madVR - high quality video renderer (GPU assisted)

[ShadowVlican]

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.
i've got a 6950 and prefer not to use OpenCL error diffusion either... the performance hit is massive

NNEDI3 doubling? not a chance. even 16 neurons makes everything unplayable.

i'll just stick with Jinc3 for everything for now

computer specs:
http://valid.canardpc.com/cache/banner/hicw7d.png (http://valid.canardpc.com/hicw7d)

[MistahBonzai]

madshi, could you explain how 'chroma upscaling = NNEDI3' and 'image doubling, double Chroma resolution' relate when upscaling a video exactly 2x?

640x360 -> 1280x720

GPU load 37%
chroma upscaling = NNEDI3
image doubling luma = NNEDI3
image doubling chroma = NNEDI3
opencl dither = enabled

.
.
.

[Edit]
Same thing happens with GPU load changing when 'image doubling luma = NNEDI3' is active for 640x360->1280x720 2x scale, depending on the setting for 'image upscaling'.

I was under the impression that 'image doubling luma' & 'image doubling chroma' would override 'image upscaling' & 'chroma upscaling' when the video is resized to a 2x multiple.

Am I misunderstanding something?

I have to be overlooking something here..or my rig (PCIE 3 2GB HD 7850 (latest Catalyst 13.2 driver pkg), i73770@3.4GH, Windows 7 64bit on a SSD) is a lot weaker that I had imagined.

When I set-up with those values I'm seeing a steady 29% GPU load. No image doubling (1280x720) is occurring and I see similar values to yours. If I select alt+3 (200% up-scale) while playing in MPC-HC I get 79% GPU load. If I scale the video via ffdshow with "Lanczos4Resize(1280,720)" in a simple script I still see GPU load similar to yours. That's because neither of the scaling triggers for Luma or Chroma have kicked in.

Bottom-line is I open the native video (640x360) with the values set as in your test (29% GPU load for me) and then double the video window size via alt+3 (1280*720) and my GPU load jumps to 79%. Otherwise the modifications to the image quality aren't active...

[pie1394]

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).

... skip ...

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)


- The rendering time is about 18ms on my system with the similar + all Jinc3AR setting for 720x480i60 16:9 DVD MV contents.
- It is dropped to 15ms with Bicubic75AR setting... (suggested by madshi. It is just enough when vector-adaptive deinterlacing is also counted for 60fps playback)
- It only takes 6.2ms with pure Jinc3AR setting.

Anyway this shows how powerful HD7970 is if compared to IGP/APU.

ps: My card is made by Lantic. Bought it in March, 2013 and price was about US$320 --- still quite expensive as a gaming / madVR usage card.

--
Core i5-3570K + Z77 + dual-ch DDR3-2400 + HD7970@925MHz Catalyst 13.12 (forced VSync ON, Vector-adaptive deinterlace mode) + Win7x64SP1 + MPC-BE 1.3.0.3 + LavFilter 0.60.1 (DXVA2) + madVR FSE (deband,angle-detect, all Jinc3AR for up-scale)

[Asmodian]

That really is odd. I rather think it's random. Try again with v0.87.4. And if the problem still occurs, please double check with v0.86.11 to make sure it's a new bug and not an old one. I'm only interested right now in bugs which were introduced in v0.87.x.

v0.87.4 fixed my queue size issues, I can use any queue sizes with any resize options. Well at least every combination I tried worked; I didn't do full factorial testing. I cannot reproduce that black screen (I also down reved my Nvidia drivers for obvious reasons, I assume that wasn't it). :D

With all NNEDI3 2x and 4x options at 64 neurons, Jinc3AR for chroma+luma upscaling, debanding low, smooth motion off, and no trade quality for performance options checked, my Titan (@1071MHz) hits 96% watching 352x240 29.97p at 1920x1440. I am also using a 3DLUT and gamma processing, queues at 32/24/8, FSW.

I will have to play around a lot to find the appropriate trade offs. I didn't expect that for a while with a Titan; until I saw all these wonderful NNEDI3 options. :)

Great work as always, amazing. :thanks:

edit: I think I like Jinc3 chroma + NNEDI3 32 neuron doubling (chroma+luma) for watching 720p at 1440p, very sharp. :) With smooth motion on I am at 92% GPU usage and the GPU reports it is pulling 345 watts. :eek:

[antonyfrn]

Thanks guys. More details would be:

(1) Can be reproduced by unchecking the "use random dithering instead of OpenCL error diffusion" option in the "rendering -> trade quality for performance" section of the madVR settings.

(2) Works fine with 327.23 and older drivers. Problem occurs with anything newer than 327.23.

(3) Technical cause is that D3D9 <-> OpenCL interop doesn't seem to work properly. All APIs return success, but the OpenCL kernels seem to produce zero/black output at all times.


Well some good news ManuelG responded about this in the new driver thread and will take a look into it.

[leeperry]

Humm it's hard to deny that a used HD7950 would really hit the spot in order to permanently force 1.2/2.4x 64neurons NNEDI :)

And to my eyes, two things improve the subjective pop effect on my 3500:1 Sammy TV: BFI, as there isn't motion blur smearing everything anymore and Error Diffusion as the dithering noise is not a just a veil of random noise thrown onto the picture(much like what a VGA connection looks like due to the increased noise level). I see the same sort of "pop effect" improvement with mVR's error diffusion as I was getting with SmoothL, too good to be true :devil:

My only problem right now is that using PotP's GUI in 2D mode flashes black frames during the FSE/FSW transitions(that wasn't the case on XP) but PotP's D3D GUI would appear to completely alleviate this problem, so no biggy.

[6233638]

I may have to edit my 3DLUT now, as values near black seem to be darker than before when using error diffusion.

http://abload.de/img/error-diffusion-2uks1c.png


Brightened:
[url]http://abload.de/img/error-diffusion-3plsex.png
http://abload.de/img/random-dither-31wssj.png

As you should be able to see though, with error diffusion enabled each level is much more distinct from the last. When brightened like this, 20-25 all look roughly the same with random dither.


