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 1st June 2015, 15:25   #30661  |  Link
aufkrawall
Registered User
 
Join Date: Dec 2011
Posts: 1,716
This should work, I used it myself that way. With OpenCL 1.2 driver, performance is even said to be improved.
I'd not buy a Radeon for NNEDI3 because AMD hasn't fixed OpenCL -> D3D interop bug for ages now.

However, now with SuperRes, you can also achieve very good results without NNEDI3, as long as you aren't watching cartoons.

If you want to play games with the card, I'd wait for AMD's Fiji release which should make the 980 less expensive. VRAM of 970 is crippled in an odd way.
aufkrawall is offline   Reply With Quote
Old 1st June 2015, 15:52   #30662  |  Link
nevcairiel
Registered Developer
 
Join Date: Mar 2010
Location: Hamburg/Germany
Posts: 9,835
Quote:
Originally Posted by aufkrawall View Post
VRAM of 970 is crippled in an odd way.
That is of no consequence for video. In fact, its of very little consequence in general.
__________________
LAV Filters - open source ffmpeg based media splitter and decoders
nevcairiel is offline   Reply With Quote
Old 1st June 2015, 16:19   #30663  |  Link
petran79
Registered User
 
Join Date: Aug 2007
Posts: 82
Finally! Nvidia GTX660 Vsync issue on Windows XP has been solved with latest Madshi and Nvidia drivers 347.88
After almost 2 years wait!

After every Nvidia driver update on Windows XP and every new madshi release I tested videos but had the same issue:

June 2013:
http://forum.doom9.org/showpost.php?...ostcount=18951

Originally Posted by petran79
for a few seconds or even for 1 minute, at that said scene "display" goes at 0000000 Hz instead of 59

Madshi:

Oh, that. Yeah, sounds like a GPU driver issue. madVR highly depends on being able to get up-to-date vsync information all the time. If the display goes to 0000000 Hz that means that the GPU driver isn't reliably providing vsync information. There's probably not much I can do about it, sadly...


This time Vsync seems to be stable at 60 hz at full screen (no exclusive mode), sometimes 61.

In windowed mode it fluctuates from 52-55 hz before stabilizing at 59-60, but never at 0 hz, thus not dropping frames

I can finally watch movies in good quality and without tearing on Windows XP like 3 years ago. Windows 7 is too sluggish.
petran79 is offline   Reply With Quote
Old 1st June 2015, 16:24   #30664  |  Link
aufkrawall
Registered User
 
Join Date: Dec 2011
Posts: 1,716
Quote:
Originally Posted by nevcairiel View Post
That is of no consequence for video. In fact, its of very little consequence in general.
I was explicitly referring to games, and consequences can be enormous (I tested this myself).
But not for video, sure. GTX 970 is a very reasonable card for madVR, probably the best (if you don't need native HEVC decoding).
aufkrawall is offline   Reply With Quote
Old 1st June 2015, 19:45   #30665  |  Link
MysteryX
Soul Architect
 
MysteryX's Avatar
 
Join Date: Apr 2014
Posts: 2,173
Quote:
Originally Posted by Braum View Post
I don't think it matters because your rig isn't powerfull enough to upscale with NNEDI.

I got a i5 3570@4,5Ghz and a 7870XT and to upscale 1280x720 to 1680x998 I can't use Double NNEDI 16x.

I can use NEDI in double and quadruple tough.
I won't use NNEDI3 for 720p videos, but will definitely for 288p videos! Then I'd have to test whether it would also work on 480p content.

I'm curious. Why were Jinc4, Jinc8 and Lanczos8 deemed "bad" algorithms?


Btw, kind of off-topic, but yesterday I realized that in a video game (with latest Catalyst driver), performance was *higher* when using full-screen windowed mode than when in exclusive mode! Not sure why. I tried that in madVR. There doesn't seem to be any performance difference between windowed mode and exclusive mode; but switching into full screen is much faster.

