Welcome to Doom9's Forum, THE in-place to be for everyone interested in DVD conversion.

Before you start posting please read the forum rules. By posting to this forum you agree to abide by the rules.

 

Go Back   Doom9's Forum > Hardware & Software > Software players

Reply
 
Thread Tools Search this Thread Display Modes
Old 27th November 2017, 06:37   #47361  |  Link
ryrynz
Registered User
 
ryrynz's Avatar
 
Join Date: Mar 2009
Posts: 3,646
You guys that like double again, try quadruple with some/more? Adaptive Sharpen.
ryrynz is offline   Reply With Quote
Old 27th November 2017, 12:02   #47362  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by ashlar42 View Post
My opinion is that NNEDI3 can go the way of the dodo.
Quote:
Originally Posted by 3x3cUt0r View Post
I stopped using NNEDI3, so i vote yes, go ahead and remove it.
Quote:
Originally Posted by foozoor View Post
You can clearly throw this opencl based sh*t in the trash.
Quote:
Originally Posted by Gopa View Post
I never use NNED13 for anything. Prefer: NGU AA.
Quote:
Originally Posted by Thunderbolt8 View Post
I dont use nnedi3 so imho it can go
Quote:
Originally Posted by ABDO View Post
I never ever use NNEDI3 since NGU AA come to madvr, so for me it is time to go.
Quote:
Originally Posted by Polopretress View Post
I never ever use NNEDI3.
Quote:
Originally Posted by cork_OS View Post
2) 3) NGU AA made NNEDI3 obslolete. No regrets.
Quote:
Originally Posted by ryrynz View Post
Double again and NNEDI3 can go.
Quote:
Originally Posted by SamuriHL View Post
Yes, NNEDI3 and double again can go.
Quote:
Originally Posted by Neo-XP View Post
IMHO, NNEDI3 is not relevant anymore with NGU AA, and if it can help you to get rid of OpenCl, remove it.
Quote:
Originally Posted by yukinok25 View Post
I have never used NNEDI3 as is terribly slow with my GPU, then I guess I'll have to try to use NGU-Anti Alias and see what's happening.
Thank you guys, that's a pretty clear vote that we can let NNEDI3 go. Rest in peace, you've served us well.

Quote:
Originally Posted by huhn View Post
and someone here reported a bug with nnedi chroma and RCA which is clear there just saying.
Yes, I've had a couple of crash reports sent to me pointing to NNEDI3 being the cause of the crash. So clearly some users are still using NNEDI3. But are they using it because they compared it to NGU AA and found NNEDI3 to be better? I doubt that. They're probably using NNEDI3 either because they still have it active from before NGU AA existed, or because some tutorial told them to use NNEDI3, or because they are familiar with the NNEDI3 name, but don't know NGU AA.

In any case, I think the vote is very clear (see above), and it matches my own opinion. So NNEDI3 will be removed from the next build. I'll first remove just the options, but leave the code there, just in case definite proof comes up that it's still needed, after all.

Quote:
Originally Posted by huhn View Post
well deinterlancing is currently a painful topic with an nvidia card and nnedi3 is a deinterlancer so maybe you could use it for that until we get proper deinterlancing without mismatch chroma positions and stuff like this...
you know give it something useful todo.

isn't there a fast and working shader version of it?
Yes, NNEDI3 was originally made for deinterlacing. However, if I wanted to, I could make a tweaked NGU variant for deinterlacing, too. Video mode deinterlacing is still on my list of things to look at. But priority for that keeps sinking because natively interlaced video content keeps becoming rarer every day. FWIW, it's not just a hour or two of making a deinterlacing solution available with NNEDI3, it would be a major effort, so not happening. I'm confident an NGU deinterlacing variant would beat NNEDI3 in terms of power vs quality. NNEDI3 was great for its time (it was actually very ahead of its time), but science/technology has caught up and overtaken it in the meanwhile.

Quote:
Originally Posted by Gopa View Post
I never use double again/twice.
Quote:
Originally Posted by Polopretress View Post
No double/again upscaling also.
Quote:
Originally Posted by cork_OS View Post
1) "Double again" is too sharp and slow, never used it since direct quadruple is appears.
Quote:
Originally Posted by ryrynz View Post
Double again and NNEDI3 can go.
Quote:
Originally Posted by Asmodian View Post
I agree, both double again and NNEDI3 are not needed going forward.
Quote:
Originally Posted by SamuriHL View Post
Yes, NNEDI3 and double again can go.
Quote:
Originally Posted by jkauff View Post
No problem here with any of your three proposal questions.
Ok, several votes for not needing "double again".