Hopefully editing my 3DLUT will not affect the results I'm seeing on a full black screen:

http://abload.de/thumb/error-diffusion5osvs.png (http://abload.de/img/error-diffusion5osvs.png) http://abload.de/thumb/random-ditherr5sl9.png (http://abload.de/img/random-ditherr5sl9.png)
(brightened significantly)

With error diffusion enabled, black is actually solid black, which means that my display can turn off the local dimming zones rather than keeping them on at a low level.
This is a significant improvement and absolutely worth the performance hit.

Of course, if you don't have that problem, it may not matter, and I really hope that changing my 3DLUT values does not affect this result.

I agree with cyberbeing that error diffusion does have a tendency to create patterns though, which random dither avoids.
But the noise level is so much lower near black, and noise is monochrome with error diffusion, rather than being multicolored with random dither.

[ryrynz]

Perhaps we could have some sort of anti pattern dithering whilst still allowing for this level of black performance?

20-25 are looking fantastic there! That's an awesome result which should have some significant impact in your late night viewing.

[6233638]

Perhaps we could have some sort of anti pattern dithering whilst still allowing for this level of black performance?I wonder if it would be possible to do something like render dithering at 2x and downsample it? (probably not?)

20-25 are looking fantastic there! That's an awesome result which should have some significant impact in your late night viewing.Yes, it's a big improvement. I actually suspect that the measured points in my 3DLUT will be the same, or roughly the same as before, it's just that each level between them is much better defined now.

[madshi]

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.

Opinion section:
Error diffusion vs going from 32 to 64 neurons? :D
FWIW, depending on the video, even 16 neurons can look almost identical to 64 neurons. And 16 neurons only consume a quarter of the preformance of 64 neurons. That said, I recommend not to go lower than 32 neurons for luma upscaling, because 16 neurons leaves too many artifacts in the image for my taste with *some* videos. Every higher neuron setting costs almost twice the performance, and only brings a minor improvement (but an improvement nonetheless).

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.
That sounds like you don't need any of the more "expensive" algorithms. So the limiting factor might be the debanding. I guess a budget GPU should do the trick for you. E.g. look at gaming benchmarks and choose the cheapest GPU which has GDDR5 and which has a good performance per price ratio. Of course choosing a more powerful GPU never hurts madVR playback. But that's your decision, really. If you choose an AMD GPU, I would definitely choose a GCN model (7xxx), though, not an older model.

If you ask for a specific model, well I don't know your exact budget. Given your needs, I guess a 7570 or 7670 should probably be fast enough. A 7750 or 7770 would be awesome, probably too fast for your needs, but with some performance headroom left for future algorithms. Of course you can also go with NVidia. If you need a silent GPU, I'd suggest that you read some reviews, so you pick a good/silent model from a specific manufacturer. In any case, pick one with GDDR5, and make sure the price/performance ratio is good.

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 correct. Although it gets more interesting when you quadruple. In that case the image shifts up/left when doing the first (doubling) pass. Then when doing the 2nd (quadrupling) pass, madVR uses a modified NNEDI3 pass which moves the pixels right/bottom instead of left/up. But the move back right/bottom is smaller in size than the move left/up. So basically both doubling and quadrupling move the image left/up, but doubling moves it more than quadrupling.

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 suppose you're only doubling the luma channel and not the chroma channels? In that case Jinc might be used to upscale the chroma channels directly from 720x576 to 1920x1080. Luma can't be upscaled by Jinc for Y because it would be a downscale. So luma X is upscaled by using Lanczos (using the same tap number you specified for Jinc), and luma Y is downscaled using Catmull-Rom. Lanczos always replaces Jinc if Jinc is selected but can't be used.

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

It is the first ("CbCr channels upscaled 1280x720->1920x1080 with 'image upscaling' setting").

It might be nice to have the ability to import/export profiles so that we can potentially share and cross-check them.
You can already import/export the whole settings - but not profiles, only. I'm not sure if exporting/importing is necessary. Anyway, even if you guys want it, it's probably not coming soon.

I guess using the same physical memory for CPU/GPU just doesn't help at all with the DX9/OpenCL interoperation overhead.
Yes, it seems so, thanks for testing. I'm not sure exactly where the overhead comes from. But everything should be happening on the GPU, so it makes sense that unified memory doesn't help.

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?
Can anybody reproduce this? Doesn't seem to occur here.

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.
No, IVTC is definitely done first. It must be because it's done on the CPU and I don't copyback, so I couldn't accidently swap the order of these two operations.

My best guess would be that your source video is not telecined film, but already contains those blended frames. Forced film mode only works correctly with properly telecined content. For field-blended or frame-blended sources you need to use video mode deinterlacing.

3. Probably known but playing video in film mode is not good.
It's very bad, yes that's known and expected.

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 :D

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 :cool:
Yeah, it seems that the GCN generation (7xxx) helps a lot with OpenCL performance.

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'.
Yes, I'm seeing the same thing. Although the checkerboard pattern should be with color values which are very near to each other, so it should be rather faint.

I agree that random dithering hides source banding better. Actually the original f3kdb debanding algorithm has a feature which does something very similar to madVR's random dithering, just to hide source banding. At some point I'm probably going to add an "add grain" or similar algorithm which you could then combine with error diffusion.

So would it be fair to say that error diffusion is more true to the source compared to random dithering?

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.
My impression is that AMD did a lot of changes in the GCN generation (7xxx) to improve compute performance - which is really what OpenCL is all about. So I suppose 5xxx and 6xxx generations are expected to underperform a bit with OpenCL.

i've got a 6950 and prefer not to use OpenCL error diffusion either... the performance hit is massive

NNEDI3 doubling? not a chance. even 16 neurons makes everything unplayable.
Ouch, that hurts! Sounds like my 7770 performs better than your 6950 with this! Not really sure, though...

I have to be overlooking something here..or my rig (PCIE 3 2GB HD 7850 (latest Catalyst 13.2 driver pkg), i73770@3.4GH, Windows 7 64bit on a SSD) is a lot weaker that I had imagined.

When I set-up with those values I'm seeing a steady 29% GPU load. No image doubling (1280x720) is occurring and I see similar values to yours. If I select alt+3 (200% up-scale) while playing in MPC-HC I get 79% GPU load. If I scale the video via ffdshow with "Lanczos4Resize(1280,720)" in a simple script I still see GPU load similar to yours. That's because neither of the scaling triggers for Luma or Chroma have kicked in.

