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 January 2014, 23:56   #22081  |  Link
trip_let
Registered User
 
Join Date: Sep 2012
Posts: 47
Quote:
Originally Posted by madshi View Post
That is very weird. I also tested with Win8.1 64. But I used a dedicated GTX 650. With 327.23 drivers I get perfectly correct image output with all NNEDI3 related options and with error diffusion. With newer drivers everything is screwed up (either black screen or green tint).
To be clear, previous versions of 0.87.x prior to 0.87.4 have the same behavior as 0.87.4 for me: the black screen on error diffusion and the red/dark thing on NNEDI3 doubling.

That said, I didn't try new drivers with 0.87.x prior to 0.87.4, but that's two variables off and I don't think that combination's going to work.

Wait, oh yeah, there was that OpenCL.dll thing. I'll go try a few different ones of these from the old drivers.

edit: looks like no luck trying different versions of the OpenCL.dll from C:\Program Files\NVIDIA Corporation\OpenCL from different driver installs. Same results.

Last edited by trip_let; 28th January 2014 at 00:09.
trip_let is offline   Reply With Quote
Old 27th January 2014, 23:56   #22082  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by cyberbeing View Post
Could you clarify this separate scaling step for the CbCr channel? Does it use the 'chroma upscaling' setting or the 'image upscaling' setting?
The "chroma upscaling" settings are only for 4:2:0/4:2:2 to 4:4:4 conversion, for nothing else. Once 4:4:4 is reached, all other scaling operations are decided by "image doubling/upscaling/downscaling" options.

Quote:
Originally Posted by cyberbeing View Post
Also when you have NNEDI3 double both luma and chroma, are you working in RGB or YCbCr?
YCbCr because that allows you to use more neurons (higher quality) for Y and less neurons (lower quality) for CbCr. Which is also the default neuron configuration.

Quote:
Originally Posted by cyberbeing View Post
I ask, since a long time ago you changed madVR so the 'chroma scaling' setting was only used for converting sub-sampled YCbCr -> 4:4:4
That is still the case.

Quote:
Originally Posted by cyberbeing View Post
Side question. What is the behavior if you set both NNEDI3 doubling, and quadrupling to 'Always'? Is madVR smart enough to only quadruple for scale factors >2x, or does it only quadruple?
I've added the " - if upscaling is needed" text to the "always" option to indicate that madVR won't (ever) double if the scaling factor is <= 1.0, and madVR won't (ever) quadruple if the scaling factor is <= 2.0.

Quote:
Originally Posted by 6233638 View Post
I don't know if you know this, but NNEDI3 scaling is still not working in the new builds on Nvidia cards.
Haha - you got me there for a second!

Quote:
Originally Posted by 6233638 View Post
In all seriousness, the new dithering and NNEDI3 image doubling looks very nice.
NNEDI can look a bit artificial sometimes, but even when that is the case, the overall image is probably better than without it.
Yeah, I guess I agree with that. What is your opinion about error diffusion? It's hard for me to compare. My projector dithers itself, and is never perfectly smooth. And my LCD monitor is too small to see any difference. So I can't really judge if error diffusion is an improvement over random dithering or not. Technically/scientifically it should be. And the lower noise level is visible on my projector, well in test patterns, at least. Can't really judge smoothness, though...
madshi is offline   Reply With Quote
Old 28th January 2014, 00:02   #22083  |  Link
Farfie
Registered User
 
Farfie's Avatar
 
Join Date: Aug 2012
Posts: 51
Okay so, signature is relevant for this post... reporting performance.

Everything is working for me, but image doubling 1280x720 to 2560x1440 is absolutely killer. With error diffusion on I can forget using anything above 32 neurons. Actually I find it rather funny, with nnedi3 at 64 neurons, mitchel on chroma, with opencl error diff turned on, it can just barely.. barely not handle it. I watched the render queue slowly creep down over the course of 6 seconds before frames start dropping. Pretty funny. I feel like if I just overclocked it a bit, it might just work! (default clocks currently). I might actually start tinkering with that soon.

So what I've settled with is: error diffusion on, nnedi3@32 neurons (720p to 1440p) on luma, jinc 3+ar on chroma (I realize the chroma is overkill but as I said before, going from 64 to 32 neurons kills almost all the overhead, so the GPU is bored. Might as well up the chroma.)
I just want to make sure this is the kind of performance you're expecting.


