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 22nd October 2012, 07:03   #14961  |  Link
sunnah
Registered User
 
Join Date: Aug 2012
Posts: 12
Quote:
Originally Posted by leeperry View Post
PS: bah, now I crave a new GPU so I can use jinc3 AR for 720p to 1080p luma upscale...my 96SP 8800GS only allows jinc3 AR for chroma and lanczos/spline AR for luma. This said, lanczos3 AR looks quite good and maybe you'll manage to optimize jinc3 AR for luma at some point
Try to set in chroma softcubic 80 and for luma upscaling jinc3 or jinc4. The supposed difference what the human eye could detect from chroma upsampling is the color if you zoom-in on the image on a red color and some other colors, you will see it becomes a little brighter with jinc. But I think, this difference is not too important if you can't use jinc for scaling, because of that.

So try that mate

Cheers.!
__________________
MSI GT683DX, i5-2140M, NVIDIA GTX 570M
sunnah is offline   Reply With Quote
Old 22nd October 2012, 07:18   #14962  |  Link
SamKook
Registered User
 
Join Date: Mar 2011
Posts: 216
Quote:
Originally Posted by cyberbeing View Post
If you set a default frame-rate like '23.976' instead of 'unknown' in Reclock, you should be able to prevent an initialization pop/gap when the estimator and your default frame-rate match.
Looks like I got bad result because of some manual framerate that were set a while back when I did my test, because that did get rid of the problem.

But now I remember why I set it to unknown in the first place instead of 23.976. It's because it can't detect properly the frame rate of some of my own videos with 59.94 fps so I wanted to avoid it using 23.976 in those cases.

Thanks for the information.
SamKook is offline   Reply With Quote
Old 22nd October 2012, 11:36   #14963  |  Link
leeperry
Kid for Today
 
Join Date: Aug 2004
Posts: 3,477
Quote:
Originally Posted by madshi View Post
In the first step I'll probably just duplicate what MPC-HC offers. A more complete solution might come at a later date.
Alright, sounds good! I see that Seb.26 released condition scripts that allow some commands to only be processed depending on the movie dimensions, so if we can process the PS scripts prior to scaling that should work fine and we could process different commands for SD and HD

I'm not sure how we could roll scripts, though? through hotkeys again I presume.

Quote:
Originally Posted by madshi View Post
I don't think I agree with that.
Oh, well it made sense to me to, for instance, sharpen onto the right colors instead of rolling colors afterwards.

Quote:
Originally Posted by madshi View Post
I wouldn't expect speed improvements for Jinc3 AR. It might even get a little slower, after some more tweaks to the AR algorithm.
OK, I need to read up more on Jinc coz I dunno whether it was placebo but it seemed to be a far superior algorithm than the others from my limited testing. Uber sharpness and zero ringing Hopefully I could get away with a $100-150 GPU.

Quote:
Originally Posted by sunnah View Post
Try to set in chroma softcubic 80 and for luma upscaling jinc3 or jinc4. The supposed difference what the human eye could detect from chroma upsampling is the color if you zoom-in on the image on a red color and some other colors, you will see it becomes a little brighter with jinc. But I think, this difference is not too important if you can't use jinc for scaling, because of that.
Actually I ran a bunch of tests yesterday and IIRC I can use jinc3 for both luma and chroma when upscaling 720p to 1080p, I can even enable AR for chroma but AR for luma is a big no-no....98% GPU load on my 8800GS and massive frames drop. Jinc3 w/o AR is an halo feast so that's a no-go.