Bottom-line is I open the native video (640x360) with the values set as in your test (29% GPU load for me) and then double the video window size via alt+3 (1280*720) and my GPU load jumps to 79%. Otherwise the modifications to the image quality aren't active...
I'd suggest to use NNEDI3 only for luma doubling, and only with 32 neurons, if your GPU can't handle more. I would not recommend using 16 neurons for luma doubling, though.

v0.87.4 fixed my queue size issues, I can use any queue sizes with any resize options. Well at least every combination I tried worked; I didn't do full factorial testing. I cannot reproduce that black screen (I also down reved my Nvidia drivers for obvious reasons, I assume that wasn't it). :D
Great!

With all NNEDI3 2x and 4x options at 64 neurons, Jinc3AR for chroma+luma upscaling, debanding low, smooth motion off, and no trade quality for performance options checked, my Titan (@1071MHz) hits 96% watching 352x240 29.97p at 1920x1440. I am also using a 3DLUT and gamma processing, queues at 32/24/8, FSW.
Ouch! I'd suggest to use NNEDI3 only for luma upscaling. IMHO using it for the chroma channels is somewhat wasted. For quadrupling you might get away with 32 or even 16 neurons. For doubling I wouldn't go lower than 32, though.

Well some good news ManuelG responded about this in the new driver thread and will take a look into it.
Great!! :) Please keep us posted, if there's any progress...

And to my eyes, two things improve the subjective pop effect on my 3500:1 Sammy TV: BFI, as there isn't motion blur smearing everything anymore and Error Diffusion as the dithering noise is not a just a veil of random noise thrown onto the picture(much like what a VGA connection looks like due to the increased noise level). I see the same sort of "pop effect" improvement with mVR's error diffusion as I was getting with SmoothL, too good to be true
Happy to hear that!

Which Sammy TV is that? Can it do 4:4:4? I might be looking into getting a new monitor for my development PC. I'm tempted with misusing a TV for that. But 40" does sound a bit large. 30" would be ideal...

I may have to edit my 3DLUT now, as values near black seem to be darker than before when using error diffusion.
Hmmmm... That almost looks like a different gamma curve. But it's not, is it? Do you think error diffusion is more exact here? Or how to do explain the brightness differences? It looks a bit weird to me. Ok, random dithering has higher spikes up and down compared to error diffusion. And the spikes down vanish in BTB. So that might push subjective brightness up. Maybe that's the explanation?

Perhaps we could have some sort of anti pattern dithering whilst still allowing for this level of black performance?
How much does the checkerboard pattern distract in real live viewing? With real movie content every frame should be (at least ever so slightly) different compared to the next, which should totally change the whole error diffusion pattern. It's different with test patterns where each frame might be 100% identical to the previous.

I wonder if it would be possible to do something like render dithering at 2x and downsample it? (probably not?)
After downsampling you end up with more than 8bit. So you'd need another dither/error diffusion pass. Or if you downsample by rounding down to 8bit, the whole supersampling processing would probably have no benefit at all.

Yes, it's a big improvement. I actually suspect that the measured points in my 3DLUT will be the same, or roughly the same as before, it's just that each level between them is much better defined now.
It is quite possible that the measurements also change a bit. So you might want to redo the calibration, just to be sure.

-------

After seeing all those NNEDI3 benchmarks I want to post a settings recommendation. I understand that you guys want to test how much your GPU can handle, so you dial up all the new settings as much as possible, and then you're a little bit let down by how slow it all gets. But most of you really do seem to go overboard with the settings IMHO. So here's my recommendation:

(1) In my tests I've found little reason to enable NNEDI3 for the chroma channels. So I would recommend to leave the right side of the image doubling settings page unchecked.

(2) The neuron number is more important for doubling than it is for quadrupling. I'd recommend 32-64 neurons for doubling and 16-32 for quadrupling. I wouldn't go lower than 16 for doubling, unless your GPU absolutely can't handle 32 neurons and you still prefer NNEDI3 with 16 neurons over Jinc/Lanczos.

(3) When using image doubling with NNEDI3, I'd recommend to use Lanczos3 AR for "image upscaling", and not Jinc3 AR.

Since NNEDI3 costs so much performance, we have to accept some compromises. Of course if you have an almighty GPU, you can dial up everything to your heart's content. But the above recommendation is what should IMHO give the best quality/performance ratio, when using NNEDI3.

[huhn]

still problems with hd4000 and deint split drops first, gpu load still at 84-95 % with 87.4. with a old test 86.11 was at ~82 %, now it is at 88-95% but working.
back buffer is set to 8 and fullscreen exclusive is disabled else it isn't working on 86.11 too rest is totally default.

[turbojet]

Originally Posted by turbojet View Post
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?
Can anybody reproduce this? Doesn't seem to occur here.

Seems to happen always in film mode not in video mode.

Originally Posted by turbojet View Post
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.
No, IVTC is definitely done first. It must be because it's done on the CPU and I don't copyback, so I couldn't accidently swap the order of these two operations.

My best guess would be that your source video is not telecined film, but already contains those blended frames. Forced film mode only works correctly with properly telecined content. For field-blended or frame-blended sources you need to use video mode deinterlacing.

For testing you can use http://www.mediafire.com/download/5l6llo814918rff/wrong_cadence_banding.mpg set display at 60hz, turn on film mode and smooth motion. It's soft telecined so must decode with lav video to find correct cadence. Maybe it's just my setup, setting a profile helps but it would be nice to use smooth motion and ivtc together.

if (srcFps=29) and (filmMode=true) "interlaced" else "progressive"
29.97 ivtcd in film mode = progressive
29.97 in video mode= interlaced

is this a bug?

[huhn]

Which Sammy TV is that? Can it do 4:4:4? I might be looking into getting a new monitor for my development PC. I'm tempted with misusing a TV for that. But 40" does sound a bit large. 30" would be ideal...

there are hybrid displays out there: german list: http://geizhals.at/de/?cat=monlcd19wide&xf=103_Tuner~99_26#xf_top

i highly recommend Philips tvs for pc usages they normally all support unlimited rgb in pc mode on >all< refresh rates.