Last edited by MysteryX; 24th June 2015 at 07:08.
MysteryX is offline   Reply With Quote
Old 1st June 2015, 19:52   #30666  |  Link
Thunderbolt8
Registered User
 
Join Date: Sep 2006
Posts: 2,171
is anyone able to tell if windows 10 has positive or negative impact on madvr (nvidia (mobile)) performance?
__________________
Laptop Acer Aspire V3-772g: i7-4202MQ, 8GB Ram, NVIDIA GTX 760M (+ Intel HD 4600), Windows 8.1 x64, madVR (x64), MPC-HC (x64), LAV Filter (x64), XySubfilter (x64)
Thunderbolt8 is offline   Reply With Quote
Old 1st June 2015, 19:59   #30667  |  Link
iSunrise
Registered User
 
Join Date: Dec 2008
Posts: 497
Quote:
Originally Posted by Thunderbolt8 View Post
is anyone able to tell if windows 10 has positive or negative impact on madvr (nvidia (mobile)) performance?
Personally, I would not risk it just yet, at least wait for RTM. For desktops Win10 might be ok already, but mobiles have their own weird problems, even on an already stable and feature-complete OS.
iSunrise is offline   Reply With Quote
Old 1st June 2015, 20:08   #30668  |  Link
huhn
Registered User
 
Join Date: Oct 2012
Posts: 5,965
i have issue on windows 10 and i will report issue with the RTM version which will be available soon.
huhn is offline   Reply With Quote
Old 1st June 2015, 20:42   #30669  |  Link
RainyDog
Registered User
 
Join Date: May 2009
Posts: 168
Quote:
Originally Posted by nevcairiel View Post
You are using Catmull-Rom for downscaling, right? With other algorithms, linear light might not look correct, its only really recommended for CR.
I've always used Spline3 for downscaling as I prefer the slightly sharper look to CR.

So I shouldn't be downscaling in linear light if using Spline?
RainyDog is offline   Reply With Quote
Old 1st June 2015, 20:58   #30670  |  Link
aufkrawall
Registered User
 
Join Date: Dec 2011
Posts: 1,716
imho one shouldn't use it at all. There's a reason why usually gamma corrected is considered as the desired result (or better said: gamma correction gives the desired result in the end).
I just played some desktop capture videos and fonts look totally awful with linear light downscaling.
aufkrawall is offline   Reply With Quote
Old 1st June 2015, 21:43   #30671  |  Link
6233638
Registered User
 
Join Date: Apr 2009
Posts: 1,019
Quote:
Originally Posted by aufkrawall View Post
imho one shouldn't use it at all. There's a reason why usually gamma corrected is considered as the desired result (or better said: gamma correction gives the desired result in the end).
I just played some desktop capture videos and fonts look totally awful with linear light downscaling.
I really don't think that looking at computer graphics is a good test for downscaling.
You should be using photographs or movies.
Something that was filmed instead of rendered.

With that, linear light downscaling should be better - at least if your display is properly calibrated.

http://forum.doom9.org/showpost.php?...ostcount=25427 (note: this is a macro image of a fly)
View the comparison images here at 100% in your browser and switch tabs.

The image scaled in linear light should look very close to the source, only lower resolution.
The image without linear light scaling is quite a bit darker than the source image, and finer details are lost.

Quote:
Originally Posted by RainyDog View Post
I've always used Spline3 for downscaling as I prefer the slightly sharper look to CR.
So I shouldn't be downscaling in linear light if using Spline?
Catmull-Rom with the anti-ringing filter enabled is the only option which does not result in ugly ringing or aliasing artifacts when linear light scaling is enabled.
I don't recommend enabling the linear light scaling option with anything else.
An old example comparing Catmull-Rom and Lanczos.
6233638 is offline   Reply With Quote
Old 1st June 2015, 22:38   #30672  |  Link
aufkrawall
Registered User
 
Join Date: Dec 2011
Posts: 1,716
Quote:
Originally Posted by 6233638 View Post
I really don't think that looking at computer graphics is a good test for downscaling.
You should be using photographs or movies.
Something that was filmed instead of rendered.
I did with that grass picture:
http://upload.wikimedia.org/wikipedi...chip_trail.jpg

Quote:
Originally Posted by 6233638 View Post
http://forum.doom9.org/showpost.php?...ostcount=25427 (note: this is a macro image of a fly)
View the comparison images here at 100% in your browser and switch tabs.

The image scaled in linear light should look very close to the source, only lower resolution.
The image without linear light scaling is quite a bit darker than the source image, and finer details are lost.
What is the point of this nearest neighbor upscaling?
It introduces tons of aliasing.
It would be better if the source's resolution would be high enough to have a realistic result in a common native resolution.
aufkrawall is offline   Reply With Quote
Old 1st June 2015, 22:48   #30673  |  Link
cyberscott
Registered User
 
Join Date: Oct 2007
Posts: 81
Quote:
Originally Posted by madshi View Post
The first log makes sense, but the 2nd doesn't. The 2nd log indicates that it wasn't from the build 2 I intended you to test. Either I uploaded the wrong file, or you tested with the wrong file. To be safe, I've uploaded the 2nd build again:

http://madshi.net/madVR8810test4.rar

Could you please test this build again? Thanks!
It is very possible I messed the 2nd one up, was short on time and tried to do it quick.
I ran test 4 and same behavior with the queues. Here is the log.
http://s000.tinyupload.com/?file_id=...22210065716807

