madVR - high quality video renderer (GPU assisted) [Archive] - Page 448

DarkSpace

2nd February 2014, 17:43

Assuming you're using the latest drivers for your card, you're not supposed to increase the NNEDI factor, but decrease it. That number the lower it is, the sooner NNEDI will be used, not the opposite.
What do you mean by "NNEDI factor"?
If you mean the image scaling factor, I strongly believe it should be dependent only on my madVR Image Doubling settings (only if factor is 2x or greater in this case), and if you mean how many neurons I use, I tried 256 neurons and 16 neurons, and both gave the same result.

iSunrise

2nd February 2014, 18:50

What do you mean by "NNEDI factor"?
If you mean the image scaling factor, I strongly believe it should be dependent only on my madVR Image Doubling settings (only if factor is 2x or greater in this case), and if you mean how many neurons I use, I tried 256 neurons and 16 neurons, and both gave the same result.
He is refering to the fact that you stated "I also tried to increase the scaling factor to something a little more than 2, but the issue stayed the same." which leads someone to believe that you meant the scaling factor description from the NNEDI3 double/quadruple settings. After re-reading that about 10 times and since you cannot go any higher than 2.0x without enabling "always...", you seem not to be refering to the scaling factor settings at all, but the actual scaling factor that you applied manually when you played back a clip.

If whether NNEDI3 4:2:2 -> 4:4:4 chroma upscaling nor image/chroma doubling/quadrupling does make a difference on low-resolution clips there is clearly something wrong, though.

DarkSpace

2nd February 2014, 19:33

He is refering to the fact that you stated "I also tried to increase the scaling factor to something a little more than 2, but the issue stayed the same." which leads someone to believe that you meant the scaling factor description from the NNEDI3 double/quadruple settings. After re-reading that about 10 times and since you cannot go any higher than 2.0x without enabling "always...", you seem not to be refering to the scaling factor settings at all, but the actual scaling factor that you applied manually when you played back a clip.
Oops. I can see what you mean, and you're right, I meant the video scaling factor. I edited my original post to prevent further misunderstandings.

If whether NNEDI3 4:2:2 -> 4:4:4 chroma upscaling nor image/chroma doubling/quadrupling does make a difference on low-resolution clips there is clearly something wrong, though.
I only tested Image scaling, because I'm not sure how obvious the difference in 4:2:0-to-4:4:4 would be, seeing how a Lanczos-3 AR is used after NNEDI automatically to compensate for the shift, while with Image Upscaling, I could specify which algorithm to use after Doubling.

Asmodian

2nd February 2014, 19:45

Specs - GTX 770 2Gb with 327.23 drivers and Windows 8.1.

I use NNEDI3 Chroma Upscaling (64 neurons) with NNEDI3 Chroma & Luma Image Doubling (16 neurons) for any SD or 720p content on my 1080p LCD TV. I keep the same settings for 1080p content, so I guess that means Image Doubling is not used for 1080p, but NNEDI3 Chroma Upscaling (64 neurons) is still used. I also use OpenCL dithering, MPC-HC, LAV Video - CUVID decoder, ArgyllCMS 3DLUT.

I see that many people prefer JINC3AR for Chroma Upscaling when using NNEDI3 Image Doubling - is that a higher quality combination than using NNEDI3 for both - Image Doubling AND Chroma Upscaling? To my eyes, 1080p videos look better with NNEDI3 than with JINC3AR, but I could be wrong...

I mentioned I was using Jinc3AR instead of NNEDI3 chroma upscaling but that was only because that was where I was maxing out my card with the NNEDI3 image doubling options I was using. Using suggestions from both madshi and 6233638 I did more testing; turning off NNEDI3 chroma doubling and turning on NNEDI3 chroma scaling. I now also believe NNEDI3 chroma scaling and only luma doubling is a better tradeoff. I agree with them (and you): I really like NNEDI3 chroma upscaling (4:2:0 -> 4:4:4).

My TV is set to 4:2:2 (HDMI to HDMI, 0-255) because 4:4:4 does not allow me to calibrate CMS properly. Does that mean I am not getting the benefits of Chroma Upscaling, OpenCL dithering, and Image Doubling??? Maybe its a placebo, but NNEDI3 Chroma Upscaling with NNEDI3 Image Doubling and OpenCL dithering + extremely accurate 3DLUT looks absolutely stunning!