[omarank]

Profiling works nicely here, but keyboard shortcuts for profiles don't work for "processing" and "rendering" settings. The shortcuts work only for "scaling algorithms".

I am using a profiling script for smooth motion too, and I found that the profile which enables smooth motion, when enabled, doesn't consider the option selected in the smooth motion settings page and always enable smooth motion.

I was wondering how to use "command line to execute when this profile is activated/ deactivated". Can someone please show me an example of using this functionality?

[iSunrise]

Just as a heads-up madshi:

Last night (at 4 o´clock in the evening/morning!) Blaire from 3dcenter already contacted ManuelG of Nvidia, who requested a step by step guidance of the problem. Since I´ve linked your last and more detailed post to him, he was actually able to reproduce the problem and gave all required information to ManuelG in step by step form.

They are looking into it right now. Only thing left now is to look for a new beta driver, where they fix the problem, but I guess Blaire will let me know soon enough. Going to keep you updated on this as soon as I get some feedback.

[Qaq]

MadVR 87.3. ATI 7750, driver 13.12, Win7 SP1 x86. OS and video driver are set to high perfomance.

NNEDI3 image doubling: dropouts ~30%
OpenCL error diffusion: dropouts ~30% (only works fine for small size videos without upscaling to full screen)
NNEDI chroma upsampling: dropouts ~30% (this one worked fine for me in 87.1b, btw)

Timings are fine, its just present queue goes to zero. Need more fine tuning of the new CCC 13.12 maybe.

[iSunrise]

I may have to edit my 3DLUT now, as values near black seem to be darker than before when using error diffusion.
...
As you should be able to see though, with error diffusion enabled each level is much more distinct from the last. When brightened like this, 20-25 all look roughly the same with random dither.
That is actually very interesting. Never thought that error diffusion (haven´t checked it myself, yet, though) would improve something in such drastic form.

When I hardware calibrated my Eizo CG243W and checked that pattern after calibration in madVR, I also wondered why at some certain brightness levels (especially the lower ones) it seemed to miss a lot of differentiation, when outside of madVR, everything was perfectly fine. At that point I actually thought that I must have been doing something wrong with the madVR settings, but doing it again carefully didn´t change the outcome. Back then I reported my results, but I guess it was quite hard for others to understand what exactly I was seeing, since you need to have a screen that is able to perfectly differentiate levels in the first place.

I am really surprised by these results now. If that translates well to the actual picture, this could mean a lot more defined and natural apperance of the source and also a lot more details in darker (and brighter?) areas. And it would also mean that every good calibrated LCD/CRT would benefit a lot from it. It would actually make it even more worth it to calibrate your screen now than it already has been before.

Are we really sure that this is only the result of random dithering vs. error diffusion? And what are the reasons for this? Why should dithering change differentiation steps in the brightness levels? Yes, it is kinda expected, that less noise would lead to a lower noise floor, and as such, would lead to somewhat more perfect and darker blacks (and probably also whites) but why would it suddenly change the differentiation between the levels? Wouldn´t there have to be some gamma change involved to come to this result? I don´t understand this at all.

Could this be a result of your 3DLUT only, which needed to compensate for the shortcomings of your display? But why did I see the same thing and outside of madVR, everything is fine. That´s confusing me to no end right now. But I think I really need to test this to see it for myself.

[leeperry]

So is there any way to avoid the black frames flashing during FSE/FSW transitions? This is happening whenever I want to open a context menu and that never occured on XP.

All this said, FSE was mandatory on XP but isn't it a bit of overkill on W7? The OS knows how to handle VSYNC for a change.

depending on the video, even 16 neurons can look almost identical to 64 neurons. And 16 neurons only consume a quarter of the preformance of 64 neurons. That said, I recommend not to go lower than 32 neurons for luma upscaling, because 16 neurons leaves too many artifacts in the image for my taste with *some* videos. Every higher neuron setting costs almost twice the performance, and only brings a minor improvement (but an improvement nonetheless).
Does it make sense to use 32 neurons for 720p to 1080p? Or would J3AR be sufficient and NNEDI only necessary for SD to HD? I guess that calls for a number of tests but you might have run some on your own.

Which Sammy TV is that? Can it do 4:4:4? I might be looking into getting a new monitor for my development PC. I'm tempted with misusing a TV for that. But 40" does sound a bit large. 30" would be ideal...
Yep, personally I find 27" too small and 32" a tad too big, this 30.5" OLED would be perfect: CES 2014: Chinese Brands Get In On The OLED Action (http://www.digitalversus.com/tv-television/ces-2014-chinese-brands-get-oled-action-n32652.html)

It's the Sammy UE32F5000, it does support 4:4:4 but then no more BFI so it becomes blurryland and there's the usual panel and backlight homogeneity lotteries....luckily you can check the panel maker in the factory menu, there's a thread on that matter on AVS if you can find the patience of trying your luck: Samsung Panel Version Thread (http://www.avsforum.com/t/1294904/)

Many ppl claim that the Sammy panels are better but not a single one shows any measurement, all I know is that the AUO panel provides 0.4 cd/m² black, acceptable blurring once BFI is enabled and 3500:1 native CR.

The best choice for 4:4:4 BFI and 5K:1 native CR is that new UV²A monitor from EIZO but it costs 500€ and it's 24" only. This Sammy TV will fit the bill until the chinese flood the market with OLED :)

i highly recommend Philips tvs for pc usage
AFAIK these don't provide any kind of R/G/B gain/offset settings so you'll have to use a CLUT on the PC in order to reach D65.

[DragonQ]

Who can actually use error diffusion though? It seems as if it's about as demanding as NNEDI3 from all the reports so far.

[ryrynz]

Who can actually use error diffusion though? It seems as if it's about as demanding as NNEDI3 from all the reports so far.

This is why I inquired about a CPU version.. unfortunately no answer from Madshi though :P So many goodies.. so much GPU.. :(

Hey Madshi, was wondering how a 2x upscale with Jinc3ar + 2x upscale with nnedi then a downscale with Lanczos3 would look in MadVR. Basically getting that nnedi sharpness and antialiasing without too much thinning happening.. could be a nice blend of both. Do you think it's worth having a preupscaling option? I've played a round a bit in Avisynth with pre Jinc and Lanczos upscaling and from what I see things look quite nice when set up before nnedi. Thoughts?

