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. |
19th November 2016, 10:32 | #40241 | Link | ||||||||||
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
Do you have an Optimus system?
Do you have any 3rd party software running which might also access the GPU, like f.lux, FRAPS, GpuZ or something like that? Or maybe some sort of debugger or something? Quote:
Quote:
About slow render times: Are you using any custom flush settings? Have you tried resetting madVR to default settings as a test? Quote:
NGU High runs slower for you than expected. But not as slow as huhn reported. Quote:
Quote:
Quote:
Are you saying that the option "use D3D11 for presentation" makes problems when using NGU, but makes no problems when using other image doublers? Quote:
Quote:
Generally, NGU was optimized for upscaling content that was downscaled by using gamma light from the original highres master. If the studio (or custom encode) actually downscaled things in linear light, NGU will produce lines that are too thin. So maybe that's what's happening here. I was planning to offer a "linear light" switch for NGU at some point, which will be optimized for images that were downscaled using linear light. But it'll be some time before I get to that. In all my tests with downscaling a "groundtruth" image (in gamma light, though!) and then upscaling it with NGU, NGU always produced lines in exactly the right width. They were never too thin. Are you sure you aren't just used to bloated lines? Quote:
Quote:
|
||||||||||
19th November 2016, 10:38 | #40242 | Link | |
Registered User
Join Date: Mar 2009
Posts: 3,646
|
Quote:
|
|
19th November 2016, 11:04 | #40243 | Link |
Registered User
Join Date: Sep 2014
Posts: 280
|
Performance Tests
Hey Madshi!
Thank you very much for the new madvr version. I did some quick performance test: Test System: i5 6600, 16GB RAM, RX480 8GB, Crimson 16.11.4, Windows 10, madvr 0.91.b File: Avatar 1920x1080p, 23,976, 8 Bit, MKV file upscaled to 3840x2160p Old Settings: Chroma: SuperXBR 125, AR Luma: image doubling Superxbr 25 AB + Super Res 2 = 28ms Changed Luma Doubling to NNEDI64, no Super Res: =30ms Changed Luma Doubling to NGU High: =84ms Changed Luma Doubling to NGU Med: =36ms Setting Chroma and Luma doubling to NGU Med: =60ms Setting Luma and Chroma to both use doubling @ NGU Med: =97ms Using Chroma Rconstuction Soft and Luma NGU med: =36ms Using Chroma Rconstuction Soft and NNEDI64: =32ms Using Chroma Rconstuction Soft and NNEDI128: =61 ms Conclusion: NGU is to performance hungry for my system. I can only use it with medium setting, with chroma set to SXBR or RecSoft
__________________
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 Last edited by Sunset1982; 19th November 2016 at 11:06. |
19th November 2016, 11:09 | #40244 | Link |
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
Ok, two RX480 users now with very bad performance results. I smell a pattern. So my current impression is:
1) NVidia: a bit slower than expected. 2) RX480: a *lot* slower than expected. 3) All older AMD generations: just as fast as expected. Will do some tests on my NVidia soon. Don't have an RX480 to test with, though. |
19th November 2016, 11:14 | #40245 | Link | ||||
Registered User
Join Date: Mar 2009
Posts: 3,646
|
Quote:
I'll post some up later. Quote:
really constricting particular areas so that they no longer match the original line thickness and I don't like that. Quote:
Quote:
|
||||
19th November 2016, 11:19 | #40247 | Link | |||
Registered User
Join Date: Oct 2012
Posts: 7,903
|
Quote:
it is most likely independent of the GPU queue it was still very high Quote:
GPU Vram usages in GPU-z is not reliable it could be out of ram even with numbers like 3.6 GB of 4 GB. Quote:
ShowRenderStep works like a handbrake for NGU. FHD -> UHD 23p luma only i get about 26-30 ms with NGU med now not sure how much NGU is eating. NGU high is using ~100 ms |
|||
19th November 2016, 11:25 | #40248 | Link | |
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
Quote:
You reported earlier 22ms for NGU med. So it seems to be even slower now?? Or are you testing a different resolution now? The most important thing is not the speed by itself, but the speed compared to NNEDI3, to have a good reference point. So how do NGU med/high compare to NNEDI3 now? |
|
19th November 2016, 11:42 | #40250 | Link | |
Registered User
Join Date: Oct 2012
Posts: 7,903
|
Quote:
FHD -> UHD normal rendertimes NGU med 25-30 ms nnedi64 32 ms NGU high ~80-100ms nnedi3 256 110ms nnedi3 128 55ms NGU med is about 10 ms faster now since i disabled showrenerstep and reset all settings. NGU high doesn't seem to care. |
|
19th November 2016, 11:46 | #40251 | Link |
Registered User
Join Date: Sep 2014
Posts: 280
|
ok, had showrendersteps activated too.