I do not understand how your TV would not allow you to calibrate CMS properly when in 4:4:4, are you using it for content other than madVR? If you are only using madVR you can do all the color management with ArgyllCMS's 3DLUT in madVR and leave the TV in a "just display what you get" mode. That said I bet you do get some benefits now but not the full glory. :)

NNEDI3 is shown to have some "other artifacts" but I have not seen any... What are they? When would I see them?

The NNEDI3 artifacts are odd details at are "different" after NNEDI3 compared to the original, they are really hard (impossible?) for me to notice unless comparing to the original. madshi posted some examples a few pages back. (http://forum.doom9.org/showthread.php?p=1665373#post1665373)

XMonarchY

2nd February 2014, 20:23

I do not understand how your TV would not allow you to calibrate CMS properly when in 4:4:4, are you using it for content other than madVR? If you are only using madVR you can do all the color management with ArgyllCMS's 3DLUT in madVR and leave the TV in a "just display what you get" mode. That said I bet you do get some benefits now but not the full glory. :)

Thank you for explaining. My Samsung LN40B550 40" CCFL SPVA LCD simply locks CMS controls when in PC Mode, which is the mode that uses 4:4:4. In that mode color accuracy is bad enough that I will need about 4300 patches to achieve the same results as with 4:2:2 mode. Almost all Samsung TVs are like that in PC 4:4:4 Mode.

At the same time, I do not use Movie 4:2:2 mode for several reasons, one them being text sharpness. In Standard 4:2:2 mode that I use, I disable sharpness completely, but text remains much much better than in Movie 4:2:2 mode, where text is just plain bad. Let's just say that there is very little difference between Standard 4:2:2 mode and PC 4:4:4 mode, but there is a huge difference between Movie 4:2:2 Mode and PC Mode.

In Standard 4:2:2 mode I calibrated CMS manually (all of CMS readings were under dE 3) and then applied 3DLUT profile on top, using 2527 patches. The result was amazing - about 3 saturation sweeps as high dE 1.1, one ColorChecker sweep as high as dE 1.5, and all the rest is under 1.0 - measured by i1 Display 3.

I still see a stunning image using madVR rendering though. I could try to create a 3DLUT for PC 4:4:4 mode using 4300 patches over night - can't think of a better way to test the new ArgyllCMS 1.6.3 stable release (not beta) and dispcalGUI 1.7.1.6.

EDIT: Actually, my TV can't use the 48Hz judder-free mode for 24p content in 4:4:4. I would have to trade judder-free smoothness for 4:4:4 subsampling... not sure which one is worth more to me...

Mano

2nd February 2014, 20:23

is it normal for mpc to use 1.5gb ram then another 500mb in exclusive mode playing a 322mb mkv anime?

Only if you have Avisynth or something similar running.i dont think i have any program like that running.

Do you use XySubFilter with a very high CPU queue setting of 64 in madVR?

i only have LAV Video, Audio decode and LAV spliter.

huhn

2nd February 2014, 21:01

EDIT: Actually, my TV can't use the 48Hz judder-free mode for 24p content in 4:4:4. I would have to trade judder-free smoothness for 4:4:4 subsampling... not sure which one is worth more to me...

just force 60 hz and use smooth motion best you can do with samsung.

Asmodian

2nd February 2014, 21:09

@XMonarchY

It sounds like you understand your TV very well and I trust you have it correctly configured for your conditions and preferences. My old 2010 LG Plasma's CMS did more harm than good.

I use NNEDI3 Chroma Upscaling (64 neurons) with NNEDI3 Chroma & Luma Image Doubling (16 neurons) for any SD or 720p content on my 1080p LCD TV.

That said I doubt NNEDI3 16 chroma doubling is worth it, I suggest trying out NNEDI3 Chroma Upscaling 32 and Luma only Image Doubling at 32 or 64 if you can.

just force 60 hz and use smooth motion best you can do with samsung.
If the TV is doing BFI or something else better than smooth motion it might be worth 4:2:2. I quite like smooth motion on a 60Hz display but it does add a lot of ms to rendering times, especially when using OpenCL error diffusion dithering.

XMonarchY

2nd February 2014, 23:12

@XMonarchY
That said I doubt NNEDI3 16 chroma doubling is worth it, I suggest trying out NNEDI3 Chroma Upscaling 32 and Luma only Image Doubling at 32 or 64 if you can.