This said, I just asked for a refund on that 46" CCFL TV coz the last ten vertical percents on the right end had some slight colorimetry drift that was almost impossible to see in 50fps digital TV but utterly annoying after a while in slow 24p pans...dunno if it came from a backlight homogeneity issue, some problem with the polarizing filters or whatever else. Strangely, everything looked fine in Nokia Monitor Test, the problem was pretty much only visible in 24p pans

And as 6233638 said, CCFL's make nasty bright green RBE's that are absolutely unbearable to me on anime, duh again

So I'm back on the market for a new TV....and the only company that doesn't con its customers with the panel lottery is Sharp so I guess a LC-40LE630E with the killer UVČA panel & 24Hz support is in good order

Last edited by leeperry; 22nd October 2012 at 12:59.
leeperry is offline   Reply With Quote
Old 22nd October 2012, 11:52   #14964  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by leeperry View Post
OK, I need to read up more on Jinc coz I dunno whether it was placebo but it seemed to be a far superior algorithm than the others from my limited testing. Uber sharpness and zero ringing Hopefully I could get away with a $100-150 GPU.
The main positive thing about Jinc is that it has a very pleasent almost analog like look with no aliasing, and a decent sharpness (though a bit less sharp than Lanczos). In comparison all other resamplers look more digital and have more aliasing. With AR turned on, there should be no difference in ringing. If you're a sharpness junky, you may want to stay with Lanczos, though, or wait until you can post-sharpen the Jinc output. A great GPU would be the 660, if you can swing it.

Quote:
Originally Posted by leeperry View Post
So I'm back on the market for a new TV....and the only company that doesn't con its customers with the panel lottery is Sharp so I guess a LC-40LE630E with the killer UVČA panel is in good order
If you plan to use it for movie playback, only, and not for any computer stuff, then I believe plasma (e.g. try Panasonic) gives the best image quality of the currently available TVs, at least if you can half way control ambient light. LCD is better when there's lots of ambient light. But we're getting a bit OT here.
madshi is offline   Reply With Quote
Old 22nd October 2012, 12:04   #14965  |  Link
leeperry
Kid for Today
 
Join Date: Aug 2004
Posts: 3,477
Quote:
Originally Posted by madshi View Post
The main positive thing about Jinc is that it has a very pleasent almost analog like look with no aliasing, and a decent sharpness (though a bit less sharp than Lanczos). In comparison all other resamplers look more digital and have more aliasing. With AR turned on, there should be no difference in ringing. If you're a sharpness junky, you may want to stay with Lanczos, though, or wait until you can post-sharpen the Jinc output. A great GPU would be the 660, if you can swing it.
Analog, exactly! It doesn't have the digital look of Lanczos

Post-sharpening the Jinc output? I've grown tired of sharpening tbh.

Quote:
Originally Posted by madshi View Post
If you plan to use it for movie playback, only, and not for any computer stuff, then I believe plasma (e.g. try Panasonic) gives the best image quality of the currently available TVs, at least if you can half way control ambient light. LCD is better when there's lots of ambient light. But we're getting a bit OT here.
Yeah, well a friend of mine who doesn't see any RBE's on his 32" CCFL Sony LCD(5K:1 CR in HCFR) tells me that he sees them very clearly on Plasma's.....he currently owns the same 4X DLP projector I used to have so I trust his eyes.

Well, if the Sharp fails as well I will try my luck with a Sammy in a brick & mortar shop that also provides a cooling-off period. Argyll's CR measurement will make the call

I'm terribly sorry for hijacking the thread, but you'll have to admit that it's still more or less mVR related because we all want the best bang for bucks gear to enjoy mVR....and it's quite a jungle out there if you want to buy a proper flat screen. But I think I've got my mind made up now, and it might help other mVR users as well

Last edited by leeperry; 22nd October 2012 at 13:00.
leeperry is offline   Reply With Quote
Old 22nd October 2012, 12:15   #14966  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by leeperry View Post
Post-sharpening the Jinc output? I've grown tired of sharpening tbh.
I don't like conventional sharpening, either. But have a look at FineSharp, that looks quite promising to me.