[huhn]

AFAIK these don't provide any kind of R/G/B gain/offset settings so you'll have to use a CLUT on the PC in order to reach D65.
are you sure? my old 2011 week 11(and not really good) pf4600 got all theses options. for my 3d lut i use r128 g122 b122 to reach d65. so newer models didn't have this anymore this would be terrible...

[huhn]

This is why I inquired about a CPU version.. unfortunately no answer from Madshi though :P So many goodies.. so much GPU.. :(

if a gpu can't handle it how can a cpu do it? and like nev? said it's the last step... ofc you can download then to the ram use the cpu and upload it agian but this would be slow really slow.

[nevcairiel]

huhn is right, trying to get the CPU to do dithering is practically impossible, as it would mean you have to download the image, process it (which isn't going to be very fast either), and upload it again. Its going to be insanely slow.

[Ver Greeneyes]

Who can actually use error diffusion though? It seems as if it's about as demanding as NNEDI3 from all the reports so far.
It doesn't seem to take much juice (2-3ms) on my laptop's Radeon Mobility HD 4650, which surprises me a little (especially given that it's stuck on the legacy driver).

madshi, I don't suppose it would be possible to test that the OpenCL dithering is working correctly? e.g. by testing a pre-dithered sample against the result of using it. I hope it either works or doesn't work at all, but with my laptop card coping so well I'm a little worried that it's just disabling dithering altogether :) (call me paranoid)

[antonyfrn]

Great!! :) Please keep us posted, if there's any progress...


Will do I posted your list for him and suggested he try's to get in contact with you hope you don't mind.

[ryrynz]

if a gpu can't handle it how can a cpu do it? and like nev? said it's the last step... ofc you can download then to the ram use the cpu and upload it agian but this would be slow really slow.

Damn, makes sense, well I would think a i7 would be faster at certain things than the openCL performance of an HD3000 or 4000.. but I guess there's that image transfer which makes it completely pointless performance wise, oh well.


How much does the checkerboard pattern distract in real live viewing? With real movie content every frame should be (at least ever so slightly) different compared to the next, which should totally change the whole error diffusion pattern. It's different with test patterns where each frame might be 100% identical to the previous.


Nah, not distracting at all.. I don't know how anyone could actually see that with their naked eyes, would have to be a hawk. I will say things look very very clean with it enabled upon close inspection, just wish it had less of an impact on the GPU.

[leeperry]

Oh well, scratch that...720p@1080p looks a hell lot better with 16 neurons NNEDI than regular J3AR, too bad I can't swing 32 neurons.

And I do get tearing at the bottom of the screen(roughly 5%) in FSW so I'll have to use FSE, I just wish the frigging black frames flashing didn't occur..

I also get tearing with other renderers but it's much lower like at 1% of the bottom of the screen, I can't believe tearing is still a problem on a mid-range H87 Haswell mobo with a HD7850....I wasn't getting any tearing on a P35 s775 mobo :mad:

are you sure? my old 2011 week 11(and not really good) pf4600 got all theses options.
Philips TV's are chinese OEM made by TCL these days FWIR, so YMMV I guess.

[noee]

It doesn't seem to take much juice (2-3ms) on my laptop's Radeon Mobility HD 4650, which surprises me a little (especially given that it's stuck on the legacy driver).

madshi, I don't suppose it would be possible to test that the OpenCL dithering is working correctly? e.g. by testing a pre-dithered sample against the result of using it. I hope it either works or doesn't work at all, but with my laptop card coping so well I'm a little worried that it's just disabling dithering altogether :) (call me paranoid)

Edit: Oh wait, Madshi stated that ErrorDiff would only work 8bit? Some many interdepedencies now, time for a spreadsheet.... ;)

[Plutotype]

I have read the last few pages and would like get answer on when and using which NNEDI configuration would one need to get 720p/1080p to 2160p NNEDI@ x neurons to achieve maximum upscaling clarity without artifacts. Can also anybody estimate which CPU and GPU would one need to achieve this?
Thanks

[huhn]

I may have to edit my 3DLUT now, as values near black seem to be darker than before when using error diffusion.

http://abload.de/thumb/error-diffusion-2uks1c.png (http://abload.de/img/error-diffusion-2uks1c.png) http://abload.de/thumb/random-dither-2ygsae.png (http://abload.de/img/random-dither-2ygsae.png)

Brightened:
http://abload.de/thumb/error-diffusion-3plsex.png (http://abload.de/img/error-diffusion-3plsex.png) http://abload.de/thumb/random-dither-31wssj.png (http://abload.de/img/random-dither-31wssj.png)

As you should be able to see though, with error diffusion enabled each level is much more distinct from the last. When brightened like this, 20-25 all look roughly the same with random dither.


Hopefully editing my 3DLUT will not affect the results I'm seeing on a full black screen:

http://abload.de/thumb/error-diffusion5osvs.png (http://abload.de/img/error-diffusion5osvs.png) http://abload.de/thumb/random-ditherr5sl9.png (http://abload.de/img/random-ditherr5sl9.png)
(brightened significantly)

With error diffusion enabled, black is actually solid black, which means that my display can turn off the local dimming zones rather than keeping them on at a low level.
This is a significant improvement and absolutely worth the performance hit.

Of course, if you don't have that problem, it may not matter, and I really hope that changing my 3DLUT values does not affect this result.

I agree with cyberbeing that error diffusion does have a tendency to create patterns though, which random dither avoids.
But the noise level is so much lower near black, and noise is monochrome with error diffusion, rather than being multicolored with random dither.
i can't reproduce this at alll... the brightness isn't changing at all with my 3d lut.

Edit: Oh wait, Madshi stated that ErrorDiff would only work 8bit? Some many interdepedencies now, time for a spreadsheet....

rendertimes jump up by ~13ms on 10 bit sources so i think 10bit works.

and errordiff gets 16 bit ycbcr 4:4:4 anyway...

[noee]

rendertimes jump up by ~13ms on 10 bit sources so i think 10bit works.

I was talking panel native bitdepth setting.

[6233638]

That is actually very interesting. Never thought that error diffusion (haven´t checked it myself, yet, though) would improve something in such drastic form. I have just finished creating a new 3DLUT and tested this again - I am not getting the same results now.