From what I understand - Image Doubling only works for content that is below the native resolution of the display, correct? My TV is set to its native 1080p, so Image Doubling would NOT be applied to any 1080p content, right? Since Image Doubling is ignored for 1080p content, then only NNEDI3 Chroma Upscaling is being used in such cases. My PC handles 64 neurons for Chroma Upscaling without any frame losses.

I see a considerable image improvement in 720p and below resolution content when using NNEDI3 Chroma Upscaling (64 neurons) in combination with NNEDI3 Image Doubling (Chroma & Luma) with 16 neurons. I also tried using 32 neurons for Chroma Upscaling and upping Image Doubling Chroma and Luma to 32 neurons, but that was a stutter-fest. But I see what you're saying - since Chroma is always upscaled with NNEDI3, there may be no point in doubling Chroma, which would also decrease performance. Instead, I could use Chroma Upscaling with only Luma Doubling but with higher neurons if I disable Chroma Doubling.

Thanks again! I really wish there was a sure way to tell what looks better with some kind of picture analysis benchmark.

xabregas

2nd February 2014, 23:43

Hi

First of all let me say i`m glad to see madshi still surprises me with new ways to improve PQ, and NNEDI3 is awesome, better than jinc i must say, i use chroma NNEDI3 32 neurons with lanczo 3 AR Luma and the PQ is amazing, better than both in jinc 3 AR. Ofc i cant use NNEDI3 chroma + jinc Luma because i will burn my graphic card LOLOL

Anyway, i dont understand the new Image Double thing, meaning is Image Doubling applied in 720p videos upscaled to 1360x768 screen resolution?

TIA

Stereodude

3rd February 2014, 00:35

Anyway, i dont understand the new Image Double thing, meaning is Image Doubling applied in 720p videos upscaled to 1360x768 screen resolution?
That depends on your settings, but probably not.

Stereodude

3rd February 2014, 00:41

From what I understand - Image Doubling only works for content that is below the native resolution of the display, correct?
Basically, but it depends how you have it set. For small scaling amounts like 720p to 768p it doesn't make any sense. You would double it from 720p to 1440p and then downscale it to 768p using the downscaling settings.

realghostvids

3rd February 2014, 01:20

I just updated all the stuff I was using for MPC including madVR and now I'm getting a flickering issue with 60fps videos only. When I fullscreen the video a green box the size of the video will flicker on and off randomly as the video plays like this:

http://fc05.deviantart.net/fs71/f/2014/033/7/e/mpc_green_2_by_realghostvids-d74valb.jpg

That's not an actual screenshot of the issue, just something I did in Photoshop to illustrate the problem.

If it's a 1920x1080 60fps video then the entire screen will flicker green.

I haven't had this issue with previous versions of madVR.

Q-the-STORM

3rd February 2014, 05:18

For 24fps FILM type contents, it should be fast enough for GTX780(900MHz,2304 CUDA, 384-bit GDDR5-6000), which is theroetically faster than HD7970 (925MHz,2048 GCN,384-bit GDDR5-5500)...

However I have posted it previously that HD7970 is even unable to handle DVD 720x480i60 16:9 contents to 1080p60 output with following options:

1. DXVA2 MPEG2 VLD (very minor mem bandwidth used on GPU)
2. Vector-adaptive deinterlacing + frame rate doubling
3. NNEDI3 Image Doubling (default --> Luma 64, Chroma16)
4. All Jinc3AR upscaling
5. debanding_with_Angle_detection

It is just fast enough to handle 54 ~ 55 fps. With OpenCL dithering, it might not even approach 50 fps ...

I only play up to 29.97p (old tv shows sometimes run at progressive 29.97) so it should be fine...

though I would like to play with SVP a little bit.. does the image go through madVR first, or SVP? nnedi3 @ 23.976 + SVP to 60fps afterwards could work, but if SVP converts to 60fps first, then NNEDI3 at 60fps would (theoretically) be too much for the GTX 780 Ti...

cyberbeing

3rd February 2014, 05:46

NNEDI3 4:2:0->4:4:4 chroma upscaling is broken on my GTX 770 with Video Mode deinterlacing. The screen turns a greenish yellow. Film Mode is unaffected.

Can anyone else with NVIDIA reproduce this?

Asmodian

3rd February 2014, 05:54

I can on a GTX Titan and madshi already knows about it, reported by iSunrise (http://forum.doom9.org/showthread.php?p=1665051#post1665051):
Ok, it seems I can reproduce this problem in video mode on my 9400 mainboard. Shouldn't be too hard to fix.

cyberbeing

3rd February 2014, 06:10

Ah, thanks for pointing that out Asmodian. I didn't realize before iSunrise was reporting this problem only with Video mode deinterlacing, but indeed his sample was interlaced.