Last edited by cyberscott; 1st June 2015 at 23:16. Reason: grammer
cyberscott is offline   Reply With Quote
Old 1st June 2015, 23:31   #30674  |  Link
6233638
Registered User
 
Join Date: Apr 2009
Posts: 1,019
Quote:
Originally Posted by aufkrawall View Post
What is the point of this nearest neighbor upscaling?
It introduces tons of aliasing.
It would be better if the source's resolution would be high enough to have a realistic result in a common native resolution.
To keep the image the same size so that you can tab through all of them in your browser and make an accurate comparison.
Though fine details are better preserved by using linear light scaling, it is the overall brightness/appearance of the image which is important.

Linear light scaling preserves the "silvery" appearance of the eyes, while scaling without that option shifts the tone more towards brown.
If you wear glasses, you may even want to take them off when comparing the three if you find the aliasing distracting. Resolution should not be a factor in this comparison.

And please ensure that the images are not being scaled at all, they must be displayed at 100% size.
6233638 is offline   Reply With Quote
Old 2nd June 2015, 04:24   #30675  |  Link
baii
Registered User
 
Join Date: Dec 2011
Posts: 180
In general, is there a consent on using higher needi3 neurons vs using SuperRes/ more pass SuperRes?

I have some 480p DVD that is like 0.5% of the content I watch, but I really like those~. My eye test seem to favor more neuron(32->64) compare than more pass, or even activating SR, but then there is the pixel shift which make the comparison kind of weird/hard ?

Last edited by baii; 2nd June 2015 at 04:33.
baii is offline   Reply With Quote
Old 2nd June 2015, 08:54   #30676  |  Link
huhn
Registered User
 
Join Date: Oct 2012
Posts: 5,965
in my test more passes with SR is very harmful to the picture unlike nnedi3 with more neurons so hard to compare.
huhn is offline   Reply With Quote
Old 2nd June 2015, 09:26   #30677  |  Link
Anima123
Registered User
 
Join Date: Jun 2005
Posts: 462
The effect of SuperRes with madVR is not exactly the same as the one with MPDN for some reason, I don't know why.

I'd prefer the MPDN version, quality speaking. Maybe madshi have some idea to tweak the SuperRes a little bit, or there's some kind of bug when SuperRes and NEDI were included within madVR?
Anima123 is offline   Reply With Quote
Old 2nd June 2015, 11:18   #30678  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,137
Quote:
Originally Posted by aufkrawall View Post
I meant P8 state, sorry.
madshi did something to fix NNEDI3 quadrupling in the past with image enhancement because there was a similar issue.
I suppose it's still related to NNEDI3 quadrupling + image enhancement.
Driver gets totally stuck in P8 state, thus it drops lots of frames if it's too slow. It's not able anymore to enter a higher P-state unless the driver is restarted.
I had fixed a bug, but that had nothing to do with power states. I have zero influence on which power states the GPU switches to.

Quote:
Originally Posted by shaolin95 View Post
Thanks a lot. Its very easy with that example indeed so I got a good idea now of how they behave. With Blurays though telling the difference is another story.
Differences are hard to see in most scenes. However, there are some scenes in which differences can be visible. It's usually when there's lots of red and black.

Quote:
Originally Posted by andybkma View Post
Hi madshi, found a bug with mVR 88.10 that is not occurring with 88.7. Whenever I try to play the attached link two AVC clips Zoom Player immediately crashes when I use either LAV or ffdshow as the AVC video decoder, whether software decoding, CUDA or Quicksync is selected. I duplicated the problem using PotPlayer and setting the same external AVC decoders as the video decoder. Oddly when I select PotPlayer's AVC built-in decoder the clips don't crash so perhaps that may help with figuring out what is wrong. Like I mentioned, this is new prob, the clips play fine with mVR 88.7. Please let me know what other info I can provide.
I think this will probably be fixed in the next build.

Quote:
Originally Posted by QBhd View Post
I have some more info for those of us having issues with D3D11 and PotPlayer.

