madVR - high quality video renderer (GPU assisted)

[billqs]

Originally Posted by Betroz

The GTX 1070 is a 6.46 TFLOPS card. An aftermarket 1080 or wait for the 1080Ti if you can afford it.

You're correct. I was misremembering. I guess I'm a little surprised that the 1070 wasn't doing more, but I am upscaling to 4k which is hugely resource-intensive.

[aufkrawall]

For this reason it makes sense to use less neurons for the quadrupling step or super-xbr instead of NNEDI3.
The doubling step is much more important than the quadrupling step, and with SuperRes in the end, the differences vanish further.
With increasing source resolution, the need for expensive scaling algorithms decreases. It's probably a pure waste of resources to use more than 32 neurons for 1080p sources.

[billqs]

Thanks, I was wondering about using an less intensive algorithm for the quadrupling. My main concern are 720p files. My 1080p files do a simple Image Doubling to 2160p with NNEDI3 and look fantastic. Its just harder with 720p since it doesn't fall into an easy multiple and so you double, quadruple then downscale. My SD has no trouble using high settings to 2160p but that doesn't mean that what comes out is a masterpiece. At some point garbage in = garbage out.

[thighhighs]

billqs
gtx1070 should be good for 4k. What madvr settings you use? If you like nnedi, use the double\quadruple luma only 32\64 + SuperRes in upscaling refinement section, no deringing, orderer dithering. And create the fps-based profiles with alternative settings for >29fps sources (if you need it). Read warner guide for more http://forum.kodi.tv/showthread.php?tid=259188

[Georgel]

Originally Posted by ryrynz

Perfectly normal.

Originally Posted by aufkrawall

Yes, it is.

Thank you!

I seem to have reached a stable setup.

Nvidia drivers are a bit of a hit and miss right now, which is pretty sad, but the GPU works well when it works.

[billqs]

Originally Posted by thighhighs

billqs
gtx1070 should be good for 4k. What madvr settings you use? If you like nnedi, use the double\quadruple luma only 32\64 + SuperRes in upscaling refinement section, no deringing, orderer dithering. And create the fps-based profiles with alternative settings for >29fps sources (if you need it). Read warner guide for more http://forum.kodi.tv/showthread.php?tid=259188

My 1080p settings to 2160 are fine.
My 720p settings took into accu]nt your suggestiom
Chroma Upscaling- Reconstruction soft
Image downscaling- SSIM ID Strength 100
Image doubling- NNedi3 32 neurons
Quadrupling NNEDI3 32 neurons
no chroma doubling
Image Upscaling- Jinc
Upscalijng Refinement- Superres 4
refine after every 2x upscaling step
Ordered Dithering

Playing a 720 source I was between 40-42ms which dropped 101 frames

I am using 3820x2160 24hz refresh rate and madvr reports display 240003 and computation 24000 movie 23.987 fps.

From this I can see I'm close to being under the threshold but not quite there yet.

My drivers are 373.06. Should I back then up to 368.81?

[Asmodian]

Originally Posted by billqs

My 1080p settings to 2160 are fine.
My 720p settings took into accu]nt your suggestiom
Chroma Upscaling- Reconstruction soft
Image downscaling- SSIM ID Strength 100
Image doubling- NNedi3 32 neurons
Quadrupling NNEDI3 32 neurons
no chroma doubling
Image Upscaling- Jinc
Upscalijng Refinement- Superres 4
refine after every 2x upscaling step
Ordered Dithering

Playing a 720 source I was between 40-42ms which dropped 101 frames

I am using 3820x2160 24hz refresh rate and madvr reports display 240003 and computation 24000 movie 23.987 fps.

From this I can see I'm close to being under the threshold but not quite there yet.

My drivers are 373.06. Should I back then up to 268.81?

Using refine after every 2x upscaling step with NNEDI3 quadrupling is very very slow. I recommend you either turn off refine after every step or use super-XBR for quadrupling. This will allow you to use more neurons for doubling as well, which is much more important than using NNEDI3 for quadrupling.