NicolasRobidoux

3rd February 2014, 08:50

I vaguely remember that for pixel art the 4-lobe version of (ImageMagick's) EWA LanczosSharp was better than the 3-lobe version. Will dig out posts when I have a minute.

pie1394

3rd February 2014, 10:04

I only play up to 29.97p (old tv shows sometimes run at progressive 29.97) so it should be fine...

though I would like to play with SVP a little bit.. does the image go through madVR first, or SVP? nnedi3 @ 23.976 + SVP to 60fps afterwards could work, but if SVP converts to 60fps first, then NNEDI3 at 60fps would (theoretically) be too much for the GTX 780 Ti...

If you mean the external SVP, it definitely requires all 60 fps workload on madVR. With the SmoothMotion feature in madVR, NNEDI3 workload only happens on source frame's processing --- unless there is frame-rate-doubling deinterlace in the processing chain..

Have tested various samples to guarantee source 60fps processing performance on HD7970 --> #22356 (http://forum.doom9.org/showthread.php?p=1665685#post1665685)

I just could say that currently it at least needs a card cost about US$300 to deal with madVR's crazy demanding features. With further optimization, I guess it might be dropped to near US$200... :D

GTX780Ti(928MHz , 2880 CUDA, 384-bit GDD5-7000) is even more powerful than most GTX or HD7970GHz cards. It is a fully functional GK110 chip. Next generation 20nm Maxwell GPU should be 100% more powerful than it -- if the fully function chip contains 6144 CUDA...

R9 290X(Hawaii core) has the on-pair 2816 GCN units, 512-bit GDDR5-5000 bandwidth to be compared. In addition each GCN's primitive processing capabibility is doubled from R9 280X(=HD7970GHz, Tahiti core) --- 4 vs 2 per clock cycle.

THX-UltraII

3rd February 2014, 10:21

wolfman2791

3rd February 2014, 10:48

Noob here. Just a FYI if it helps:

My rig is MSI GTX 660 2GB OC, I5-3470 running at 40x, 2x4 1866 DRAM and a MSI mobo;

NNED13 brings my system to its knees :D On 720p anime, I can run 16 neur chroma, 16 neur luma doubling and 16 neur luma quadrupling with Lanzcos AR. I have none of "trade quality for performance boxes" checked. I also run smooth motion. BUT, the rendering times are 40+ ms :scared:

In order to get the rendering times down a bit, i run Jinc 3AR on Chroma while keeping the rest the same as above... it gets me down to 35 ms :(

Basically, NNED13 is killing me lol.

Ver Greeneyes

3rd February 2014, 11:20

A bit of wishful thinking, but when Nvidia fix the D3D9-OpenCL interop bugs, I hope they also spend some time optimizing their drivers for the algorithms madVR uses like they do for many games ;)

James Freeman

3rd February 2014, 11:41

A bit of wishful thinking, but when Nvidia fix the D3D9-OpenCL interop bugs, I hope they also spend some time optimizing their drivers for the algorithms madVR uses like they do for many games ;)

Dream on.

They still don't have a proper way to switch HDMI between Full/Limited range.
Although it does appear that in the latest version 60Hz is 0-255 and 59Hz is 16-235, without the need of external soft.

realghostvids

3rd February 2014, 12:59

I just updated all the stuff I was using for MPC including madVR and now I'm getting a flickering issue with 60fps videos only. When I fullscreen the video a green box the size of the video will flicker on and off randomly as the video plays like this:

[img]http://fc05.deviantart.net/fs71/f/2014/033/7/e/mpc_green_2_by_realghostvids-d74valb.jpg[img]

That's not an actual screenshot of the issue, just something I did in Photoshop to illustrate the problem.

If it's a 1920x1080 60fps video then the entire screen will flicker green.

I haven't had this issue with previous versions of madVR.

Turns out the issue wasn't madVR, it was MPC-BE. I swapped back to MPC-HC and it's all good now.

DragonQ

3rd February 2014, 14:11

Dream on.

