madVR - high quality video renderer (GPU assisted)

[Zachs]

Originally Posted by x7007

Hey, how can I use the OpenCL NNEDI3 ? with CUVID or DXVA2 write-back ?
Can you explain .
Thanks

My reply was specifically for madshi only. I have no idea how to answer your question for madVR as I don't use it. If you mean to ask how you can use OpenCL NNEDI3 with MPDN, kindly post in the MPDN thread instead.

Cheers.

[madshi]

Originally Posted by Zachs

Well FWIW on NVidia at least, anything less than 128 neurons (i.e. 64 and below), the SM5.0 NNEDI3 trumps the OpenCL version using avoid branches optimization. On Intel with P4600 (which is still very modern), the SM5.0 version is ~2x faster than the OpenCL version across the board with the "Alt weight access + scalar small code" optimization option. For example, upscaling 720x576 to 1080p the P4600 renders in 31ms (16/16 neurons) making it perfectly acceptable to watch anything 30fps and below. With the OpenCL version, it needs 58ms. At the end of the day it really comes down to how well the driver/compiler maps the code to its underlying hardware.

From reading the MPDN thread my impression was that older NVidia GPUs were roughly the same with SM5.0 and OpenCL and that newer NVidia GPUs (Maxwell) still performed faster with OpenCL. Has this changed with recent drivers? I'm aware of that Intel likes SM5.0 much more. Thanks...

[Sunset1982]

@xAURUS:

Originally Posted by Xaurus

Ok so I tried to figure out the mysterious dropped frames in Win 10.

Tried to replace madvr with EVR, no difference. Tried moving back to 32-bit mpc-hc/lav/madvr. No difference.
Tried using ReClock. No difference. Went back to madvr 88.16. No difference.

The only thing that actually works is to restart the computer, then typically the first movie I watch has no mysterious dropped frames.
After that, it's the same thing.

So I guess this must be something with Win 10. In either case, it's extremely annoying.

did you try the latest insider build 10525 of Win10? maybe this works better?

[aufkrawall]

128 neurons NNEDI3 quadrupling on GTX 980 with 355.60
OpenCL (seems to be working fine with Zachs' new implementation):
GPU load: ~53%
clock: 1421Mhz
power consumption: ~72%
rendering time: ~25.6ms

SM 5.0 Avoid Branches:
GPU load: ~63%
clock: 1421Mhz
power consumption: ~67%
rendering time: ~31.8ms

Prefer Vector:
GPU load: ~80%
clock: 1421Mhz
power consumption: ~80%
rendering time: ~39.5ms

100% power consumption would be ~180 Watts.

[madshi]

So OpenCL does seem to be still faster with Maxwell GPUs, right?

Since OpenCL vs. SM5.0 is like compute power vs shader power (so to say), I suppose that the more the GPU architecture is optimized for compute performance, the higher the probability that OpenCL will be faster than SM5.0. Which also means that newer GPU architectures are likely to favor OpenCL. But then, a lot depends on compilers, too. And Intel seems to be quite different, in any case.

[aufkrawall]

Originally Posted by madshi

So OpenCL does seem to be still faster with Maxwell GPUs, right?

Yes. However, the advantage naturally shrinks when using less neurons and further effects like AdpativeSharpen or SuperRes. So, in many situations there probably won't be a difference "perceived".

Now questions is, does the way Zachs found for OpenCL look secure for the future?

[Arm3nian]

Can someone explain why d3d9 has to be used? Why not 10, 11 or 12? Does dxva not support those or something? Isn't this microsoft's fault for not updating their stuff instead of nvidia removing backward compatibility or incompatibility in general.

Also, I don't think new gpu's like the 980ti or 950 even have drivers, on any operating system, that support d3d9 interop.

[Xaurus]

Originally Posted by Sunset1982

@xAURUS:



did you try the latest insider build 10525 of Win10? maybe this works better?

I did not try that yet. But I am surprised, if this is a Win 10 issue there should be much more info on this already. But I can't find much.

What I will do today is to downgrade the Nvidia drivers to the version that can use NNEDI3.

Right now, my HTPC feels useless. Can't watch anything without pulling my hair out because of these mysterious dropped frames.

[kalston]

Originally Posted by Arm3nian

Also, I don't think new gpu's like the 980ti or 950 even have drivers, on any operating system, that support d3d9 interop.

Well my 980 ti does NNEDI3 256 neurons for 720p to 1440p with ~25ms render times on my Win 7 64 machine so I'm pretty sure that's incorrect.

[Zachs]

Originally Posted by madshi

So OpenCL does seem to be still faster with Maxwell GPUs, right?

Since OpenCL vs. SM5.0 is like compute power vs shader power (so to say), I suppose that the more the GPU architecture is optimized for compute performance, the higher the probability that OpenCL will be faster than SM5.0. Which also means that newer GPU architectures are likely to favor OpenCL. But then, a lot depends on compilers, too. And Intel seems to be quite different, in any case.

Well it would seem Maxwell is the odd one out - the other simpler explanation could be that they simply haven't optimised the driver as much as they have for the more mature architectures.

[madshi]

Well, it's Maxwell and AMD. So basically those who're strong at computing, no? But I don't really know, I'm just guessing...

[nevcairiel]

Originally Posted by Zachs

Well it would seem Maxwell is the odd one out - the other simpler explanation could be that they simply haven't optimised the driver as much as they have for the more mature architectures.

If by "odd one out" you mean OpenCL is stronger, then thats expected as NVIDIA changed the architecture in Maxwell to cover a lot of ground in computing.

[Zachs]

By "odd one out" I mean it is different to Fermi and Kepler. Fermi was strong in computing too.
But, is shader not computing? Is it not possible for Nvidia to map the shader version to the OpenCL version in its entirety in the future with driver optimizations?
We don't know any of that - all we're doing here is pure speculation, and a bit of a waste of time to be honest since there's no way to find out, or tell the future.

[Asmodian]

Originally Posted by Arm3nian

Also, I don't think new gpu's like the 980ti or 950 even have drivers, on any operating system, that support d3d9 interop.

Using 853.30 or .49 d3d9 interop worked fine on my 980ti in Windows 10.

[amayra]

i manage to make madvr work in Wine 1.7.50 somehow but is crash later i will post log after they release "CSMT" and "staging" edition

[Akeno]

This has been spanning on for the past couple of pages but can anyone give a short run down on what problems madVR has with windows 10?
So far, it sounds like nnedi3 still doesn't work on nvidia cards due to missing opencl implementation and something about d3d9 interop although I don't know what effects that has on madVR.

[Arm3nian]

Originally Posted by Asmodian

Using 853.30 or .49 d3d9 interop worked fine on my 980ti in Windows 10.

Thanks worked for me also. Now that I've tried it, super-xbr looks identical to nnedi3 in chroma upscaling and image chroma and luma doubling watching bluray on a 4k screen. The only difference is that super-xbr doesn't turn my room into an oven.

[noee]

Originally Posted by amayra

i manage to make madvr work in Wine 1.7.50 somehow but is crash later i will post log after they release "CSMT" and "staging" edition

Start another thread, maybe in the Linux section....I'm trying the same thing....

[Fadeout]

Quick question: does anyone know why MadVR deinterlacing looks much better than Yadif on VLC and other applications?

Windows Media Player gives a similar result to MadVR, it's only Yadif that makes things less smooth. And this is a problem because I'm trying to encode the video to progressive, but the only options for the decoding only produce results that look like Yadif, and so everything is less smooth.

[nussman]

madVR can use GPU Deinterlacing like WMP does.