Edit: This is an odd synergy between the two options that is much slower than either of them separately. I think it is something about converting back and forth between RGB to YUV.

[thighhighs]

billqs
Not use nnedi quadrupling for 720p hd. Imho, good quality nnedi-based preset 720p->2160p should be:
Chroma: super-xbr100 + AR
Image: Jinc + AR
Double Luma: nnedi 32 (or 64 m.b)
Double Chroma: no
Upscaling refinement: SuperRes 3 (refine the image only once after upscaling is complete)
Artifact removal - Debanding: low+high
Artifact removal - Deringing: Off
Image enhancements: Off
Dithering: Ordered
Jinc resource intensive for 4k upscaling, it can be reduced to lanczos3+AR, chroma super-xbr -> Lanczos3+AR or Bicubic 75 no AR for max performance.

For nnedi quadrupling 720p try like this:
Chroma: Bicubic 75 no AR
Image: Lanczos3 + AR
Quadruple Luma: nnedi 32
Quadruple Chroma: no
Downscaling: Bicubic 100 +AR (soft)
Upscaling refinement: SuperRes 3 (refine the image only once after upscaling is complete)
Artifact removal - Debanding: low+high
Artifact removal - Deringing: Off
Image enhancements: Off
Dithering: orderer

If you reduce settings more, you lose quality, imho. Better replace nnedi -> super-xbr.

[billqs]

Thanks everyone! With the GTX 1070 I was able to cut down quadrupling to Superxbr sharpness 75 and go with NNEDI3 128 Neurons for the doubling. I put Superres on 3 and only rendered once after the upscale. This produced an image very similar to the 1080p to 2160 image I watched earlier today. There was really great details in the picture, quite amazing, really. I may still end up with a 1080 and sell the 1070 but its nice to know a list of settings that really work well together!

[70MM]

Does image doubling do anything good to the image if one is only using 1080p.
Can someone explain what exactly is Image Doubling doing?
I know what it sounds like, but not really sure of the true explanation...

[huhn]

image doubling 1080p will scale it to 2160p.
if you do that for a 1080p screen it give you the opportunity to refine the image at 2160p before it get scaled down again. this is quite useful for some filter.

this is called super sampling this is just a different way to process an image. you have to ask your eyes if you want that result or if it is just a waste of GPU power.

this shouldn't be compared to video game supersampling because they are quite different.

[70MM]

Originally Posted by huhn

image doubling 1080p will scale it to 2160p.
if you do that for a 1080p screen it give you the opportunity to refine the image at 2160p before it get scaled down again. this is quite useful for some filter.

this is called super sampling this is just a different way to process an image. you have to ask your eyes if you want that result or if it is just a waste of GPU power.

this shouldn't be compared to video game supersampling because they are quite different.

Sorry Im not 100% sure what you are saying...
I think you mean this but correct me if Im wrong.

If you use 1080p BDs with Image Doubling only at 1080p, the benefit is lost when it downscales again.....

If you set the video card to 4k + image doubling, then the downscale this is much better than the above....

Am I saying this right or not?

I have a large screen 145" dia scope and doing just 1080p with image doubling looks great to me. When I change the video card to 4K it doesn't seem to look any better than when the card is at 1080 with image doubling...

Your thoughts on this please as Im still trying to get a grip on all this.

[Georgel]

Something is wrong with present times - and I can't pinpoint what. The present time starts low, and grows to 8ms after the video has been playing for a while.

Anyone knows why or how to solve?

EDIT:: Maybe this helps pinpoint the problem.

Using those settings, the problem doesn't seem to appear anymore, but the present time is 0.4 ms, which is a bit high still.

Should I look for another GPU driver?

[CarlosCaco]

Originally Posted by Georgel

Something is wrong with present times - and I can't pinpoint what. The present time starts low, and grows to 8ms after the video has been playing for a while.

Anyone knows why or how to solve?

EDIT:: Maybe this helps pinpoint the problem.

Using those settings, the problem doesn't seem to appear anymore, but the present time is 0.4 ms, which is a bit high still.

Should I look for another GPU driver?

Try lower the gpu queue you don't need much try the standard 8 should be fine...