Quote:
Originally Posted by leeperry View Post
Yeah, well a friend of mine who doesn't see any RBE's on his 32" CCFL Sony LCD tells me that he sees them very clearly on Plasma's.....he currently owns the same 4X DLP projector I used to have so I trust his eyes.
Did he have a plasma at his home? Plasmas at shops are usually put into torch mode, so they can half way compete with the super bright LCDs. Torch mode doesn't really do any favors to plasmas. You know, stores are usually brightly lit (much brighter than your living room is), and there plasmas don't look so good compared to LCDs. It's the same as with audio volume: Louder sounds "better". In the same way brighter looks "better". But in your living room, with controlled lighting, everything is very different. I've had a plasma myself and I couldn't see any hint of RBE. Plasmas do not draw images sequentially like DLPs do. 1-chip DLPs draw color by color. Plasmas have red/green/blue subpixels and all 3 are active at the same time. So there's no reason plasmas should have any RBE. What they do have is dithering noise, because plasma pixels can only be on or off, but they can't be half on. So they simulate half on by pulsating very quickly. Why don't you go to a shop and check yourself whether you can see any hint of RBE with plasmas? They really do have the best image quality, IMHO, if you can control ambient light.
madshi is offline   Reply With Quote
Old 22nd October 2012, 12:51   #14967  |  Link
NicolasRobidoux
Nicolas Robidoux
 
NicolasRobidoux's Avatar
 
Join Date: Mar 2011
Location: Montreal Canada
Posts: 269
Quote:
Originally Posted by madshi View Post
... I wouldn't expect speed improvements for Jinc3 AR. It might even get a little slower, after some more tweaks to the AR algorithm.
Dr. Minglun Gong of Memorial University will confirm that I can work miracles refactoring HLSL code (which I assume is what you're using) when paired with an expert.
NicolasRobidoux is offline   Reply With Quote
Old 22nd October 2012, 13:31   #14968  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by NicolasRobidoux View Post
Dr. Minglun Gong of Memorial University will confirm that I can work miracles refactoring HLSL code (which I assume is what you're using) when paired with an expert.
Maybe I can hire you when I earned my first million with madVR. That might be hard to achieve with freeware, though... I fear the main performance cost is reading all the source pixels, and you can't really optimize that away with HLSL code. There might be a higher chance of optimization by using CUDA/OpenCL. For now my main priority is on adding all the missing features, though, not on squeezing out 10% more performance.
madshi is offline   Reply With Quote
Old 22nd October 2012, 13:39   #14969  |  Link
NicolasRobidoux
Nicolas Robidoux
 
NicolasRobidoux's Avatar
 
Join Date: Mar 2011
Location: Montreal Canada
Posts: 269
Put a flattr button on your download site. This may get you to your first thousand faster than you think.
(And of course it's features first.)
P.S.... or somewhere obvious. (I just realized the download site is this thread, basically.)
P.S.... and it makes complete sense that moving things through memory is the bottleneck, not operations. But the fact that Jinc is slower than, say, Lanczos4, suggests that there is headroom.

Last edited by NicolasRobidoux; 22nd October 2012 at 13:45.
NicolasRobidoux is offline   Reply With Quote
Old 22nd October 2012, 13:47   #14970  |  Link
NicolasRobidoux
Nicolas Robidoux
 
NicolasRobidoux's Avatar
 
Join Date: Mar 2011
Location: Montreal Canada
Posts: 269
Not totally sure this is useful, but here are some news, courtesy of the GIMP dev. mailing list:
Vladimir Yuzhikov has release his SmartDeblur program under the GPL at
https://github.com/Y-Vladimir/SmartDeblur

Yuzhikov worked on two common cases: blurring due to incorrect focus and blurring due to movement. "Each of you knows very well," he points out, they "are very difficult to repair." In fact, the examples used as samples by Yuzhikov are impossible to fix using current commercial technology. You can see these amazing results here:
http://gizmodo.com/5953601/incredible-csi-enhance-technology-fixes-unfocused-photos
NicolasRobidoux is offline   Reply With Quote
Old 22nd October 2012, 14:05   #14971  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by NicolasRobidoux View Post
Put a flattr button on your download site. This may get you to your first thousand faster than you think.
(And of course it's features first.)
I'm planning to do that (or maybe releasing a "pro" version with additional features) when madVR is feature complete, but not before.