Opinion section:
Error diffusion vs going from 32 to 64 neurons?
Farfie is offline   Reply With Quote
Old 28th January 2014, 00:07   #22084  |  Link
leeperry
Kid for Today
 
Join Date: Aug 2004
Posts: 3,477
Yeah, NNEDI does wonders with 64 neurons when upscaling VCD to 1080p

Too bad mVR doesn't align chroma properly for MPG1 but that's nothing ffdshow can't fix. I'm just not sure whether it makes any sense to output P216 as I'm doubtful about ffdshow's ability to output >8bit.

I would really appreciate some sort of indication in the OSD to know what kind of exact NNEDI upscaling is going on but anyway, I've got them set to 1.2x/2.4x with 64 neurons and my hopeless low-res files have never looked so crispy
leeperry is offline   Reply With Quote
Old 28th January 2014, 00:07   #22085  |  Link
mimi123
Registered User
 
Join Date: Sep 2011
Posts: 5
Htpc card

Quote:
Originally Posted by djfred93 View Post
The only improvement is deep color who is a smooth gradation.
I think that your projector works in 12bit but can only display 8 bit. Madvr works in 16bit but can only output 8bit.

Maybe the next physical disk will do more than 8bit.
If you want to use OpenCL, you have to choose an ati card until madshi resolve all the problem with the Nvidia card.

My card is a ati radeon hd 5770 OC and i can't do all the OpenCL stuff. I can do Jinc3 AR/Jinc3 AR/ Catmull-Rom AR+SL with debanding. I think a ATI Radeon R9 280X is good. You can compare GPU here.
Quote:
Originally Posted by madshi View Post

That's hard to so. It all depends on which madVR features you want to use, and at which resolutions and framerates. Some algorithms are very hard on the GPU. But the question is if you have to use them. I'd suggest that you first play with the various madVR settings with your current GPU to find out which algorithms you want to use. Then maybe we can recommend a GPU that can do what you need.

My primary play content is Bluray, then 1920*1080 and a frame rate of 23.976Hz.
My display is a JVC X3 projector and 110" screen.

At this moment I have not any discrete card, I use the intel hd3000.

I want to remove banding as much as possible and add a bit of sharpness.
I want a card that can do all this flawlessly and without artifacts and if someday Madvr can send more than 8 bits I want a compatible card.
And finally the card must to be as silent as possible.


Very thanks.
mimi123 is offline   Reply With Quote
Old 28th January 2014, 00:13   #22086  |  Link
Chortos-2
Registered User
 
Join Date: Mar 2012
Location: Riga, Latvia; Cambridge, the UK
Posts: 6
0.87 looks amazing! I have just a few quick questions about nnedi3.

First: do you correct for the half-sample shifts it introduces? (One step goes x x x x -> x y x y x y x y, and the image has now shifted by half a new sample to the left. And don’t forget about chroma positioning!)

Second: I gather from the nnedi3ocl thread that you mirror edge pixels? Doing that causes lines that disappear off the edge of the screen suddenly make a sharp turn after upscaling. Of course, no edge padding is perfect, but I think replicating the edge pixel works better as a default and is normally used as such, probably in your own pre-nnedi3 rescalers too.
Chortos-2 is offline   Reply With Quote
Old 28th January 2014, 00:23   #22087  |  Link
cyberbeing
Broadband Junkie
 
Join Date: Oct 2005
Posts: 1,859
I know you've said the NNEDI3 shifts pixels to the left, but looking at some screenshots it appears the image is shifted both up & left (i.e. diagonally)? Is that expected?
cyberbeing is offline   Reply With Quote
Old 28th January 2014, 00:33   #22088  |  Link
6233638
Registered User
 
Join Date: Apr 2009
Posts: 1,019
Quote:
Originally Posted by madshi View Post
Yeah, I guess I agree with that. What is your opinion about error diffusion? It's hard for me to compare. My projector dithers itself, and is never perfectly smooth. And my LCD monitor is too small to see any difference. So I can't really judge if error diffusion is an improvement over random dithering or not. Technically/scientifically it should be. And the lower noise level is visible on my projector, well in test patterns, at least. Can't really judge smoothness, though...
I need to spend more time with it to draw a "real world" conclusion.
There does seem to be less noise, and looking up close, error diffusion is a lot better than random dithering - but whether that actually translates to a visible difference that's worth the performance hit, I'm not sure yet.