Enviado do meu iPhone usando Tapatalk

[billqs]

Perhaps the long CPU and GPU queue sizes (well above the defaults) could be linked to the long present time? Also, you have checked two things that delay playback, maybe that's causing the presentation lag? I'm sure some much more knowledgeable folks will be able to chime in soon. That's just what I noticed.

[Georgel]

Thank you for your help~!

Windowed overlay and desktop composition seem to be related to the problem.

Setting the delay until queu is full will only make the video pause for a second or so when seeking to make sure that the playback queue is full, this helps with less dropped frames when seeking or when opening a file. Still have 3-4 frames dropped when going full screen.

[CarlosCaco]

Originally Posted by Georgel

Thank you for your help~!



Windowed overlay and desktop composition seem to be related to the problem.



Setting the delay until queu is full will only make the video pause for a second or so when seeking to make sure that the playback queue is full, this helps with less dropped frames when seeking or when opening a file. Still have 3-4 frames dropped when going full screen.

Don t seens to be a overlay problem, on the first picture is showing d3d11 windowed fullscreen... overlay mode is d3d9 if you tick the d3d11 overlay do not activate... maybe you could try untick d3d11 and fullscreebn exclusive and try overlay mode... it's a great mode, as you said previously you use a icc profile, in overlay if you enable gpu gamma ramps and enable gamma processing madvr process the icc in 16bit which have greater quality...


Ps: also in the first picture shows you are using dxva2 and the second not... this can be the issue...
also dxva2 native with nvidia its not good with madvr causes some blur on chroma layer... if I was you i give a try to software decoding... even my old i5 2500 can decode hevc 1080p sources 8bit and 10bit... BTW the option delay playback until... don t work with dxva2 native... this also can be the problem




Enviado do meu iPhone usando Tapatalk

[Asmodian]

Originally Posted by 70MM

Sorry Im not 100% sure what you are saying...
I think you mean this but correct me if Im wrong.

If you use 1080p BDs with Image Doubling only at 1080p, the benefit is lost when it downscales again.....

No, you can run upscaling refinement after doubling but before downscaling. Running sharpening at a higher resolution and then downscaling often looks better than simply running sharpening at the native resolution. NNEDI3 doubling can also "repair" minor aliasing that might be present in the source. Try SSIM downscaling.

Originally Posted by 70MM

If you set the video card to 4k + image doubling, then the downscale this is much better than the above....

Am I saying this right or not?

No, generally everything has worse downscaling than madVR. It is better to set your GPU to the native resolution of your screen and do any supersampling with madVR.

Games are different in that they actually render the content at the higher resolution, not just scaling up lower resolution content and then scaling it back down again. The only reason to use super sampling with video, IMO, is to use some of the upscaling refinement options at the doubled resolution.

Originally Posted by 70MM

I have a large screen 145" dia scope and doing just 1080p with image doubling looks great to me. When I change the video card to 4K it doesn't seem to look any better than when the card is at 1080 with image doubling...

Your thoughts on this please as Im still trying to get a grip on all this.

You do have a 1920x1080 display? As I said above, always output your screen's native resolution from the GPU.

[nussman]

Originally Posted by CarlosCaco

Don t seens to be a overlay problem, on the first picture is showing d3d11 windowed fullscreen... overlay mode is d3d9 if you tick the d3d11 overlay do not activate... maybe you could try untick d3d11 and fullscreebn exclusive and try overlay mode... it's a great mode, as you said previously you use a icc profile, in overlay if you enable gpu gamma ramps and enable gamma processing madvr process the icc in 16bit which have greater quality...


Enviado do meu iPhone usando Tapatalk

Seems to be no (madVR) problem at all ...
@Georgel is maxing "every" setting and then complaining about 3-4 dropped frames when going to full screen? Come on ...

[iSeries]

Hi, this might seem like a stupid question. If I switch to windowed overlay (was using D3D11 FSE), could this mode result in output being different? Will I need to create a new 3dlut? (No ICC profile is being used, defaults in MadTPG were used for 3dlut creation - disable videoluts etc).