They still don't have a proper way to switch HDMI between Full/Limited range.
Although it does appear that in the latest version 60Hz is 0-255 and 59Hz is 16-235, without the need of external soft.
Yep, which is ludicrous decision because it doesn't help those viewing 50 fps material. It would be so god damned easy to add a toggle to the nVidia Control Panel, just like AMD did years ago, but they refuse to do it for whatever reason.

nVidia also can't get HDMI audio streaming right, which prevents WASAPI from being useful and requires third party software to get AVRs to work properly.

nevcairiel

3rd February 2014, 14:23

I have been using WASAPI with HDMI audio for years, and never had any issues. Maybe your AVR is at fault? Or whatever audio renderer you are using? :)
To offset your biased opinions, I was running an AMD card (a 7750) for a while, and had various issues over the period, so I replaced it with a NVIDIA again, and all is smooth sailing now!

In any case, the range issue is so easily fixed, i don't get why getting annoyed by it is even worth it.

pie1394

3rd February 2014, 15:16

When I choose NNEDI3 as Chroma Upscaler MPC-BE crashes after some time of playback (10 minutes or so). When I choose another Chroma Upscaler eg. JINC3+AR i have no problems at all. How can I find out what could be the problem?

sidenotes:
- It does not matter if I choose 16, 32, 64, 128 or 256 neurons
- My GPU (oc-ed 280x) can handle NNEDI3 easily
- I use the latest AMD driver (dec. 2013)
- I use the latest version of LAV filters
- I use MPC-BE v1.3.1.0 build 4146
- unchecked all boxes in trade quality for performance expect 'use random dithering instead of OpenCL error diffusion' (checking this option makes my video stutter like hell!)
- Running on W7 X64 with all latest updates installed
- Latest version of DirectX
- Content is 23,976 Blu-Ray outputted at 3840x2160@23 to my 3840x2160 display device

Maybe you can try the debug build version to capture the crash point -- althrough the log file will be terrible big. :D

Is it fine with 3D game + the most GPU stressful setting on 4K output? I suspect it could be the overheated issue since it is random and the Jinc3AR mode is fine... Maybe you can monitor the GPU's temperature during playback to verify that...

My card's GPU is the same with yours in theory and the playback setup is also similar except the 4K output, MPC-BE 1.3.0.3. Yesterday I did the very GPU stressful tests with NNEDI3 128 neurons Chroma 2x scaling, all Jinc3AR up-scaling, Debanding_with_angle_detection settings. Played a DVD 576i50 to 1080p50 more than 1 hour and it did not show the issue which happened to you.

michkrol

3rd February 2014, 15:29

When I choose NNEDI3 as Chroma Upscaler MPC-BE crashes after some time of playback (10 minutes or so). When I choose another Chroma Upscaler eg. JINC3+AR i have no problems at all. How can I find out what could be the problem?
Checking if it happens in MPC-HC or any other player would be the first step. Especially after realghostvids' report that MPC-BE breaks playback for him. Another simple thing to do is checking system RAM usage at start and when it crashes and GPU VRAM usage (with GPU-Z or similar). For other suggestions you'll have to wait for someone more knowledgeable than me.

(cut)
On 720p anime, I can run 16 neur chroma, 16 neur luma doubling and 16 neur luma quadrupling with Lanzcos AR. I have none of "trade quality for performance boxes" checked. I also run smooth motion. BUT, the rendering times are 40+ ms :scared:

What resolution are you using for output? Setting NNEDI quadrupling to "always" is pointless. If you have it set to "only if scaling factor is ...x" then it's probably not used for 720p or you have quite big screen ;)
Not sure how powerful your GPU is, but you might get better quality with the first 5 trade quality options enabled (default setting) and using 32 neurons for luma doubling and NOT using chroma doubling at all, if your GPU can handle it.

EDIT: Looks like I don't type fast enough.

DragonQ

3rd February 2014, 15:38

I have been using WASAPI with HDMI audio for years, and never had any issues. Maybe your AVR is at fault? Or whatever audio renderer you are using? :)
The nVidia "silent stream" bug is pretty well documented. It probably doesn't affect all AVRs but is common. Basically when there is no audio transmitted, the AVR will "lose" the audio stream. It will then "find" it again once audio starts playing, but that causes a gap of a few seconds. This makes gapless music playback impossible, as well as other less important issues. I use AVR Audio Guard to solve this problem, which basically makes sure audio is always sent to the AVR (by mixing in an all-zero audio track). However, this doesn't work for WASAPI because it relies on mixing.

I do not know why only nVidia cards suffer from this problem.

EDIT: Oh apparently it is an issue with Intel HDMI too. AMD seem to be the only company with a proper solution.