I use a 72 Hz refresh rate for 23.976 playback (with ReClock) and a 60 Hz refresh rate for 29.970/59.940.... PotPlayer's onscreen info (TAB) will show 23.976-->72 (actually starts a little above 72, then falls to 72 and stays there). When the video is paused (or something else pulls focus away) there is a flicker (like going out of FSE) however ctrl-J still says FSE... and now the OSD of PP will show 23.976-->70 (and the 70 keeps falling until it hits 24)... ctrl-J shows empty queues (with no dropped frames, oddly enough) and playback gets super jittery. I have been racking my brains trying to find a way for this to work, but no setting in PP, ReClock or madVR allows me to pause the video and resume without this very very odd behavior. Playback is awesome as long as nothing interrupts it, but many times a video can be paused (can't miss anything for a beer run or to take a leak!!) and having to stop and restart each time is not an option. So I keep going back to D3D9.

To me, it seems that once the stream is interrupted, the refresh that is being reported gets altered in some way to that of the video and not the target... I just tried without ReClock, and everything I described above happens except the playback seems smooth, even with the last two queues not filling up. It's all very odd
Those PotPlayer frame rate reports don't have a lot of meaning.

Does this problem only occur with PotPlayer? Or also with e.g. MPC-HC? I think there is a setting in PotPlayer which causes madVR to rerender frames all the time in paused state, causing rather high GPU consumption in paused state. I'm not very familiar with PotPlayer, though. If you disable that feature (in case you know which it is), does that make the problem go away?

Quote:
Originally Posted by aufkrawall View Post
Could it be that downscaling in linear light can totally shift an image in an unwanted way?
Some time ago, I did some tests with downscaling game sreenshots, and with linear light everything seemed to be getting too bright.
Linear light should produce the "correct" result, unlike gamma light. The reason is that blending pixels together in gamma light is mathematically not correct. And downscaling blends a lot of source pixels together for each target pixel. For upscaling it's a bit different. In theory linear light should produce more correct results for upscaling, too, but it adds so much ringing that the positive benefits are outweighed.

Quote:
Originally Posted by nevcairiel View Post
You are using Catmull-Rom for downscaling, right? With other algorithms, linear light might not look correct, its only really recommended for CR.
All 2-tap algos should be fine. However, 3-taps and more produce a lot of ringing when using linear light which the anti-ringing filter doesn't completely erase. Other than the ringing, using 3-taps/4-taps for linear light downscaling is ok, too. It's really only the ringing which is the key problem.

Quote:
Originally Posted by Shiandow View Post
FWIW you can actually prove mathematically that using linear light (with a gamma > 1) will make the image brighter. For more information see Jensen's inequality.
Hmmmm... Isn't that backwards? I mean don't we have to think of linear light as a straight line, and of gamma light as a curve? Jensen's inequality should prove that *gamma light* produces incorrect results, not linear light. Or am I missing something? You seem to suggest that gamma light is a straight line and that linear light is a curve. But I think it's exactly the other way round? Images/videos are using gamma light because that's how CRTs used to work, and also because it happens to approximate the way our eyes perceptually work. But nature/physics really is linear light, not gamma light.

Quote:
Originally Posted by dansrfe View Post
Is it possible to have an option to use D39 overlay in windowed mode and D311 in exclusive mode?
Possible yes. But too much work.

Quote:
Originally Posted by XRyche View Post
I know this is the madVR thread but since Shiandow's debanding algorithm is being tested along side madVR's, I hope the thread will forgive me for the following.
In my tinkering with the deband algorithm's this weekend I have confirmed for myself a rather happy alternative use for Shiandow's algorithm set at 1.00 strength and .01 margin with grain. It does an excellent job of deblocking and denoising my old dvr rips at sub 480p resolution without totally killing detail (in tandem with nnedi3) and making it a fuzzy mess. Well more fuzzy than it already was.
This is just my opinion, but I think it would really be beneficial (for me and possibly even a small minority of users) if Shiandow might be able to optimize his debanding algorithm further down that path and possibly madshi could use it as a denoiser/general deblocker for madVR somewhere in the near future.
Deblocking should be handled separately from debanding, IMHO. Why? Because a deblocking algorithm can actually assume that blocks appear at typical distances (e.g. at 8x8 pixel boundaries). This information should allow a deblocking algorithm to work more efficiently than trying to make a debanding algorithm remove block compression artifacts.

That said, yes, maybe Shiandow's debanding algorithm could be used as a starting point for a new deblocking algorithm. But there's no room in my time schedule at least for that anytime soon.