Quote:
Originally Posted by Neo-XP View Post
I am not happy with the "direct quadruple" option. It looks not natural at all ("plastic" look), not as sharp as the "double again" option and adds a lot of aliasing.
Quote:
Originally Posted by ABDO View Post
ok, since NGU direct quadruple Implementation, i did some comparison with double again, and i found some Reasons make me prefer to stay using double again In spite of the diffrent performance between them, for me double again look somehow a little bit best than direct quadruple, a little details look deeper, some lines here or here treated best, and the overall picture look little best for my eyes.
Ok, votes noted. Can you provide some screenshots (only the original small image suffices)? I've tested with my favorite test images and to my eyes direct quadruple usually looks better or at least not worse overall. In the "castle" image some lines are better and some worse with quadruple vs double again, so no clear winner there. But in the "clown" and "small" images, to my eyes direct quadruple clearly wins. See here:

http://www.madshi.net/clown.png
http://www.madshi.net/small.png

Quote:
Originally Posted by Gopa View Post
NGU Sharp + Free RCA!!!
Told you you'd love it!

Quote:
Originally Posted by JarrettH View Post
Why is RCA so important to be built into NGU?
Because we're not photo editors who get RAW out of our cameras. We're getting highly compressed videos from the studios.

Quote:
Originally Posted by ABDO View Post
with double again i can make double high and quadruple meduim and set chroma to high, but with direct quadruple i forced to use medium at all, i think this would be a Problem to weak gpu owners like me.
Are you sure that double high + double medium is faster for you than direct quad high? That seems surprising to me.
madshi is offline   Reply With Quote
Old 27th November 2017, 12:03   #47363  |  Link
foozoor
Registered User
 
foozoor's Avatar
 
Join Date: Feb 2012
Posts: 116
Hey madshi,
What config (mpv.conf) did you use to test FSRCNNX?
Did you use the D3D11, Vulkan or the old and slow Angle backend/api?
foozoor is offline   Reply With Quote
Old 27th November 2017, 12:11   #47364  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
I used shinchiro's 2017-11-19 64bit build with "profile=gpu-hq". I didn't specify which backend to use. I trusted in mpv auto-picking the best option automatically. Bad idea? If D3D11 or Vulkan are better, why are they not the default backend/api?
madshi is offline   Reply With Quote
Old 27th November 2017, 12:23   #47365  |  Link
foozoor
Registered User
 
foozoor's Avatar
 
Join Date: Feb 2012
Posts: 116
Vulkan doesn't work on old GPUs and the d3d11 renderer is not documented.
Igv released FSRCNNX_x2_r1_16-0-4-3.glsl that is quite faster too.

To get proper benchmark with vulkan, use this:
https://haasn.xyz/posts/2017-10-05-h...hroughput.html

Last edited by foozoor; 27th November 2017 at 13:06.
foozoor is offline   Reply With Quote
Old 27th November 2017, 12:50   #47366  |  Link
Asmodian
Registered User
 
Join Date: Feb 2002
Location: San Jose, California
Posts: 4,406
Quote:
Originally Posted by madshi View Post
Are you sure that double high + double medium is faster for you than direct quad high? That seems surprising to me.
On my Titan X (Pascal) at 1569 MHz scaling 528x288 to 2048x1152 NGU Sharp high 2x -> medium 2x is ~1.8 ms faster than high 4x.

However, a medium-high 4x option seems better if all we want are more steps to maximize the power available in a GPU.
__________________
madVR options explained
Asmodian is offline   Reply With Quote
Old 27th November 2017, 13:25   #47367  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by Asmodian View Post
On my Titan X (Pascal) at 1569 MHz scaling 528x288 to 2048x1152 NGU Sharp high 2x -> medium 2x is ~1.8 ms faster than high 4x.
With my Nvidia 1070, when upscaling 720p by 400%, I'm getting 13.3ms with 4x High, and about 14.0ms with 2x High + 2x Medium. In any case, it seems rather close, so is this really a reason to keep "double again"?