Quote:
Originally Posted by cyberbeing View Post
I know you've said the NNEDI3 shifts pixels to the left, but looking at some screenshots it appears the image is shifted both up & left (i.e. diagonally)? Is that expected?
I'm seeing this too.
6233638 is offline   Reply With Quote
Old 28th January 2014, 00:34   #22089  |  Link
flashmozzg
Registered User
 
Join Date: May 2013
Posts: 77
Quote:
Originally Posted by cyberbeing View Post
I know you've said the NNEDI3 shifts pixels to the left, but looking at some screenshots it appears the image is shifted both up & left (i.e. diagonally)? Is that expected?
Maybe it's because NNEDI scales only one side (X or Y), so when it scales X it shifts left, and when Y - up.
flashmozzg is offline   Reply With Quote
Old 28th January 2014, 00:43   #22090  |  Link
Werewolfy
Registered User
 
Join Date: Feb 2013
Posts: 137
Quote:
Originally Posted by madshi View Post
madVR v0.87.4 released

http://madshi.net/madVR.zip

Code:
* workaround added: NNEDI3 upscaling failed/froze with newer NVidia GPUs
* fixed: NNEDI3 chroma upscaling produced wrong colors with 10bit sources
* got rid of some unnecessary texture sharing
All OpenCL features (except DXVA NV12 processing) should now also work with newer NVidia GPUs. However, the current official NVidia drivers are broken and always produce zero/black output when used with D3D9 interop inside of madVR. I've tried to make it work somehow, but failed. You have to downgrade to NVidia 327.23 drivers (or older) to make the OpenCL features work with NVidia. Would be nice if someone could report the OpenCL driver bug with the current official drivers to NVidia.

Maybe the 4K downscaling performance issue could be fixed, too, but I don't know. Give it a try and let me know.
Thanks it works on my Geforce GTX660! NNEDI3 does a really good job on very aliased videos, it's really impressive. And it hides ringing too. It's definitely someting I'm going to use on some videos. My GPU can handle it for 24 fps SD videos but not 60 fps videos so profiles are going to be very useful

Quote:
Originally Posted by madshi View Post
NNEDI3 image doubling is completely separate from the other scaling. Think of NNEDI3 as manipulating the source video into a different resolution. The other scaling algorithms don't care what the original source video res was, they just look at the resolution of the image doubling output, and go from there.

Jinc currently doesn't support downscaling, and it always handles both X and Y. So if either X or Y need to be downscaled, Jinc can't be used (and should silently be disabled by madVR).
I tested it with a 750*576 video on a 1080p TV screen and it does seem that when I use NNEDI3, Jinc upscales in one direction and Catmull-Rom downscales in the other direction simultaneously because rendering times go up when I select Jinc 8 for upscaling for example and go up too when I select Lanczos 8 for downscaling.

I also noticed that on this video, if you select the option "if scale factor is 2x or bigger" rendering time is lower than with the option "if scale factor is 1.5x or bigger" . NNEDI3 is enabled on both case, if I disable it, rendering time go down.

Last edited by Werewolfy; 28th January 2014 at 01:08.
Werewolfy is offline   Reply With Quote
Old 28th January 2014, 00:50   #22091  |  Link
6233638
Registered User
 
Join Date: Apr 2009
Posts: 1,019
Quote:
Originally Posted by Werewolfy View Post
Thanks it works on my Geforce GTX660! NNEDI3 does a really good job on very aliased videos, it's really impressive. And it hides ringing too. It's definitely someting I'm going to use on some videos. My GPU can handle it for 24 fps SD videos but not 60 fps videos so profiles are going to be very useful
I can see that I'm going to end up with a lot of profiles to manage too. I already had to split up my NTSC and PAL settings.

Quote:
Originally Posted by Werewolfy View Post
rendering times go up when I select Jinc 8 for upscaling for example and go up too when I select Lanczos 8 for downscaling.
For what it's worth, 3-tap is generally considered the best for image quality. (though some argue for 4-tap Lanczos over 3-tap)

Quote:
Originally Posted by Werewolfy View Post
I also noticed that on this video, if you select the option "if scale factor is 2x or bigger" rendering time is lower than with the option "if scale factor is 1.5x or bigger" . NNEDI3 is enabled on both case, if I disable it, rendering time go down.
It's only scaling in one dimension if you select 2x.