Quote:
Originally Posted by MSL_DK View Post
With v0.88.9 and v0.88.10 Average stats > rendering never stabilize, it jumps up and down and i've have frame drops. It applies to Jinc and jumps from 6ms to 22ms. If I choose Bicubic it settles at around 18ms after a few seconds.
In older versions rendering times were averaged "forever". That means you got nicely stable times after a while. *However*, directly after switching algos or changing zoom factors or things like that, rendering times are usually much higher than a couple of seconds later. This means that old way of long-time averaging often started too high and needed a looooong time to settle down to the correct values. The new method averages rendering times over just 1 second of playback. So naturally the rendering times can jump quite a bit. However, if the rendering times over multiple consecutive measurement periods are relatively near to each other, madVR averages them together again. So as long as the jumps are not too large from one measurement second to the next, the times should calm down over time, too, and then get much quicker to the correct end result compared to before.

Quote:
Originally Posted by petran79 View Post
Finally! Nvidia GTX660 Vsync issue on Windows XP has been solved with latest Madshi and Nvidia drivers 347.88
After almost 2 years wait!
FWIW, I don't think it's the drivers. I've added a number of changes for Optimus systems, and I think those probably have helped with your specific problem, too. The drivers are probably still as buggy as ever, but madVR has added a number of driver workarounds to be more forgiving with bad drivers.

Quote:
Originally Posted by baii View Post
In general, is there a consent on using higher needi3 neurons vs using SuperRes/ more pass SuperRes?

I have some 480p DVD that is like 0.5% of the content I watch, but I really like those~. My eye test seem to favor more neuron(32->64) compare than more pass, or even activating SR, but then there is the pixel shift which make the comparison kind of weird/hard ?
Quote:
Originally Posted by huhn View Post
in my test more passes with SR is very harmful to the picture unlike nnedi3 with more neurons so hard to compare.
Quote:
Originally Posted by Anima123 View Post
The effect of SuperRes with madVR is not exactly the same as the one with MPDN for some reason, I don't know why.

I'd prefer the MPDN version, quality speaking. Maybe madshi have some idea to tweak the SuperRes a little bit, or there's some kind of bug when SuperRes and NEDI were included within madVR?
There are small differences in the SuperRes implementation. But you guys are too fast. We're still discussing debanding atm. We'll get to SuperRes soon enough.

Quote:
Originally Posted by James Freeman View Post
I find Shiandows 0.80, 0.0, Grain On; vs High, identical in terms of detail loss and shape retention.

IMO, They are so close, any one of them will do perfectly.
As the old saying goes, if it's working don't touch it.
Quote:
Originally Posted by vivan View Post
On debanding:
"Add grain". It adds grain, yes, but it's 10x stronger than madVRs/f3kdb's and I don't really feel it being useful. In some cases it even makes an opposite effect and makes banding more visible.
Overall I still strongly prefer madVR debanding over Shiandow's (on anime sources).
Thank you for your feedback, once again!

Any more votes about debanding?
madshi is offline   Reply With Quote
Old 2nd June 2015, 11:20   #30679  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,137
Quote:
Originally Posted by cyberscott View Post
It is very possible I messed the 2nd one up, was short on time and tried to do it quick.
I ran test 4 and same behavior with the queues. Here is the log.
http://s000.tinyupload.com/?file_id=...22210065716807
Ok, thanks. One more try, if you don't mind:

http://madshi.net/madVR8810test5.rar

Please activate debug mode first (by double clicking "activate debug mode.bat"), then afterwards drop in the new madVR.ax. If you do it the other way round, but you're replacing the "wrong" madVR.ax file and you're testing with the official v0.88.10 again. I know, it's a bit confusing. Basically, after you dropped the replacement madVR.ax file into the madVR folder, don't double click any batch files, until you've completed testing. Thanks!
madshi is offline   Reply With Quote
Old 2nd June 2015, 14:10   #30680  |  Link
Shiandow
Registered User
 
Join Date: Dec 2013
Posts: 752
Quote:
Originally Posted by madshi View Post
Hmmmm... Isn't that backwards? I mean don't we have to think of linear light as a straight line, and of gamma light as a curve? Jensen's inequality should prove that *gamma light* produces incorrect results, not linear light. Or am I missing something? You seem to suggest that gamma light is a straight line and that linear light is a curve. But I think it's exactly the other way round? Images/videos are using gamma light because that's how CRTs used to work, and also because it happens to approximate the way our eyes perceptually work. But nature/physics really is linear light, not gamma light.
Jensen's inequality doesn't really care about which one is "correct" or "linear". It just uses the fact that the function used to go from gamma light to linear light is convex. If you go the other way around then the function you use is concave, so the inequality will be the other way around. You can really only use it to show that the values will (always) be higher when you use linear light, you can't use it to show whether this is better or not.
Shiandow 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 18:24.


Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2019, vBulletin Solutions Inc.