In any case, the range issue is so easily fixed, i don't get why getting annoyed by it is even worth it.
It used to be very easy to fix, but nVidia actually made changes to their drivers a month or two back that makes the third-party range setters not work any more.

nevcairiel

3rd February 2014, 16:21

Basically when there is no audio transmitted, the AVR will "lose" the audio stream. It will then "find" it again once audio starts playing, but that causes a gap of a few seconds.

Not sure from the top of my head if thats the player and its audio renderer I use, or my Pioneer AVR, but I at least don't suffer from this.
Its easy enough to fix in the player or audio renderer, as long as you're not bitstreaming.

It used to be very easy to fix, but nVidia actually made changes to their drivers a month or two back that makes the third-party range setters not work any more.

The tool which comes with madVR seems to still work fine for me.

THX-UltraII

3rd February 2014, 17:00

Maybe you can try the debug build version to capture the crash point -- althrough the log file will be terrible big. :D

Is it fine with 3D game + the most GPU stressful setting on 4K output? I suspect it could be the overheated issue since it is random and the Jinc3AR mode is fine... Maybe you can monitor the GPU's temperature during playback to verify that...

My card's GPU is the same with yours in theory and the playback setup is also similar except the 4K output, MPC-BE 1.3.0.3. Yesterday I did the very GPU stressful tests with NNEDI3 128 neurons Chroma 2x scaling, all Jinc3AR up-scaling, Debanding_with_angle_detection settings. Played a DVD 576i50 to 1080p50 more than 1 hour and it did not show the issue which happened to you.Where can I find the debug version?

I don t game or have any games installed on my system. I m 100% sure it is no overheating issue. JINC3+AR runs at 48 degrees and 30%GPU usage. NNEDI3 16 runs at 51 degrees celsius and 36%GPU usage.

THX-UltraII

3rd February 2014, 17:01

Checking if it happens in MPC-HC or any other player would be the first step. Especially after realghostvids' report that MPC-BE breaks playback for him. Another simple thing to do is checking system RAM usage at start and when it crashes and GPU VRAM usage (with GPU-Z or similar). For other suggestions you'll have to wait for someone more knowledgeable than methxz. Ill test that tonight too

huhn

3rd February 2014, 17:42

Where can I find the debug version?

I don t game or have any games installed on my system. I m 100% sure it is no overheating issue. JINC3+AR runs at 48 degrees and 30%GPU usage. NNEDI3 16 runs at 51 degrees celsius and 36%GPU usage.

you may just run out of ram? i mean there is a magical 4gb wall for 32 bit programms.

for debug run the "activate debug mode.bat" file in madvr folder. now when playback is strted madvr will create a HUGE file on your desktop

XMonarchY

3rd February 2014, 19:23

THX-UltraII

3rd February 2014, 19:51

you may just run out of ram? i mean there is a magical 4gb wallbut would we not all have this same problem??

iSunrise

3rd February 2014, 20:05

but would we not all have this same problem??
No, because we don´t upscale to 4K (2160p) like you do, as that´s still pretty uncommon for the majority of people.

michkrol

3rd February 2014, 20:20

If I use Image Doubling for 720p content on a 1080p screen, then the content image is doubled to 1440p using NNEDI3, Chroma is upscaled also using NNEDI3, and then image is downscaled using the recommended Catmull-Rom settings, right? Wouldn't the end-result still be of higher quality than disabling NNEDI3 Image Doubling and only using NNEDI3 Chroma Upscaling to prevent downscaling from 1440p?
It should be of higher quality, but NNEDI3 really shines with low resolution and low quality sources. As always best to test it yourself and then decide ;) Profiles make it possible to use NNEDI only for low resolution sources for full flexibility, if you're performance limited.
I also noticed that Image Upscaling still has only Jinc available - would it make any sense to introduce NNEDI3 Image Upscaling? Or maybe Image Upscaling is ignored when Image Doubling is used/enabled?
NNEDI has one major drawback aside from performance, you can only scale by a factor of 2. That's why it's "image doubling" and why you need to use "normal" image scaling afterwards. Image is either downscaled or upscaled after doubling, depending on the source and target resolution. You really should read madshi's release post for more info 1663611

but would we not all have this same problem??
It might be a bug with memory usage, triggered only sometimes or rather if some conditions are met. Also most of us (probably) haven't jumped the 4K displays bandwagon yet ;)
When you get to testing maybe try lowering the queues and see if the crashes persist, since you might be running out of system RAM or GPU VRAM.