FWIW, once I switch to D3D11 rendering, NGU High + Very High will probably get a small speed boost. NGU Medium + Low probably not.
madshi is offline   Reply With Quote
Old 27th November 2017, 13:36   #47368  |  Link
huhn
Registered User
 
Join Date: Oct 2012
Posts: 7,903
Quote:
Originally Posted by madshi View Post
Yes, NNEDI3 was originally made for deinterlacing. However, if I wanted to, I could make a tweaked NGU variant for deinterlacing, too. Video mode deinterlacing is still on my list of things to look at. But priority for that keeps sinking because natively interlaced video content keeps becoming rarer every day. FWIW, it's not just a hour or two of making a deinterlacing solution available with NNEDI3, it would be a major effort, so not happening. I'm confident an NGU deinterlacing variant would beat NNEDI3 in terms of power vs quality. NNEDI3 was great for its time (it was actually very ahead of its time), but science/technology has caught up and overtaken it in the meanwhile.
if you can use NGU for deinterlancing sometime in the future well rip nnedi.
huhn is offline   Reply With Quote
Old 27th November 2017, 13:51   #47369  |  Link
Sunset1982
Registered User
 
Join Date: Sep 2014
Posts: 280
Quote:
Originally Posted by madshi View Post
Thank you guys, that's a pretty clear vote that we can let NNEDI3 go. Rest in peace, you've served us well.


But are they using it because they compared it to NGU AA and found NNEDI3 to be better? I doubt that. They're probably using NNEDI3 either because they still have it active from before NGU AA existed, or because some tutorial told them to use NNEDI3, or because they are familiar with the NNEDI3 name, but don't know NGU AA.
I would also vote for deleting NNEDI3.

And I doubt it too. I think there are a lot of outdated How to's out there where NNEDI3 was the best solution, before the era of NGU started. I think it will be fine to let NNEDI3 rest in peace.
__________________
Intel i5 6600, 16 GB DDR4, AMD Vega RX56 8 GB, Windows 10 x64, Kodi DS Player 17.6, MadVR (x64), LAV Filters (x64), XySubfilter .746 (x64)
LG 4K OLED (65C8D), Denon X-4200 AVR, Dali Zensor 5.1 Set
Sunset1982 is offline   Reply With Quote
Old 27th November 2017, 14:08   #47370  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by foozoor View Post
Vulkan doesn't work on old GPUs and the d3d11 renderer is not documented.
Igv released FSRCNNX_x2_r1_16-0-4-3.glsl that is quite faster too.

To get proper benchmark with vulkan, use this:
https://haasn.xyz/posts/2017-10-05-h...hroughput.html
Ok, I've done some more tests. It seems for my Nvidia 1070, there's not much of a speed difference between Angle and d3d11, for some reason. Didn't test Vulkan. Doesn't make much sense if I can't get comparable measurement results.

The new FSRCNNX_x2_r1_16-0-4-3.glsl is faster than FSRCNNX 32-0-2-2, but slower than FSRCNN 32-0-2. But at least it runs (although not smoothly). FSRCNNX 32-0-2-2 just freezes on my PC, for some reason, when trying to play video.

So speed on my PC from fastest to slowest is:
FSRCNN 32-0-2 > FSRCNNX 16-0-4-3 > FSRCNNX 32-0-2-2

Comparing quality, FSRCNN vs FSRCNNX has no clear winner, IMHO. Either one beats the other in some image areas. Overall I might prefer FSRCNN by a tiny amount. However, this is just my personal opinion, and only based on the one "Spirited Away" test image. So take it with a pinch of salt.
madshi is offline   Reply With Quote
Old 27th November 2017, 14:09   #47371  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by huhn View Post
if you can use NGU for deinterlancing sometime in the future well rip nnedi.
Quote:
Originally Posted by Sunset1982 View Post
I would also vote for deleting NNEDI3.

And I doubt it too. I think there are a lot of outdated How to's out there where NNEDI3 was the best solution, before the era of NGU started. I think it will be fine to let NNEDI3 rest in peace.
Ok, thanks!
madshi is offline   Reply With Quote
Old 27th November 2017, 17:07   #47372  |  Link
leeperry
Kid for Today
 
Join Date: Aug 2004
Posts: 3,477
Quote:
Originally Posted by ryrynz View Post
Double again and NNEDI3 can go
+1, I'm not willing to waste CPU cycles on "Double again" even if it looked a bit better.

but non-RCA NGU Sharp will still be possible and medium/high will still be an option please?
leeperry is offline   Reply With Quote
Old 27th November 2017, 18:53   #47373  |  Link
Neo-XP
Registered User
 
Neo-XP's Avatar
 
Join Date: Mar 2016
Location: Switzerland
Posts: 140
Quote:
Originally Posted by madshi View Post
Ok, votes noted. Can you provide some screenshots (only the original small image suffices)? I've tested with my favorite test images and to my eyes direct quadruple usually looks better or at least not worse overall. In the "castle" image some lines are better and some worse with quadruple vs double again, so no clear winner there. But in the "clown" and "small" images, to my eyes direct quadruple clearly wins. See here:

http://www.madshi.net/clown.png
http://www.madshi.net/small.png
Those tiny images do not represent the real content I use madVR for.