1920/720 = 2.67x
1080/576 = 1.88x

Last edited by 6233638; 28th January 2014 at 00:54.
6233638 is offline   Reply With Quote
Old 28th January 2014, 00:53   #22092  |  Link
cyberbeing
Broadband Junkie
 
Join Date: Oct 2005
Posts: 1,859
And to make sure I'm no longer confused, what is the result of the following?

640x360 4:2:0 -> 1920x1080
Chroma Upscaling = NNEDI3
NNEDI3 double Luma = Enabled
NNEDI3 double Chroma = Enabled
NNEDI3 quadruple Luma = Enabled
NNEDI3 quadruple Chroma = Disabled

  • Conversion from 4:2:0 YcbCr (320x180) to 4:4:4 RGB (640x360) with 'chroma upscaling' setting NNEDI3
  • Conversion from 640x360 4:4:4 RGB -> 640x360 4:4:4 YCbCr.
  • Y & CbCr channels doubled to 1280x720 4:4:4 YCbCr with NNEDI3
  • Y channel only doubled to 2560x1440

Is it then:
  • CbCr channels upscaled 1280x720->1920x1080 with 'image upscaling' setting
  • Y channel downscaled 2560x1440->1920x1080 with 'image downscaling' setting
  • Conversion from 4:4:4 YCbCr to RGB

Or is it:
  • CbCr channels upscaled 1280x720->2560x1440 with 'image upscaling' setting
  • Conversion from 4:4:4 YCbCr to RGB
  • 2560x1440 RGB -> 1920x1080 RGB with 'image downscaling' setting

Last edited by cyberbeing; 28th January 2014 at 00:57.
cyberbeing is offline   Reply With Quote
Old 28th January 2014, 01:00   #22093  |  Link
Werewolfy
Registered User
 
Join Date: Feb 2013
Posts: 137
Quote:
Originally Posted by 6233638 View Post
For what it's worth, 3-tap is generally considered the best for image quality. (though some argue for 4-tap Lanczos over 3-tap)
I don't use Jinc 8 usually, it was just to test if madVR upscales or downscales or both after NNEDI3 doubles the resolution on this particular scenario. I use Jinc 4 because it's a little bit sharper without introducing more ringing (to my eyes) but I agree, the difference is negligible. With the new NNEDI3 I will probably use Jinc 3 or a cheaper algorithm.

Quote:
Originally Posted by 6233638 View Post
It's only scaling in one dimension if you select 2x.

1920/720 = 2.67x
1080/576 = 1.88x
Thanks, I think I understand.

Last edited by Werewolfy; 28th January 2014 at 01:02.
Werewolfy is offline   Reply With Quote
Old 28th January 2014, 01:06   #22094  |  Link
dansrfe
Registered User
 
Join Date: Jan 2009
Posts: 1,210
It might be nice to have the ability to import/export profiles so that we can potentially share and cross-check them.
dansrfe is offline   Reply With Quote
Old 28th January 2014, 01:15   #22095  |  Link
mandarinka
Registered User
 
mandarinka's Avatar
 
Join Date: Jan 2007
Posts: 729
In case anybody was wondering - I just tried the openCL dithering on the Llano APU (A8-3850, so the model with best IGP) with 1600MHz/CL11 DDR3 RAM (dual channel).

It definitely works, but perhaps contrary to what one would expect (given it's called APU/Fusion etc), the performance hit is very high.

Testing with an 1856 x 1004 resolution video, scaled to 1920x1080 (windowed mode, not FSE), the rendering time went from 34,5 ms to 57 ms.
Using nnedi3 luma/chroma doubling for a 704x480 -> 1920x1080 upscaling but without OpenCL dithering raises rendering time "just" to 156 ms or so, BTW.
(Note: othere settings: Jinc3AR luma, lanczos3AR chroma)

I guess using the same physical memory for CPU/GPU just doesn't help at all with the DX9/OpenCL interoperation overhead. Has anybody tried with less ancient APU (Trinity/Richland and Kaveri)?

------------------------

Obvious performance costs aside, I find these new features great, awesome job!

Last edited by mandarinka; 28th January 2014 at 01:20.
mandarinka is offline   Reply With Quote
Old 28th January 2014, 01:19   #22096  |  Link
turbojet
Registered User
 
Join Date: May 2008
Posts: 1,840
A few issues on 87.4 that I don't think have been reported:

1. With nvidia 250 and 650 and error diffusion on 87.4 the screen goes blank when paused and takes a few seconds to come back. Is this intentional or a bug?