Quote:
Originally Posted by NicolasRobidoux View Post
P.S.... and it makes complete sense that moving things through memory is the bottleneck, not operations. But the fact that Jinc is slower than, say, Lanczos4, suggests that there is headroom.
Roughly calculated, when upscaling, Lanczos4 reads 16 source pixels for every destination pixel, while Jinc3 reads 36 source pixels to do the same. So it's pretty clear that Jinc3 must be slower than Lanczos4.

Ok, reading a square of 36 pixels is not optimal for Jinc3. It should be possible to shave off some of the corner pixels because they'll definitely lie outside of the Jinc circle. But anyway, we'll never get down to those 16 source pixels Lanczos4 is happy with. The simple reason why Jinc3 is slower than Lanczos4 is that we can't 2-pass Jinc, as we can do with Lanczos.

One problem with HLSL is that a true branching "if" instruction is usually slower than doing extra work. That can make optimizations quite difficult.

Quote:
Originally Posted by NicolasRobidoux View Post
Not totally sure this is useful, but here are some news, courtesy of the GIMP dev. mailing list:
Vladimir Yuzhikov has release his SmartDeblur program under the GPL at
https://github.com/Y-Vladimir/SmartDeblur

Yuzhikov worked on two common cases: blurring due to incorrect focus and blurring due to movement. "Each of you knows very well," he points out, they "are very difficult to repair." In fact, the examples used as samples by Yuzhikov are impossible to fix using current commercial technology. You can see these amazing results here:
http://gizmodo.com/5953601/incredible-csi-enhance-technology-fixes-unfocused-photos
This looks very cool, but he definitely needs an AR filter... And it's probably too slow for realtime video playback. Another problem is that you probably need to play with the sliders to get the right parameters for every image, while with video you just want to playback, you don't want to play with sliders all the time. Anyway, looks quite interesting, and maybe in the long run it can be used to create better sharpening algorithms.
madshi is offline   Reply With Quote
Old 22nd October 2012, 14:13   #14972  |  Link
leeperry
Kid for Today
 
Join Date: Aug 2004
Posts: 3,477
Quote:
Originally Posted by madshi View Post
Did he have a plasma at his home? Plasmas at shops are usually put into torch mode
I think he did actually, and there are the usual burn-in issue, highly glossy panel and brightness loss IIRC?

Also, some plasma's output ridiculous CR, such as this 40" LG with 800:1 CR yay, even worse than a TN.

But anyway, I forgot to try mirroring the picture this morning(and it was the last day I could ask for a refund), but apparently many 24p movies have the last 10 vertical percents on the right hand marked as hell, such as this sample: test.mkv

I can mirror it, roll VR's or decoders, move the window in the middle of my CRT and it's still terribly marked, duh! And many/most movies, whatever from the US or EU seem to have this problem IME.

I never noticed how noisy some telecine's were on my previous DLP pj, I guess LCD is indeed merciless with defects

I still don't regret asking for a refund coz the RBE's and backlight buzzing noise were darn annoying anyway. This said, apparently edge LED LCD's also suffer from buzzing noise when their backlight is dimmed but at least their colorimetry should be more stable

Last edited by leeperry; 22nd October 2012 at 15:09.
leeperry is offline   Reply With Quote
Old 22nd October 2012, 14:53   #14973  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Every current Panasonic plasma has a *much* better native contrast than any LCD. Of course LCD uses some tricks to level the playing field like e.g. local dimming, but these tricks can introduce artifacts on their own. LCD backlights lose light, just as much as plasmas do. Current Panasonic plasmas are rated to reach half brightness after 100.000 hours of use. If you leave on your plasma 8 hours every day, it will reach half brightness after 34 years of use. That's much better than CRTs used to be. And LCD backlights won't hold any longer than this, either. There are valid and good reasons for choosing LCD over plasma, but lifetime is not one of them. Just reading reviews in the latest "audiovision" print magazine right now:

LG LCD "65 LM 620 S":
- on/off contrast: 1522:1
- brightness at 10-60° viewing angle: 50%
- contrast at 10-60° viewing angle: 36% (= 548:1 !!)
- 24p quality: 2/3
- motion sharpness: 2/4
- cost: 2800 EUR

Panasonic Plasma TX-P 65 VT 50E:
- on/off contrast: 10963:1
- brightness at 10-60° viewing angle: 80%
- contrast at 10-60° viewing angle: 95% (= 10415:1)
- 24p quality: 3/3
- motion sharpness: 4/4
- cost: 4000 EUR

Ok, price wise this is not a fair comparison, but these are the most comparable reviews in the current issue of "audiovision". You should get an LCD instead of a plasma if (and only if, IMHO):

- you want to use the display as a computer monitor, too (word, programming, browsing)
- you have lots of ambient light in your room and no way (or will) to control it
- power consumption is extremely important to you

If none of these apply, and if you truly value image quality, then you should seriously consider giving Plasma another chance. 'nough said. This is my last comment on LCD vs Plasma in this thread.
madshi is offline   Reply With Quote
Old 22nd October 2012, 14:58   #14974  |  Link
SamuriHL
Registered User
 
SamuriHL's Avatar
 
Join Date: May 2004
Posts: 5,351
I've already decided on a Panasonic plasma for my b-day present in February. Absolutely wonderful consumer products. I can't wait to replace my dying SXRD.
__________________
HTPC: Windows 11, AMD 5900X, RTX 3080, Pioneer Elite VSX-LX303, LG G2 77" OLED
SamuriHL is offline   Reply With Quote
Old 22nd October 2012, 15:03   #14975  |  Link
NicolasRobidoux
Nicolas Robidoux
 
NicolasRobidoux's Avatar
 
Join Date: Mar 2011
Location: Montreal Canada
Posts: 269
Quote:
Originally Posted by madshi View Post
...
One problem with HLSL is that a true branching "if" instruction is usually slower than doing extra work. That can make optimizations quite difficult...
I know all about this, and it makes optimization more interesting

How about this: When I run out of wealthy consulting clients, I'll let you know that I'm ready to take improbably repaid IOUs based on an internet handshake.

Note: With Jinc deblurred with the value that I most recommend with AR (0.981...), you need a 7x7 square to make sure to hit all the locations with nonzero weights. Within this 7x7 square, at most 38 pixels actually have nonzero weights, and it is possible to only use those by exploiting symmetry and arithmetic branching. If you deblur just a little more, so that the disc has radius sqrt(10), which is not a bad thing given that Jinc3 is a little soft anyway, at most 36 pixels actually have nonzero weight.

But of course, the ones that stick out of the 6x6 have minuscule weights, so there is no harm in getting rid of them, esp. since you are actually already approximating via a LUT. If you actually get rid of all those which are at a distance of 3 or more, and use symmetry and arithmetic branching, you are down to 33.

OK, this is not a big reduction.

However, what you can do, is write efficient code that increases the density by 4 in both directions, and either finish off by the built in bilinear interpolation, or with cheap Lagrange-type interpolation. Because increasing the density by 4 in both directions involves only three types of relative positions w.r.t. the original ones when you exploit symmetry, you can program this cheaply (you know the coefficients ahead of time, as well as the relevant pixels, for each type), and then you can use arithmetic branching to "select" the values you need to assemble.

Just sayin'...
NicolasRobidoux is offline   Reply With Quote
Old 22nd October 2012, 15:13   #14976  |  Link
leeperry
Kid for Today
 
Join Date: Aug 2004
Posts: 3,477
Quote:
Originally Posted by madshi View Post
I'm planning to do that
Good, coz it's about time we show you our gratitude

OK, got it! I will check plasma's but I'm not shelling out 4 grand

Anyway, I'm still shocked to see that my problem came from marked telecine's.......darn it, even the new Spiderman movie had this very same problem in the exact same location and several french movies too....the bigger your LCD screen, the more merciless it gets, I can now understand why many ppl hate 24p played in 60Hz. This marked telecine issue is already extremely distracting in a 24Hz multiple..