With the new 3DLUT the two images look almost identical, other than the dithering. Again, these have been brightened considerably, as these bars are very near black. edit: and I now realize that I left debanding on. (low preset)

http://abload.de/thumb/new-3dlut-error-diffuczlhr.png (http://abload.de/img/new-3dlut-error-diffuczlhr.png) http://abload.de/thumb/new-3dlut-random-dithh6axt.png (http://abload.de/img/new-3dlut-random-dithh6axt.png)

The only thing I can think of which may have caused the differences I saw before, is that I simply hit PrtScn when the bars were on-screen, and perhaps I managed to pick a frame that was different?
Or worse, perhaps I disabled the 3DLUT without realizing.

This time I counted the number of flashes and restarted the video when I did the comparison.

I am very happy to report that black is still black with the new 3DLUT though.

Who can actually use error diffusion though? It seems as if it's about as demanding as NNEDI3 from all the reports so far.I have no problem using it on my system. If I recall correctly, it adds about 5ms to the rendering times with 1080p video displayed at 100%.

So is there any way to avoid the black frames flashing during FSE/FSW transitions? This is happening whenever I want to open a context menu and that never occured on XP.I've only ever used madVR on Windows 7/8, and this has always been the case. I was not aware that XP didn't do this.

All this said, FSE was mandatory on XP but isn't it a bit of overkill on W7? The OS knows how to handle VSYNC for a change.FSE is also for performance, and preventing other windows overlaying the madVR image.

And I do get tearing at the bottom of the screen(roughly 5%) in FSW so I'll have to use FSE, I just wish the frigging black frames flashing didn't occur..Have you disabled Aero/Desktop Composition? (switching the Windows theme might also do this?)

[pie1394]

- Video screen becomes totally black when the 1st subtitle arrived from XySubFilter -- if SmoothMotion logic is NOT engaged. It happens to any 24-frames or 60-fields contents regardless of resolution. Debanding option ON/OFF does not matter, either.

Hi madshi,

This issue which was introduced since 0.87 is finally fixed in 0.87.4! :thanks:

So the remain old & weird case is the black box under ASS subtile text with SmoothMotion feature on ION-LE(GF9300) platform + GeForce 320.49 since 0.86.x. It happens to all of my samples with external ASS file. Thus I also wonder why it does not happen to your ION(GF9400) machine. :confused:

--
C2D P9700 + ION-LE(GF9300) + dual-ch 8GB DDR3-1333 + Dell U2412M + Win7x64SP1 + MPC-BE 1.3.0.3 + LavFilter 0.60.1 + madVR (deband low, all Bilinear, all for performance) + XySubFilter 3.1.0.546

[MSL_DK]

It has been hell to get OpenCL to work with the latest driver from AMD

Here's what I've done, hope it can help others.


DOWNLOAD and RUN AMD Catalyst Uninstall Utility

GO TO device manager> Display adapters> right click> properties> driver> uninstall

Go to Start–>Search type in gpedit.msc

Click the file to open the Local Group Policy Editor

Go to ->Computer Configuration->Administrative Templates->System->Device Installation. Click on the subfolder Device Installation on the left and on the right side you will see the possible restrictions.
Right Click on Prevent Installation of Devices not described by other policy settings and edit this option, set it on ENABLED.

Reboot

Go to ->Computer Configuration->Administrative Templates->System->Device Installation. Click on the subfolder Device Installation on the left and on the right side you will see the possible restrictions.
Right Click on Prevent Installation of Devices not described by other policy settings and edit this option, set it on DISABLED.

INSTALL AMD Catalyst drivers

Reboot

[leeperry]

Have you disabled Aero/Desktop Composition?
Yes I did as I don't like transparent GUI's, out of desperation I tried to set the Windows GUI to "best appearance" and I found out the hard way that "desktop composition" was the fancy name for Aero :D

As soon as I uncheck it in mVR, the tearing is back so I wonder what that mVR option might be good for?

I've only ever used madVR on Windows 7/8, and this has always been the case. I was not aware that XP didn't do this.
Oh, so for once XP beats W7 hah.......I presume that madshi didn't find a way to avoid the flashing on Vista+? :(

FSE is also for performance, and preventing other windows overlaying the madVR image.
Well, the flashing is quite a PITA.......if Aero can work as effectively in FSW as mVR on its own in FSE then I don't think I'll bother with the latter anymore.

It has been hell to get OpenCL to work with the latest driver from AMD
On x64 or something? I simply disabled the drivers signature check before installing the 13.12 AMD drivers on W7SP1 et voilà ))

[madshi]

I had in madVR settings image upscaling and image downscaling set to DXVA2. If I change the settings to Bilinear the video plays.
I can confirm setting image upscaling and/or downscaling to DXVA2 freezes the image whether it's up/downscaled. On an Nvidia 8200 onboard (that's right; I use bilinear and works :D )
I can confirm the problem. Will look at that for v0.87.5. But not today, anymore...
Strange. I could reproduce it yesterday, but not today, anymore. Can anybody still reproduce this problem? If so, please upload your settings somewhere so I can try to reproduce again.

still problems with hd4000 and deint split drops first, gpu load still at 84-95 % with 87.4. with a old test 86.11 was at ~82 %, now it is at 88-95% but working.
back buffer is set to 8 and fullscreen exclusive is disabled else it isn't working on 86.11 too rest is totally default.
Hmmmm... I thought it was confirmed by some other users that deinterlacing was back to v0.86.11 levels. Now you say it isn't? FWIW, I've forgot (once again) to reenable optimization for v0.87.4, so CPU consumption is slightly higher than necessary. But I don't think that's the cause of whatever issue you're seeing.

there are hybrid displays out there: german list: http://geizhals.at/de/?cat=monlcd19wide&xf=103_Tuner~99_26#xf_top

i highly recommend Philips tvs for pc usages they normally all support unlimited rgb in pc mode on >all< refresh rates.
I'm not interested in a TV due to the tuner. What interests me most is having a bigger display (30"), a good panel (not TN), and HDMI 1.4 3D playback ability. I can't seem to find a good monitor which ticks all these boxes. Most 3D capable monitors only support 3D when using an NVidia GPU which I find a terrible idea.

