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. |
12th May 2017, 00:22 | #43702 | Link | |
Visual Novel Dev.
Join Date: Nov 2015
Location: Bucharest
Posts: 200
|
Quote:
I don't even have a 4k display, barely a 1440p display and GTX 1080 cannot keep up with NGU high for all content. Even using Cubic smooth can strain the GPU on 1440p for 1080 with only a few tweaks enabled...
__________________
Visual Novels! https://www.facebook.com/seventh.heart.studios/ https://audiophile-heaven.blogspot.com/ https://www.youtube.com/watch?v=27VWU3ydRY8 |
|
12th May 2017, 00:24 | #43703 | Link | |
Registered Developer
Join Date: Mar 2010
Location: Hamburg/Germany
Posts: 10,346
|
Quote:
madVR uses higher precision internally though, so thats at least 16 bits per component, if not even 32, that would be 48 or 96MB per image. And madVR also doesn't use only one image, it has several layers of queues to ensure smooth playback. At the default queue size of 8 thats at least 16 images in internal queues (upload+render), and a couple more in the present queue (say also 8 for default exclusive mode). Now add to that GPU decoding, which is almost a necessity for 4K playback, which also has to store a bunch of frames on the GPU to be used as reference frames and some other tasks (say an average of 10, it depends how the video was encoded, can be more or less, note not all of those are references, before anyone complains that 10 is too high of an average, with LAV there is about 4-6 frames in use that are not references). That makes a total of 34 image so far, at 48MB per 4K 16-bit image thats already >1600MB of video memory, not to mention what the OS uses (and some images may even be using 32-bit bpc instead of "just" 16, further increasing it). In short, 2GB is quickly going to run out. You can try to make it work by lowering queues and whatnot, but overall its definitely a limiting factor.
__________________
LAV Filters - open source ffmpeg based media splitter and decoders |
|
12th May 2017, 00:28 | #43704 | Link |
Registered User
Join Date: Jun 2003
Location: Australia
Posts: 190
|
Madshi has stated elsewhere that 3D assumes a 1080p display and thus the left and right eye frames are downscaled to 1920x540 before being combined for Top/Bottom or Interleaved presentation. This is terribly inefficient for those of us with 4k passive 3D displays as we can theoretically view the entire left and right eye frames intact in the 3840x2160 screen.
Would Madshi consider modifying MadVR to provide a 4k and 1080p major option (with SBS, Top/Bottom, Line Interleaved, etc minor options), with the 4k option simply stacking or arranging the left and right eye frames without any pre-scaling? The 4k option is actually simpler, less processing intensive and better quality than the 1080p option as there is no detail reduction through downscaling then upscaling. Whilst it is possible to obtain a similar result using 3D frame-packing, it removes the capability to perform additional processing in the HTPC and relies on the fixed capabilities of the TV. If effort has been made to provide the 3D output options for 1080p displays, does it not make sense to also provide them for 4k displays too for consistency? |
12th May 2017, 01:34 | #43705 | Link |
Registered User
Join Date: Oct 2012
Posts: 7,925
|
looking at the osd line alternative should work properly without downscaling.half res and combining are the last part in the OSD right before final steps. a downscaling algorithm is never used in the OSD but only madshi knows what is used in his renderer.
|
12th May 2017, 01:55 | #43706 | Link | |
Registered User
Join Date: Apr 2017
Location: Hong Kong
Posts: 41
|
Quote:
In 1080p to 2160p double scaling situation, I have High/Medium debanding on, Bicubic60 + AR in Chroma, and NGU Sharp High in Luma, ordered dithering. With defalut CPU/GPU queue settings, my rendering time never exceed 38ms. In 4K HDR playback, since it doesn't need luma double scaling, I use Jinc AR in chroma and get around 30ms in rendering time. I can actually use NGU Sharp High in chroma below 40ms but I can hardly see any difference with Jinc AR. So I rather keep more headroom to be safe. |
|
12th May 2017, 02:14 | #43707 | Link |
Registered User
Join Date: Oct 2012
Posts: 7,925
|
than check your Vram usages.