huhn

3rd February 2014, 20:25

A few more questions to put my mind at ease:

If I use Image Doubling for 720p content on a 1080p screen, then the content image is doubled to 1440p using NNEDI3, Chroma is upscaled also using NNEDI3, and then image is downscaled using the recommended Catmull-Rom settings, right? Wouldn't the end-result still be of higher quality than disabling NNEDI3 Image Doubling and only using NNEDI3 Chroma Upscaling to prevent downscaling from 1440p? I know many popular game screenshot takers/"artists" actually use higher resolutions (1440p/1600p) to take screenshots and then downsize/downscale them to 1080p because native 1080p does not look as good.
the game source is true 1440 or true 1600 in this case it is not. you can use this as an anti alising filter not that important for vid
I noticed that Catmull-Rom provides less sharpness and more aliasing than Spline or Jinc - why is it still recommended to use Catmull-Rom then? If I remember correctly - Jinc was available for Image Downscaling in previous versions, but not anymore - why is that?
was never available, and sharpness is not all
What is special about the Exclusive mode? Does it improve performance in any way?
no tearing at all mo problems with composition rate from windows and its a but faster or slow not easy to say 23 hz with 23 p content is fast 30p on 60 hz is slow.

I also noticed that Image Upscaling still has only Jinc available - would it make any sense to introduce NNEDI3 Image Upscaling? Or maybe Image Upscaling is ignored when Image Doubling is used/enabled? nnedi can only upscale in factor 2, it's an interpolation not a normal resizing. you can't ignore the rest scaling, when image doubling is used only image up/down scaling is used to correct the rest

madshi

3rd February 2014, 21:30

R U guys sure the 260GTX won't do NNEDI 2x with 16neurons?
My GTS 250 can't do NNEDI 2x 16 neurons even for 25p SD material.
FWIW, I'm playing with DirectCompute now, and it looks promising. But I will limit support to D3D11 capable GPUs, only, so Nvidia 2xx and AMD 4xxx GPUs will not be supported (only for DirectCompute features, of course).

From what I can see error diffusion works linearly in the errors, since transforming the errors from LL to LL YCbCr is a linear operation this shouldn't give you a different result.
Let's say for pixel X the red and blue components are spot on 8bit, but the green component is 0.5 pixel off 8bit. In this case when using RGB, the 0.5 green error will be spread only to the green channel. However, when spreading the error to the neighbor pixels in YCbCr, that should also affect the red and blue pixels. Or am I wrong? To be honest, I haven't really fully thought this through yet...

Yes that is correct. I have uploaded the debug log and video sample here:
http://www.mediafire.com/?132v654vu5zfw

It's a 29.97 fps interlaced video. I am not using forced film mode, but video mode deinterlacing here.
I keep being astonished how anyone could upload a text file without zipping it first. Such a waste of bandwidth...

Anyway, according to your log madVR seems to think that forced film mode is activate (so 60Hz, 24fps). However, after the first frame comes in, actually video mode deinterlacing is performed. Right now I'm a bit confused about this. Will have to look at it in more detail later...

Maybe you can try the debug build version to capture the crash point
A debug log does not help with crashes - ever.

-------

Here's a first test build using DirectCompute instead of OpenCL for Error Diffusion. I've only tested this on my AMD HD7770 so far, so I don't know if it also works on NVidia or even Intel. What I can say is that first initialization seems to be a bit slower, but actual rendering performance seems to be noticeably better, mainly due to interop between D3D9 and DirectCompute being almost free. Give it a try and let me know whether it works for you and how performance compares to v0.87.4 with OpenCL Error Diffusion. Thx.

http://madshi.net/madVRdirectCompute1.rar

Oh yes, this requires D3D11 to be installed, and a D3D11.0 capable GPU.

nevcairiel

3rd February 2014, 21:35

Here's a first test build using DirectCompute instead of OpenCL for Error Diffusion. I've only tested this on my AMD HD7770 so far, so I don't know if it also works on NVidia or even Intel. What I can say is that first initialization seems to be a bit slower, but actual rendering performance seems to be noticeably better, mainly due to interop between D3D9 and DirectCompute being almost free. Give it a try and let me know whether it works for you and how performance compares to v0.87.4 with OpenCL Error Diffusion. Thx.

http://madshi.net/madVRdirectCompute1.rar

Oh yes, this requires D3D11 to be installed, and a D3D11.0 capable GPU.