Profiling works nicely here, but keyboard shortcuts for profiles don't work for "processing" and "rendering" settings. The shortcuts work only for "scaling algorithms".
I can confirm that. Will be fixed in the next build.

I am using a profiling script for smooth motion too, and I found that the profile which enables smooth motion, when enabled, doesn't consider the option selected in the smooth motion settings page and always enable smooth motion.
I can't confirm that. That seems to work here.

I was wondering how to use "command line to execute when this profile is activated/ deactivated". Can someone please show me an example of using this functionality?
Oh well, I actually didn't get around implementing this yet... :o Will add that in the next build, too. What do you mean with an example? You mean you're wondering what the purpose of this functionality is and what it could be good for? Well, I don't know, to be honest, with the current feature set. But one of the next planned features will make the command line stuff useful for some users.

Last night (at 4 o´clock in the evening/morning!) Blaire from 3dcenter already contacted ManuelG of Nvidia, who requested a step by step guidance of the problem. Since I´ve linked your last and more detailed post to him, he was actually able to reproduce the problem and gave all required information to ManuelG in step by step form.

They are looking into it right now. Only thing left now is to look for a new beta driver, where they fix the problem, but I guess Blaire will let me know soon enough. Going to keep you updated on this as soon as I get some feedback.
Will do I posted your list for him and suggested he try's to get in contact with you hope you don't mind.
Great! Thanks, guys, that sounds promising. Of course I wouldn't mind at all if any NVidia guy contacted me.

MadVR 87.3. ATI 7750, driver 13.12, Win7 SP1 x86. OS and video driver are set to high perfomance.

NNEDI3 image doubling: dropouts ~30%
OpenCL error diffusion: dropouts ~30% (only works fine for small size videos without upscaling to full screen)
NNEDI chroma upsampling: dropouts ~30% (this one worked fine for me in 87.1b, btw)

Timings are fine, its just present queue goes to zero. Need more fine tuning of the new CCC 13.12 maybe.
Not sure what's going on there for you. On my PC with Win 8.1 x64 and HD 7770 all queues are perfectly full, as long as rendering time is a bit (not much) lower than the frame interval. No problems whatsoever produced by OpenCL, both NNEDI3 and error diffusion increase the rendering times but have no negative effect on queues and produce no frame drops.

So is there any way to avoid the black frames flashing during FSE/FSW transitions? This is happening whenever I want to open a context menu and that never occured on XP.
XP was a bit faster and more seamless when doing FSE/FSW transitions. There's nothing we can do about that. Well, maybe I could find a way (or maybe not), but I consider that a cosmetical problem and as such not very important at the moment.

All this said, FSE was mandatory on XP but isn't it a bit of overkill on W7? The OS knows how to handle VSYNC for a change.
Windows never had the intention to guarantee that every thread is getting enough CPU time without any serious interruptions. When talking about 60fps, every frame is only visible for ~16.7ms. If you don't use FSE, playback can only be smooth if madVR always gets enough time in every 16.7ms slot to prepare and present the required frame. If only once Windows fails to give madVR enough CPU time for a duration of about 16.7ms, automatically you'll get a frame drop. This is where FSE helps: By presenting several frames in advance, you increase the critical time window from 16.7ms to a multiple of 16.7ms. E.g. with 8 presented frames in advance playback will still be perfect even if Windows stops giving madVR CPU time for 100ms.

It's the Sammy UE32F5000, it does support 4:4:4 but then no more BFI so it becomes blurryland and there's the usual panel and backlight homogeneity lotteries....
Does it support 3D playback via HDMI 1.4? And does it support 4:4:4 at every refresh rate, or just at 60Hz?

Who can actually use error diffusion though? It seems as if it's about as demanding as NNEDI3 from all the reports so far.
WHAT!?

On my PC (Windows 8.1 x64, AMD HD7770) I get the following rendering times:

1080p24 playback, 1680x1050 monitor:
- random dithering: 2.0 ms
- error diffusion: 8.3 ms

1080p60 playback, 1680x1050 monitor:
- random dithering: 2.6 ms
- error diffusion: 10.1 ms

With a 1920x1080 monitor cost of error diffusion would be about 15% higher, but that's it. Sure, error diffusion is not free. But it's not *that* demanding, either. And we're talking about a HD7770 here. There are some much more powerful beasts out there.

This is why I inquired about a CPU version.. unfortunately no answer from Madshi though
I did answer (though a bit later than normal). You seem to have missed it.

Hey Madshi, was wondering how a 2x upscale with Jinc3ar + 2x upscale with nnedi then a downscale with Lanczos3 would look in MadVR. Basically getting that nnedi sharpness and antialiasing without too much thinning happening.. could be a nice blend of both. Do you think it's worth having a preupscaling option? I've played a round a bit in Avisynth with pre Jinc and Lanczos upscaling and from what I see things look quite nice when set up before nnedi. Thoughts?
What would be the purpose of this? Just to remove aliasing? I don't think such a solution would make the image sharper, it would just remove aliasing from the source. I think it would be more effective to run a specialized antialiasing filter, if that's what you need. Most sources don't have aliasing problems, though...

It doesn't seem to take much juice (2-3ms) on my laptop's Radeon Mobility HD 4650, which surprises me a little (especially given that it's stuck on the legacy driver).
Do you have 13.12 drivers installed? As I wrote in the v0.87.0 announcement post, OpenCL only works with AMD GPUs if you update to the latest drivers.

I have read the last few pages and would like get answer on when and using which NNEDI configuration would one need to get 720p/1080p to 2160p NNEDI@ x neurons to achieve maximum upscaling clarity without artifacts. Can also anybody estimate which CPU and GPU would one need to achieve this?
My HD7770 can do 1080p -> UHD upscaling with NNEDI3 - but only with 24fps content and 16 neurons, and only if I disable everything else. And I don't recommend using 16 neurons for luma doubling. So you'd probably need something which is a bit more than twice as fast as the HD7770. Maybe 2.5x as fast, to be on the safe side. Then you should be able to do 1080p -> 2160p luma doubling with 32 neurons. Probably the "HD 7870 XT" could do the job, but this is just a guess, I can't guarantee it. CPU doesn't matter.

