but i'm quite sure it affects all GPU->CPU transfers. AMD/ATI isn't all that famous for their HPC/GPGPU capabilities.
If they want to fix something, they should work on the transfer speed issue, and make using DXVA2-renderless support a viable alternative.

It seems that AMD wants to go after GPGPU with their next generation. Maybe there's hope they'll have improved their transfer speeds then.

If only they implemented OpenCL <-> D3D9 interop. That would make a lot of things easier. And it should be possible with the current hardware, too. Some earlier version of their SDK actually had limited OpenCL <-> D3D9 interop, but then they suddenly dropped it, without any explanations.

Originally Posted by pankov

in the last few days I've noticed some random dropping of frames without any of the queues dropping low. They moved from 7-8/8 to 6/8 but not lower and despite this I got 5-9 dropped frames maybe in series of 3 for 2-3 seconds.
Sadly the problem is random and not reproducible

I haven't change any settings in past few months and I'm pretty sure this problem didn't exist in v0.73.

I'll try to confirm if it's only v0.74 that has it but does anybody else experiencing the same and have any idea what could be wrong?

I doubt it could be v0.74. Try creating a log. The log contains the exact reason for why the frames were dropped.

Originally Posted by betaking

If madvr.ax where the folder does not contain libmfxsw32.dll avcodec-53.mvr avutil-51.mvr these three files, then do not enable internal decoder! or put this seeting to Registry not to bin file! thanks!

Ah, I understand! Will do that...

[TheElix]

Originally Posted by nevcairiel

Its not like its a small "bug", their overall design by choosing OpenCL is severly limited, its doubtful that they'll design a new interface just like that. You can do the same with DXVA2, including deinterlacing, even without a renderer - except that its damn slow for ATI/AMD cards because of the limited GPU->CPU transfer speed.

I don't know if its a driver or a hardware problem, but i'm quite sure it affects all GPU->CPU transfers. AMD/ATI isn't all that famous for their HPC/GPGPU capabilities.
If they want to fix something, they should work on the transfer speed issue, and make using DXVA2-renderless support a viable alternative.

Oh, if only you were in AMD development team

[Thunderbolt8]

I'd guess at least it wouldnt getting any worse then

[ajp_anton]

Originally Posted by madshi

Depends on the algorithm. When using linear resamplers (as offered by madVR) the original pixels will not be there. Linear resamplers resample every single pixel. There are algorithms which do integer multiple upscales and leave the original pixels untouched and just interpolate the "missing" pixels. Examples for such algorithms are Nearest Neighbor, NEDI, NNEDI3, ICBI etc...

Well, in Avisynth for example you can shift by .25 pixels while resizing in order to preserve the original pixels, which IMO results in better quality.

[madshi]

Originally Posted by ajp_anton

Well, in Avisynth for example you can shift by .25 pixels while resizing in order to preserve the original pixels, which IMO results in better quality.

That shift trick would work only with Bilinear sampling, but not with any other linear filter (Bicubic, Lanczos, Spline etc). And I think Lanczos3 with no shift will still look noticeably better than Bilinear with a 0.25 shift. Preserving the original pixels is very much overrated, IMHO. But if you think you can improve on madVR scaling quality with a simple Avisynth linear resampler with a 0.25 shift, then post some comparison screenshots that proof that and I'll consider implementing it.

[pirlouy]

Originally Posted by nevcairiel

AMD/ATI isn't all that famous for their HPC/GPGPU capabilities.

Neither nVidia.
Your GPU decoder is surely nice, but nVidia drivers also have lots of problems.

[nevcairiel]

Originally Posted by pirlouy

Neither nVidia.
Your GPU decoder is surely nice, but nVidia drivers also have lots of problems.

Eh? That comment wasn't referring to consumers and their issues, or video decoding at all.
NVIDIA GPUs are used for HPC (High-performance computing) and GPGPU tasks all over the world. They even have a dedicated card just for that (the Tesla).
The key point was that GPGPU tasks typically also need good performance of GPU->CPU transfers, similar to how a CUDA decoder works, or a DXVA2-renderless decoder would - but its nothign you need for "normal" video playback or games, usually you just sent it to the GPU for display, you don't want to get it back.

Between AMD/ATI and NVIDIA, thats a clear win for NVIDIA. Which is why AMD trys to push into that world with their next generation, but IMHO they are too late. CUDA is basically an industry standard by now, and OpenCL is only very slowly catching up (and is handy-capped to some degree)

PS:
Regarding driver issues - why is it that i always read on eg. AVSForums about the latest Catalyst breaking someones playback chain, but very rarely about NVIDIA drivers doing something like that?

Sure, the drivers are not perfect (which software is?), and i could name several key points that need fixing myself, but at least there is some progress without huge regressions.

[STaRGaZeR]

Originally Posted by nevcairiel

PS:
Regarding driver issues - why is it that i always read on eg. AVSForums about the latest Catalyst breaking someones playback chain, but very rarely about NVIDIA drivers doing something like that?

Because you don't read enough. Or because you don't have enough first hand experience with both camps

[madshi]

There is a *HUGE* regression with newer NVidia drivers, though: Presentation glitches when playing back content at a higher refresh rate. E.g. playing back 25fps content at 50Hz is really bad here on my 9400. I have to switch to D3D11 presentation and play with the tweaks to make the presentation glitches go away. From reading through this thread it appears that AMD users don't have this problem.