Honestly, I was kinda wondering why you never mentioned DirectCompute, since it looks to be a half-decent API. :)

Anyhow, glad to report that it works on my NVIDIA card right away, rendering time increases by about 2.5ms on a full 1080p frame.
At least I think it works from the rendering time. No black screens or anything! (on latest drivers, so OpenCL was a no-go)

PS:
Seems to also work on the Intel iGPU, but has a really strong performance hit here. Like 10ms -> 25ms.

madshi

3rd February 2014, 21:42

DirectCompute has one disadvantage compared to OpenCL: It cannot handle DXVA NV12 surfaces. That's why I wanted to use OpenCL. And yes, the Intel iGPU has real trouble with this.

2.5ms doesn't sound too bad. Which exact NVidia GPU is that? And for comparison purposes, how much cost did OpenCL Error Diffusion have with the official v0.87.4? I suppose it was more than 2.5ms? (Even if you're using "too new" drivers for OpenCL and get a black screen, performance numbers are probably still valid.)

Thanks!

huhn

3rd February 2014, 21:42

ED works fine with intel and my nvidia 760

nevcairiel

3rd February 2014, 21:47

2.5ms doesn't sound too bad. Which exact NVidia GPU is that? And for comparison purposes, how much cost did OpenCL Error Diffusion have with the official v0.87.4? I suppose it was more than 2.5ms? (Even if you're using "too new" drivers for OpenCL and get a black screen, performance numbers are probably still valid.)

Its a 780 GTX, so high-end.
Can't really tell you performance numbers when the OSD doesn't draw. :)

iSunrise

3rd February 2014, 21:50

Anyhow, glad to report that it works on my NVIDIA card right away, rendering time increases by about 2.5ms on a full 1080p frame.
At least I think it works from the rendering time. No black screens or anything! (on latest drivers, so OpenCL was a no-go).
Nev, what driver version are you testing this with? Still with 327.23?

I´m on 334.67 beta currently and I have the same problems as with OpenCL, both NNEDI3 doubling/quadrupling and NNEDI3 4:2:2 -> 4:4:4 upscaling don´t work. To clarify that further, I get higher GPU usage, but I have swapped chroma? channels, not a completely black picture.

Ver Greeneyes

3rd February 2014, 21:51

Give it a try and let me know whether it works for you and how performance compares to v0.87.4 with OpenCL Error Diffusion. Thx.
It seems to work well on my GeForce GTX 580! (driver version 327.23)

Using the same (SD quality) video, I compared rendering times for two window sizes.Random dither: 5.26ms
DirectCompute: 6.88ms
OpenCL: 7.80ms

Random dither: 8.54ms
DirectCompute: 11.84ms
OpenCL: 13.80msSo DirectCompute seems to have less overhead and/or scale better!

nevcairiel

3rd February 2014, 21:52

Nev, what driver version are you testing this with? Still with 327.23?

I´m on 334.67 beta currently and I have the same problems as with OpenCL, both NNEDI3 doubling/quadrupling and NNEDI3 4:2:2 -> 4:4:4 upscaling don´t work. To clarify that further, I get higher GPU usage, but I have swapped chroma? channels.

You should really read what madshi writes. He only did Error Diffusion in DirectCompute (so far? :D)

And i'm on the latest beta as well.

huhn

3rd February 2014, 21:55

Asmodian

3rd February 2014, 21:56

I also noticed that Image Upscaling still has only Jinc available - would it make any sense to introduce NNEDI3 Image Upscaling? Or maybe Image Upscaling is ignored when Image Doubling is used/enabled?

The reason NNEDI3 is available for chroma upscaling is because the chroma is always 1/2 the resolution of the video, or doesn't need scaling.

Remember the "chroma upscaling" setting is only for the 4:2:0/4:2:2 -> 4:4:4 conversion. After this step is done further resizing (up or down) is done with the image scaling options (or NNEDI3 doubling if set).

I think this means when watching 720p on 1080p with NNEDI3 chroma upscaling and luma only doubling:
Chroma is scaled to 720 using NNEDI3.
Luma is scaled to 1440 using NNEDI3.
Chroma is scaled to 1080 using image upscaling.
Luma is scaled to 1080 using image downscaling.

But it might be {EDIT: But is NOT}:
Chroma is scaled to 720 using NNEDI3.
Luma is scaled to 1440 using NNEDI3.
Chroma is scaled to 1440 using image upscaling.
Everything is downscaled to 1080 using image downscaling.