2. On a 60mhz display when ivtcing with smoothmotion on there is a lot of good frames removed and blended frames left. Almost as if smoothmotion is running before ivtc.

3. Probably known but playing video in film mode is not good. It see's the correct cadence, 2:2, but it still looks like frames are being removed. Would it make sense for it to detect cadence in the middle of the video (if possible, to avoid title which is often a different cadence) and if 2:2 is detected force video if 3:2 is detected force film (and lock in 3:2 with a new option). I know madshi isn't working on ivtc now just want to mention it while it's on my mind.

yok833: you can find lumasharpen at http://dropcanvas.com/l46x9
__________________
PC: FX-8320 GTS250 HTPC: G1610 GTX650
PotPlayer/MPC-BE LAVFilters MadVR-Bicubic75AR/Lanczos4AR/Lanczos4AR LumaSharpen -Strength0.9-Pattern3-Clamp0.1-OffsetBias2.0
turbojet is offline   Reply With Quote
Old 28th January 2014, 01:39   #22097  |  Link
kasper93
MPC-HC Developer
 
Join Date: May 2010
Location: Poland
Posts: 586
Quote:
Originally Posted by madshi View Post
That is rather weird. It shouldn't be that slow. With D3D9 <-> OpenCL interop taken out of the picture, my HD7770 can do error diffusion on 1080p images with > 300fps. Measured in madVR it costs about 12ms rendering time, when talking about 1080p24 playback.
Apparently AMD made great improvements in this matter. On my old HD5870 I have rendering time ~35ms with 720p23. Of course with disabled smooth motion and not upscaled. After upscaling to full hd I have ~93ms, smooth motion (to 60fps) adds ~74ms to the process So OCL performance on my gpu is dreadful, but I bet it's better than NV

I know this post doesn't help you in any way, but I just wanted to show you how performance is scaled when it comes to older generations
kasper93 is offline   Reply With Quote
Old 28th January 2014, 01:47   #22098  |  Link
cyberbeing
Broadband Junkie
 
Join Date: Oct 2005
Posts: 1,859
Quote:
Originally Posted by madshi View Post
Yeah, I guess I agree with that. What is your opinion about error diffusion? It's hard for me to compare. My projector dithers itself, and is never perfectly smooth. And my LCD monitor is too small to see any difference. So I can't really judge if error diffusion is an improvement over random dithering or not. Technically/scientifically it should be. And the lower noise level is visible on my projector, well in test patterns, at least. Can't really judge smoothness, though...
I did a bit of testing on this:

madVR 'opencl error diffusion dither':
  • Flat colors without noise, but sometimes creates a uniform checkerboard pattern.
  • Sharper hard edges.
  • Source banding is more visible.

madVR 'random dither':
  • Noisy flat colors, without any patterns.
  • Hard edges softened by noise.
  • Source banding is less visible.

To get similar visible smoothness on gradients containing banding, I need to use higher debanding settings with 'opencl error diffusion dither' than I do with 'random dither'.
cyberbeing is offline   Reply With Quote
Old 28th January 2014, 01:55   #22099  |  Link
noee
Registered User
 
Join Date: Jan 2007
Posts: 530
Quote:
Originally Posted by mandarinka
Has anybody tried with less ancient APU (Trinity/Richland and Kaveri)?
I have a Kaveri (GIGABYTE GA-F2A88XM-D3H/A10-7700K) build coming together (pieces are arriving this week) and I hope to have some results by the weekend.
noee is offline   Reply With Quote
Old 28th January 2014, 03:56   #22100  |  Link
mrcorbo
Registered User
 
Join Date: Jan 2005
Posts: 171
Quote:
Originally Posted by kasper93 View Post
Apparently AMD made great improvements in this matter. On my old HD5870 I have rendering time ~35ms with 720p23. Of course with disabled smooth motion and not upscaled. After upscaling to full hd I have ~93ms, smooth motion (to 60fps) adds ~74ms to the process So OCL performance on my gpu is dreadful, but I bet it's better than NV

I know this post doesn't help you in any way, but I just wanted to show you how performance is scaled when it comes to older generations
Enabling any of the OpenCL options (except process NV12 surfaces) causes my GPU usage to spike to 99% with full clocks on my HD 6970 (win8.1/Cat 13.12) even when everything else is turned off/down as low as it will go.

Is anyone else seeing this issue with a 69XX Radeon? Up until now my GPU has handled any kind of processing that madVR could throw at it without breaking a sweat.
mrcorbo 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 18:52.


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