try a GPU queue of 4 and a present 3 and with this rendertimes should get lower. this kind of worked with my 960 as long as no other program was run like a web browser. my 1060 can do 1080p23-2160 NGU sharp very high (chroma bicubic)and this card is only about 2.5 as fast as a 1050. my queues are default 8 and i need about 3-3.5 GB of Vram with this. i had to replace my 960 2 Gb because it was running out of Vram. |
12th May 2017, 03:09 | #43709 | Link |
Registered User
Join Date: Oct 2012
Posts: 7,925
|
i don't know but not every file is 23p.
and 4 render queue is not save in general deinterlancing will most likely not work. deinterlancing and UHD should be close to impossible on a 2 Gb card anyway. maybe pascal got a lot better at handling Vram limiting situation. but the is no question that madVR uses more than 2 Gb of Vram. i don't know what happens when you run out of Vram with an pascal GPU but the GPU has to do something about this situation. delta color compression is nothing new on nvidia and cards maxwell got this to but this is more for bandwidth not memory. maybe they add a lossy version for such situation. lossless compression over ~10% for the presentation frames should be impossible. |
12th May 2017, 09:36 | #43710 | Link | |
Registered User
Join Date: May 2017
Posts: 3
|
Quote:
No way to watch 3D movies. It would be nice if you could find a solution for this as there is no alternative for a small HTPC than use the Intel graphics. |
|
12th May 2017, 11:19 | #43712 | Link | |
Registered Developer
Join Date: Mar 2010
Location: Hamburg/Germany
Posts: 10,346
|
Quote:
There is a trade quality option to disable that (scale chroma separately, if it saves performance), in which case it would keep the chroma on the original resolution, and scale only Luma.
__________________
LAV Filters - open source ffmpeg based media splitter and decoders Last edited by nevcairiel; 12th May 2017 at 11:22. |
|
12th May 2017, 13:08 | #43714 | Link | |
Registered User
Join Date: Jun 2007
Posts: 656
|
Quote:
|
|
12th May 2017, 13:38 | #43715 | Link | |
Registered User
Join Date: Jan 2009
Posts: 625
|
Quote:
|
|
12th May 2017, 18:49 | #43716 | Link |
Registered User
Join Date: Sep 2013
Posts: 919
|
Question about creating a 3DLUT using DisplayCAL;
What "Source colorspace" should I choose when creating a 3DLUT for UHD movies with 2020->P3? EDIT: Found it. Yes, it makes a difference if I choose BT.2020 or REC.709 as "source colorspace".
__________________
System: i7 3770K, GTX660, Win7 64bit, Panasonic ST60, Dell U2410. Last edited by James Freeman; 13th May 2017 at 11:34. |
12th May 2017, 20:43 | #43717 | Link | |
Registered User
Join Date: Oct 2012
Posts: 7,925
|
Quote:
and to be fair the chroma is >not< at the target resolution when scaling 2160 to 1080. |
|
12th May 2017, 23:12 | #43719 | Link |
Registered User
Join Date: Oct 2012
Posts: 7,925
|
the chroma or luma channel needs to be moved my half a pixel this is only shown in the advanced OSD.
it could be done directly but that is not a normal x2 downscale. and in term of quality it is most likely better to move the chroma. |
13th May 2017, 10:30 | #43720 | Link |
Soul Seeker
Join Date: Sep 2013
Posts: 715
|
A question of image enhancements and upscaling refinement, the difference for me comes from image enhancements, instead of upscaling refinement, I have no idea why. Which one of the options do you guys use for 1080p content (movies), (etc. sharpen edges, adaptive sharpen and more)?
|
Tags |
direct compute, dithering, error diffusion, madvr, ngu, nnedi3, quality, renderer, scaling, uhd upscaling, upsampling |
|
|