With a 1280x694 source upscaled to UHD ( Source / NGU very High + Direct quadruple very High / NGU very High + Double again very High ) :



Do you want to know what I find ugly ? Look at the external edges of his right ear :

http://screenshotcomparison.com/comparison/124784

It looks like plastic, very fake, and less sharp.
Even NGU + Jinc looks better than Direct quadruple IMHO.
NGU + Double again looks a little better than NGU + Jinc.

I do not use it now (not enough power), but I will with my next GPU because it is the best option.

Last edited by Neo-XP; 27th November 2017 at 19:01.
Neo-XP is offline   Reply With Quote
Old 27th November 2017, 19:05   #47374  |  Link
Asmodian
Registered User
 
Join Date: Feb 2002
Location: San Jose, California
Posts: 4,406
I have to say I don't see it in your screen shot, double again looks a little sharper but no plastic look. I find direct quadruple results in a more natural look with my content.
__________________
madVR options explained
Asmodian is offline   Reply With Quote
Old 27th November 2017, 19:09   #47375  |  Link
Neo-XP
Registered User
 
Neo-XP's Avatar
 
Join Date: Mar 2016
Location: Switzerland
Posts: 140
The edges are blurred, no texture, ugly.

It is as if the worst aspects of NGU were increased, but not the good ones.

Last comparison to try to save the Double again algo ( Source / Direct quadruple / Double again ) :



Look at this :

http://screenshotcomparison.com/comparison/124788

I let you judge.

Last edited by Neo-XP; 27th November 2017 at 19:51.
Neo-XP is offline   Reply With Quote
Old 27th November 2017, 19:23   #47376  |  Link
aufkrawall
Registered User
 
Join Date: Dec 2011
Posts: 1,812
Quote:
Originally Posted by madshi View Post
I used shinchiro's 2017-11-19 64bit build with "profile=gpu-hq". I didn't specify which backend to use. I trusted in mpv auto-picking the best option automatically. Bad idea? If D3D11 or Vulkan are better, why are they not the default backend/api?
It likely was already with D3D11 since it's higher in the autoprobe order list. OGL -> ANGLE -> DX11 presentation is not the default anymore with recent git builds.
profile=gpu-hq btw. also activates deband (and dithering).
Vulkan doesn't show lower GPU usage than D3D11 on Nvidia.
aufkrawall is offline   Reply With Quote
Old 27th November 2017, 19:52   #47377  |  Link
JarrettH
Registered User
 
Join Date: Aug 2004
Location: Canada
Posts: 860
Quote:
Originally Posted by madshi View Post
Because we're not photo editors who get RAW out of our cameras. We're getting highly compressed videos from the studios.
How apt, because I am a photographer
JarrettH is offline   Reply With Quote
Old 27th November 2017, 20:02   #47378  |  Link
cork_OS
Registered User
 
cork_OS's Avatar
 
Join Date: Mar 2016
Posts: 160
Quote:
Originally Posted by Neo-XP View Post
Shirt on the shoulder is definitely better with double again. "Double edges" is the ugliest NGU effect.
__________________
I'm infected with poor sources.
cork_OS is offline   Reply With Quote
Old 27th November 2017, 20:02   #47379  |  Link
Braum
Registered User
 
Join Date: Jan 2015
Posts: 48
Quote:
Originally Posted by Neo-XP View Post
The edges are blurred, no texture, ugly.

It is as if the worst aspects of NGU were increased, but not the good ones.

Last comparison to try to save the Double again algo ( Source / Direct quadruple / Double again ) :



Look at this :

http://screenshotcomparison.com/comparison/124788

I let you judge.
Double again looks better indeed, there's more details and the edges are sharper, less blurred.
Braum is offline   Reply With Quote
Old 27th November 2017, 22:39   #47380  |  Link
Asmodian
Registered User
 
Join Date: Feb 2002
Location: San Jose, California
Posts: 4,406
Quote:
Originally Posted by cork_OS View Post
Shirt on the shoulder is definitely better with double again. "Double edges" is the ugliest NGU effect.
I agree, on both counts. In this scene double again helps reduce the "double edges" artifact which is very beneficial.

Some of my content has this issue with NGU Sharp and some doesn't, for anything that shows double edges with NGU Sharp I use NGU AA.
__________________
madVR options explained
Asmodian is offline   Reply With Quote
Reply

Tags
direct compute, dithering, error diffusion, madvr, ngu, nnedi3, quality, renderer, scaling, uhd upscaling, upsampling

Thread Tools Search this Thread
Search this Thread:

Advanced Search
Display Modes

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off

Forum Jump


All times are GMT +1. The time now is 15:44.


Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2024, vBulletin Solutions Inc.