Performance is a bit better now: can use NGU med with chroma and luma. 30ms choma reconstruction soft and double luma with NGU med: 23ms choma reconstruction soft and double luma with NGU high: 59ms sadly, NGU high is still to much for my rx480. I will do some more testings tomorrow. To bad, I have to work tonight...
__________________
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 |
19th November 2016, 11:48 | #40252 | Link | ||
Registered User
Join Date: Mar 2009
Posts: 3,646
|
Quote:
Quote:
Madshi, http://screenshotcomparison.com/comparison/191145 This is with Lanczos 3 as second upscaler, normally I'd use Jinc but w/e, this kinda helps exacerbate the issues anyway.. This is just clean NNEDI3 64 vs NGU HQ SE1, SE2 is just too blurry to be used for any real comparison. That sucks, my 960 breezes through it, doesn't use much more performance than NNEDI3 64 neurons, extra few ms render time that's it =/ I've only tested 480 > 1080 though.. Last edited by ryrynz; 19th November 2016 at 11:53. |
||
19th November 2016, 11:48 | #40253 | Link | |
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
Quote:
|
|
19th November 2016, 11:52 | #40254 | Link | ||
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
Quote:
Quote:
|
||
19th November 2016, 11:57 | #40255 | Link | |
Registered User
Join Date: Oct 2012
Posts: 7,903
|
Quote:
"don't flush" no change. "use separate device for presentation" no change using D3D9 present in advance with FSE instead of D3D11 gave me maybe an 1 ms but NGU rendertimes are unstable anyway. |
|
19th November 2016, 12:00 | #40256 | Link | |
Registered User
Join Date: Mar 2009
Posts: 3,646
|
Quote:
SuperRes4 looks even nastier but then it doesn't thin the lines, naturally. So you could see why I'd prefer NNEDI3, after Awarpsharp etc has taken care of it I get a very nice sharp & clean image. |
|
19th November 2016, 12:05 | #40257 | Link |
Registered User
Join Date: Sep 2014
Posts: 280
|
some quick tests:
Chroma & Luma doubling both @ NGU med: 32ms Chroma & Luma doubling both @ NGU high: 79 ms Chroma & Luma doubling both @ NNEDI32: 25 ms Chroma & Luma doubling both @ NNEDI64: 36ms Chroma @ NGU med & Luma doubling @ NGU High: 60ms luma & Chroma doubling set to NGU high (luma) and NGU med: 78 ms So, still NGU high is no option on rx480 yet.
__________________
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 |
19th November 2016, 12:13 | #40258 | Link |
Registered User
Join Date: Mar 2009
Posts: 3,646
|
I wouldn't even bother with chroma doubling with this in it's current state, but that's me.. not sure if it'll make enough of a change in performance for you to run it. But like madshi said, there isn't a huge visual difference between medium and high. But I guess you're just testing rather than anything else..
|
19th November 2016, 12:20 | #40259 | Link | |||
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
Quote:
In order to properly upscale low-res images to hi-res in a way to get as near to the groundtruth as possible, NGU needs to decide which lines in the source were likely bloated through downscaling and were much thinner in the original hi-res image, and which lines were already thick in the hi-res source. NGU can't just thin all lines by a certain factor (as awarpsharp does). While that might produce pleasing results for very low quality anime sources with high compression artifacts, it would produce results very far from the groundtruth for high quality sources. Quote:
Quote:
I'll try my luck with D3D11 shaders, maybe that solves the issue. |
|||
Tags |
direct compute, dithering, error diffusion, madvr, ngu, nnedi3, quality, renderer, scaling, uhd upscaling, upsampling |
Thread Tools | Search this Thread |
Display Modes | |
|
|