madVR - high quality video renderer (GPU assisted)

[ryrynz]

Originally Posted by Hyllian

The algorithm is adjustable and I think you all video experts can extract much more from this than me in that task.

Exactly. I have no doubt madshi will find some acceptable settings and or allow us to fine tune it based on feedback.

[madshi]

Originally Posted by Hyllian

They were created with the three passes, though the third pass is optional.

I think the third pass improves a bit the IQ, in exchange of power, of course. I don't know if it's worth. Anyway, it's optional and can be skipped if power is a problem.

As you'll see, my implementation has much room for optimization. I calc the luminance inside all shaders and it could be done in a first pass (see NEDI shaders, for example) only.

The first pass must work at 2x scale factor (the same way NEDI shaders), and the second and third passes work at 1x scale factor. Use nearest sampling on all texture lookups, as it doesn't need precise texture sample locations.

You'll see too that my blendings aren't done in linear light. I didn't test this algorithm in linear light.

Ok, thanks. I'll try linear light and with/without the 3rd pass.

Originally Posted by MysteryX

Personally I find the image unnatural. Too "cartoonish" for my taste.

If you upscale that much (400%), the image must become either very very soft, or a little bit cartoonish, that's just the way it is. NNEDI3 has a similar look.

Originally Posted by MysteryX

Simple good sharpening doesn't work with material that has noise and artifacts in it. SuperRes works GREAT to neutralize noise and artifacts while bringing out the real details. Even when using low-quality material with bilinear resizing, with SuperRes it still looks good. That can't be denied.

You can filter out noise and artifacts before sharpening. Yes, madVR can't do that yet. But that doesn't mean that SuperRes is superior in concept. Well, it *may* be, but I don't take that as a given yet.

Originally Posted by MysteryX

Hey I have an idea to simplify SuperRes. Instead of having all these settings and asking whether to apply once or every 2x scaling, why not simply ask how many passes he wants to run? If he runs 2 passes with a 3.5x upscaling factor, it will do 1 pass mid-way and 1 pass after. The user generally selects the number of passes based on what his graphic card can handle, but he really doesn't need to know when each pass actually gets applied. You can determine the best distribution of passes yourself.

If you scale an image 200%, you need much less sharpening than when you scale an image 400%. Your suggestion does not take that into account at all.

[Sunset1982]

Feature request:


@madshi:

are you planning the use of directx 12 for madvr in near future?


The use of internal GPU would be a nice fauture to think about:

http://winfuture.de/videos/Spiele/Di...men-14444.html

just a thing to think about. Keep on your great work! you give a great time watching movies on our htpcs!

[nevcairiel]

People really need to stop assuming that improvements for game rendering will automagically make video rendering better....
Personally, I don't think DX12 will offer anything for Video. Its not a replacement for DX11 afterall, its an alternative for developers that want absolute low-level control.

For multi-adapter, thats probably not viable for video rendering, as you can't off-load parts of the processing chain to another GPU. Its either the entire image, or nothing. And copying the entire image around is a slow process.

[dansrfe]

Is there even a significant performance advantage of DX11 over DX9 assuming an identical madVR setup? (no additional features)

[huhn]

AFAIK DX 11 is slower and dx9 old path is the fastest. but dx 12 is made for performance unlike 10/11. I'm pretty sure he will add DX 12 if it is gives better performance and is needed for a great new feature so just wait.

[aufkrawall]

It would give him quite a task for memory management etc. Not sure if this is worth it for a video renderer.

[pirlouy]

Ahah, Nevcairiel = dream breaker.

Everywhere they say Dx12 is a major improvement, threading , low level, etc. And you break everything.

[MysteryX]

Originally Posted by madshi

You can filter out noise and artifacts before sharpening. Yes, madVR can't do that yet. But that doesn't mean that SuperRes is superior in concept. Well, it *may* be, but I don't take that as a given yet.

I tried optimizing videos with denoising the image before upscaling to the best I could. It gives a good result, but SuperRes alone gives a result almost as good as all the processing I've done. If I could use somehow use SuperRes in AviSynth, then I could know for sure and compare. Denoising the image first and applying SuperRes at the end most likely would give an excellent result even on very bad material.

Originally Posted by madshi

If you scale an image 200%, you need much less sharpening than when you scale an image 400%. Your suggestion does not take that into account at all.

We need to configure profiles for each types of videos, so that's taken into account there.

[madshi]

http://www.anandtech.com/show/7889/m...s-to-directx/2

"As with Mantle, the primary goal for Direct3D 12 is to greatly reduce the CPU overhead [...]"