And for some reason, mVR doesn't want to work in FSW anymore on my CRT....will try to find out why

Last edited by leeperry; 22nd October 2012 at 15:15.
leeperry is offline   Reply With Quote
Old 22nd October 2012, 15:22   #14977  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by SamuriHL View Post
I've already decided on a Panasonic plasma for my b-day present in February. Absolutely wonderful consumer products. I can't wait to replace my dying SXRD.
Yeah, if I weren't a projector guy, I'd get one of those 65" Panasonic plasmas, myself, for serious movie watching. However, maybe for the living room, for sports and news watching, an LCD might be better suited (due to ambient light). Argh, I wanted to *stop* talking about plasma vs. lcd. <sigh>

Quote:
Originally Posted by NicolasRobidoux View Post
How about this: When I run out of wealthy consulting clients, I'll let you know that I'm ready to take improbably repaid IOUs based on an internet handshake.
Haha, thanks, that sounds good to me... I have some funny ideas about some other video processing algorithms, too, where I lack the necessary math skills...

Quote:
Originally Posted by NicolasRobidoux View Post
Note: With Jinc deblurred with the value that I most recommend with AR (0.981...), you need a 7x7 square to make sure to hit all the locations with nonzero weights. Within this 7x7 square, at most 38 pixels actually have nonzero weights, and it is possible to only use those by exploiting symmetry and arithmetic branching. If you deblur just a little more, so that the disc has radius sqrt(10), which is not a bad thing given that Jinc3 is a little soft anyway, at most 36 pixels actually have nonzero weight.

But of course, the ones that stick out of the 6x6 have minuscule weights, so there is no harm in getting rid of them, esp. since you are actually already approximating via a LUT. If you actually get rid of all those which are at a distance of 3 or more, and use symmetry and arithmetic branching, you are down to 33.
I think I'd prefer to keep using 0.981... to keep anti-aliasing power at max. I'm already limiting myself to 36x36 right now. It seems that ignoring those border pixels doesn't harm image quality much.

Quote:
Originally Posted by NicolasRobidoux View Post
However, what you can do, is write efficient code that increases the density by 4 in both directions, and either finish off by the built in bilinear interpolation, or with cheap Lagrange-type interpolation. Because increasing the density by 4 in both directions involves only three types of relative positions w.r.t. the original ones when you exploit symmetry, you can program this cheaply (you know the coefficients ahead of time, as well as the relevant pixels, for each type), and then you can use arithmetic branching to "select" the values you need to assemble.
My math skills are not good enough for this...
madshi is offline   Reply With Quote
Old 22nd October 2012, 15:30   #14978  |  Link
NicolasRobidoux
Nicolas Robidoux
 
NicolasRobidoux's Avatar
 
Join Date: Mar 2011
Location: Montreal Canada
Posts: 269
Quote:
Originally Posted by madshi View Post
...
My math skills are not good enough for this...
Mine are
NicolasRobidoux is offline   Reply With Quote
Old 22nd October 2012, 15:54   #14979  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by NicolasRobidoux View Post
Mine are
Ok, but can you also speak Elvish?
madshi is offline   Reply With Quote
Old 22nd October 2012, 15:55   #14980  |  Link
NicolasRobidoux
Nicolas Robidoux
 
NicolasRobidoux's Avatar
 
Join Date: Mar 2011
Location: Montreal Canada
Posts: 269
Quote:
Originally Posted by madshi View Post
...
I think I'd prefer to keep using 0.981... to keep anti-aliasing power at max. I'm already limiting myself to 36x36 right now. It seems that ignoring those border pixels doesn't harm image quality much.
Yes, but because you restrict the filter to a 6x6, this value may have drifted a little.
The "radius sqrt(10)" deblur is not that different anyway: 0.97651932791303247285

Last edited by NicolasRobidoux; 22nd October 2012 at 15:57.
NicolasRobidoux 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 14:41.


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