[nevcairiel]

Originally Posted by madshi

There is a *HUGE* regression with newer NVidia drivers, though: Presentation glitches when playing back content at a higher refresh rate. E.g. playing back 25fps content at 50Hz is really bad here on my 9400. I have to switch to D3D11 presentation and play with the tweaks to make the presentation glitches go away. From reading through this thread it appears that AMD users don't have this problem.

I also get the glitches in the OSD, but i don't see them (video remains smooth for me). I did however apply some tweaks to get rid of them anyway.
Do you actually get non-smooth video?

Speaking of DX11, i noticed the desktop/mpc-hc is not properly re-drawn if i exit FSE mode directly to windowed mode via hotkey (or double-click in MPC-HC). It works properly if i bring up a context menu in full-screen first to exit exclusive mode, and then double click to get into windowed mode.
Works all fine with DX9.

PS:
I don't consider that huge, huge would be breaking playback completely, or making any kind of material unwatchable, without a screw i can turn to simply fix it.

[madshi]

Originally Posted by nevcairiel

I also get the glitches in the OSD, but i don't see them (video remains smooth for me). I did however apply some tweaks to get rid of them anyway.
Do you actually get non-smooth video?

In DX9, or without tweaks: Yes. With DX11 and carefully tweaked options I can get playback smooth, though.

Originally Posted by nevcairiel

Speaking of DX11, i noticed the desktop/mpc-hc is not properly re-drawn if i exit FSE mode directly to windowed mode via hotkey (or double-click in MPC-HC). It works properly if i bring up a context menu in full-screen first to exit exclusive mode, and then double click to get into windowed mode.
Works all fine with DX9.

Yeah, DX11 is a bit unstable, will have to work on that sooner or later.

Originally Posted by nevcairiel

I don't consider that huge, huge would be breaking playback completely, or making any kind of material unwatchable, without a screw i can turn to simply fix it.

Well, you have to consider how much time I spent to implement all those nasty tweaks to work around the problem. It was probably around 2-3 full weeks of my programming time, no exaggeration, just to work around NVidia driver problems. And even through the tweaks do seem to help, users being forced to play around with those tweaks is very ugly, especially because it seems that everybody needs different tweaks. Personally, I do consider this problem "huge". In theory exclusive mode presentation was supposed to be rock stable. In real life it's quite the opposite, with NVidia at least.

[nevcairiel]

Originally Posted by madshi

Well, you have to consider how much time I spent to implement all those nasty tweaks to work around the problem. It was probably around 2-3 full weeks of my programming time, no exaggeration, just to work around NVidia driver problems. And even through the tweaks do seem to help, users being forced to play around with those tweaks is very ugly, especially because it seems that everybody needs different tweaks. Personally, I do consider this problem "huge". In theory exclusive mode presentation was supposed to be rock stable. In real life it's quite the opposite, with NVidia at least.

Wait, didn't this regression appear quite recently (270 series drivers or such?), well after you implemented all those options?
Before those drivers, i didn't really need any tweaks, default settings worked quite smooth.

[madshi]

Originally Posted by nevcairiel

Wait, didn't this regression appear quite recently (270 series drivers or such?), well after you implemented all those options?
Before those drivers, i didn't really need any tweaks, default settings worked quite smooth.

I'm not sure with which driver version exactly the issue was introduced. When I first implemented the "final exclusive mode rendering path" several months ago, some people already had problems and some people did not. Of course different people were using very different driver versions. Personally, I was using the latest CUDA beta driver versions. Anyway, my general impression was that AMD users had no big trouble with presentation glitches, while NVidia users had. That was a few months ago. Maybe the situation got even worse with 270 drivers now? Or maybe those months ago some people were already using very new beta drivers? I don't really know for sure...

[Qaq]

Its just because AMD users are "scared to death" to change the drivers. I am at least.

[robpdotcom]

Originally Posted by madshi

First of all you should leave your GPU set to RGB full in either case. Furthermore, of course madVR's level setting must match the display's level setting. If these two requirements are met, image quality should be just fine. Now should you use video levels or PC levels? That's a good question. There's an argument for both. If you use video levels, the processing is somewhat less extensive, so there's an ever so slight chance that the output might be smoother. Thanks to dithering this should not really make any visible difference in the real world though. There's also a good argument for using PC levels: Using PC levels allows you to have correct blacks and whites for video playback, gaming and photos/applications at the same time. So pick your poison. Personally, I'm currently using PC levels.

Thanks for the clarification, madshi.

[dansrfe]

Originally Posted by pankov

I believe that he is not changing them manually but letting Windows 7 do it automatically (like I did before). There is such an option in the Desktop background configuration window.
Sadly it interferes with madVR even in Exclusive mode so I disabled it ... it was a nice feature otherwise.

Yeah it's automatic every 1 minute for me.

[mzso]

Originally Posted by madshi

The AMD API has 2 big problems:

(1) It doesn't support any kind of deinterlacing.
(2) It requires the use of OpenCL, and AMD doesn't support OpenCL <-> D3D9 interop.

Add to that the fact that DXVA deinterlacing is based on D3D9. The end result is that using the AMD API makes no sense at all for interlaced content. I don't really know what AMD is thinking. The missing OpenCL <-> D3D9 interop is a major problem, but AMD doesn't seem to care. They don't seem to know what they're doing, IMHO.

Layman question. Isn't OpenCL designed to interoperate with OpenGL?