No benefit for madVR. There are many differences between what a game typically does and what madVR does. madVR just draws 2 triangles per frame. It's a whole different world. I don't see anything in DX12 that would be interesting to madVR. All the improvements I can use are already in DX11.

[aufkrawall]

Could someone please explain to me again what the reason was why NNEDI3 via DirectCompute was much slower than OpenCL?

[madshi]

Originally Posted by MysteryX

We need to configure profiles for each types of videos, so that's taken into account there.

So you want to simplify the SuperRes options, and in the same step require users to use profiles (to achieve a comparable sharpness for all scaling factors)? Sorry, but this would be a step in the wrong direction, if our goal is to achieve a simple and easy to understand madVR configuration at some point in the future. I don't want SuperRes to produce different levels of sharpness, depending on which scaling factor you're using. That would be confusing and difficult to setup for users.

Originally Posted by aufkrawall

Could someone please explain to me again what the reason was why NNEDI3 via DirectCompute was much slower than OpenCL?

I don't know, maybe the DirectCompute compiler didn't optimize the code as well as the OpenCL compiler did? Or maybe I was too stupid to write well optimized DirectCompute code? I tried to directly convert the OpenCL kernel to DirectCompute, and while it produced correct image output, it was very slow. Could very well have been my fault, for all I know.

[Siso]

An odd thing madvr causing big dpc latencies, I've tested with evr-cp - no latency problems

madvr


evr-cp

[iSunrise]

madshi, I found a bug with 0.88.11 that should be pretty easy to fix.

1) Go to image doubling
2) Enable double luma and chroma resolution
3) Use a scaling factor of at least 2.0x in your media player
4) Now, if you enable quadruple luma, the picture suddenly has a green tint over it. It goes away again if you either disable quadruple luma or if you additionally enable quadruple chroma.
5) As an additional note, it only happens when you have enabled "always - if upscaling is needed" on quadrupling, if you enable "only if scaling factor is XX" this does not happen. Seems like a corner case related to that "always ..." option that isn't handled correctly.

I accidentally found that out, since I enabled SuperRes after image doubling and then the green tint also goes away, irregardless of the doubling settings.

[omarank]

Originally Posted by madshi

Feedback

Questions:

1) Do you prefer linear light on or off?
2) In my own very short tests I found that FineSharp sometimes introduces aliasing artifacts. These seem to be mostly fixed by setting the "repair" option to rather high values. Personally, I've tried setting "repair" to 1.0, and liked the result. But what is your opinion about this? Do you find "repair" at 1.0 works for you? Or would you prefer it at a lower value?
3) Do you see a difference worth noting between the 3 different modes? Please note that these modes will make more of a difference if the sources have stronger grain. So in order to judge which modes work best and which worst, it might make sense to also test with a source with a lot of grain in it. FWIW, mode 3 is slower, modes 1 and 2 are faster. So if you like mode 3 best, but not much better than 1 and 2, then it would still be useful to know whether you prefer 1 over 2 or the other way round.
4) Which combinations of strength and thinning would you suggest for low/medium/high presets?

Thanks!

1. On
2. Yes “repair” at 1.0 looks good. I didn’t see it doing any harm to the images
3. I like mode 3 much better than mode 1 and 2. Between mode 1 and mode 2, I can’t decide which one is better
4. The “thinning” parameter makes so slight changes that are hard to make out. While I was playing with this parameter, I settled on 0.0032. Below are my preferences for the presets:
Low preset: “strength” 1.5, Medium preset: “strength” 2.0, High preset: “strength” 2.3

Whether FineSharp is selected in image enhancements or upscaling refinements, my preferences are the same.

[Neeto]

Originally Posted by Siso

An odd thing madvr causing big dpc latencies, I've tested with evr-cp - no latency problems

madvr


evr-cp

Have found that it's different on different motherboards.
Also BIOS update can make it better or worse.... SIGH

[Siso]

Originally Posted by Neeto

Have found that it's different on different motherboards.
Also BIOS update can make it better or worse.... SIGH

The bios update cut my latency in half, so I guess it helped, but the madvr is acting weird in that matter of latency...

[6233638]

Use LatencyMon to get a better idea of what's causing it.

[Siso]

Originally Posted by 6233638

Use LatencyMon to get a better idea of what's causing it.

This is my result with highest dpc latency while playing a mkv with madvr

[Braum]

Will it be possible, in the future, to share Madvr processing load between gpu and cpu ?

I got a 7870XT and a i5 3570K@4.5ghz and quickly reach the limits of my gpu, it's a bit frustrating to see my cpu doing practically nothing.