For testing you can use http://www.mediafire.com/download/5l6llo814918rff/wrong_cadence_banding.mpg set display at 60hz, turn on film mode and smooth motion. It's soft telecined so must decode with lav video to find correct cadence. Maybe it's just my setup, setting a profile helps but it would be nice to use smooth motion and ivtc together.
Seems to work fine here. No cadence breaks detected and frame stepping looks alright. Can't see anything wrong when playing it back at 60Hz with smooth motion frc activated, either.

Have you tested in FSE mode? And there are no frame drops? Does smooth motion FRC work fine for you with other videos?

if (srcFps=29) and (filmMode=true) "interlaced" else "progressive"
29.97 ivtcd in film mode = progressive
29.97 in video mode= interlaced

is this a bug?
I've already explained this before. "srcFps" is information coming from the decoder, and it's not terribly exact. It could be 29.970 or 29.969 or 29.971. It will definitely not be 29. So I'd suggest using ">" or "<" or something like that. E.g. you could do "if (srcFps>29.9) and (srcFps < 30.1) and (filmMode) "interlaced" else "progressive"".

So the remain old & weird case is the black box under ASS subtile text with SmoothMotion feature on ION-LE(GF9300) platform + GeForce 320.49 since 0.86.x. It happens to all of my samples with external ASS file. Thus I also wonder why it does not happen to your ION(GF9400) machine. :confused:
Maybe I missed something. But since this is an old bug, please report it to the bug tracker and I'll have another look at it when I find the time. Thanks.

It has been hell to get OpenCL to work with the latest driver from AMD
Strange, worked just fine without any problems here.

[huhn]

i think the hd4000 issue is just a very very small performance degradation.

after i set my ram at 1600 it works fine and i can't really see a huge difference in gpu laod so the hd4000 was at his limited with 1333 mhz ram.

I'm not interested in a TV due to the tuner. What interests me most is having a bigger display (30"), a good panel (not TN), and HDMI 1.4 3D playback ability. I can't seem to find a good monitor which ticks all these boxes. Most 3D capable monitors only support 3D when using an NVidia GPU which I find a terrible idea.

the big point was that Philips pc mode works with all refresh rate. and the information is still easy to find in the manuels.

and samsung is know with his problems in pc mode (1 hdmi only 60 hz need a special name...) but time moves on and products should improve so they may fix that general problem.


As soon as I uncheck it in mVR, the tearing is back so I wonder what that mVR option might be good for?

performance! intel and nvidia doesn have tearing problems i only get them with amd and disabled aero.

you can disable most of aero in the system properties -> performance options just disable everything except "enable desktop composition" aero is now running with no transparent effect.

[tFWo]

Madshi, big thanks for the new features!

I don't know if this information about Error Diffusion was already mentioned. Might have missed it.
Error Diffusion performance is highly dependent on having Smooth Motion turned on.

For example, a 1080p24 (SM 24->60) file on my 270X gives these rendering times (I have nnedi64 chroma upscaling turned on and I’m downscaling with lanzos3arll to 1680X1050 so the rendering times are already high :) ):
23ms (SM off, ED off)
36ms (SM off, ED on)
24ms (SM on, ED off)
51ms (SM on, ED on)

I don’t know if this is the correct behavior. ED is applied after SM, so the performance hit is massive if you are doing 24->60. Is it necessary to dither after SM or it can be done before it?

[turbojet]

Quote:
Originally Posted by turbojet View Post
For testing you can use http://www.mediafire.com/download/5l...ce_banding.mpg set display at 60hz, turn on film mode and smooth motion. It's soft telecined so must decode with lav video to find correct cadence. Maybe it's just my setup, setting a profile helps but it would be nice to use smooth motion and ivtc together.
Seems to work fine here. No cadence breaks detected and frame stepping looks alright. Can't see anything wrong when playing it back at 60Hz with smooth motion frc activated, either.

Have you tested in FSE mode? And there are no frame drops? Does smooth motion FRC work fine for you with other videos?\

It works fine in FSE and overlay, problem only in windowed mode. It happens on all videos with the same scenario. It could be a gpu load issue, goes from 55% to 60% after enabling frc. Deinterlacing 60fps at 30 fps shows 85% gpu and plays flawlessly on 250, on 650 the same scenario doesn't look good either in windowed, gpu is about 30% in that case. So I doubt it's an issue with the 250 gpu load.

I've noticed another oddity when it comes to gpu usage. on 650 bob deinterlacing 60fps is only using about 50% gpu but does not play well in any mode, deinterlacing at 30fps looks fine in every mode, could this be an issue with the display?. Also using nnedi3 to upscale sd to 1080p works fine with about 50% load, however upscaling 720p drops many frames but the gpu load is only about 60%. I can crank up the gpu load to about 90% with jinc ar and it doesn't drop any frames, any ideas?

Quote:
Originally Posted by turbojet View Post
if (srcFps=29) and (filmMode=true) "interlaced" else "progressive"
29.97 ivtcd in film mode = progressive
29.97 in video mode= interlaced

is this a bug?
I've already explained this before. "srcFps" is information coming from the decoder, and it's not terribly exact. It could be 29.970 or 29.969 or 29.971. It will definitely not be 29. So I'd suggest using ">" or "<" or something like that. E.g. you could do "if (srcFps>29.9) and (srcFps < 30.1) and (filmMode) "interlaced" else "progressive"".

After some sleep I figured out my mistake.

[ryrynz]

What would be the purpose of this? Just to remove aliasing? I don't think such a solution would ma

On the one hand NNEDI gives beautiful sharp edges but does so at the expense of thinning things out a little more than they should be on these edges.. Quite noticeable on lettersfor example, they look fine in Jinc and Lanczos.
I wanted to experiment to see if I could get some of that NNEDI sharpness retained but keeping with as much Jinc style upscaling as possible and hopefully avoiding some of that funky pattern thing NNEDI does too.

[omarank]

I am using a profiling script for smooth motion too, and I found that the profile which enables smooth motion, when enabled, doesn't consider the option selected in the smooth motion settings page and always enable smooth motion.

I can't confirm that. That seems to work here.

Can you try this script?

if (deintFps > 24.5) and (deintFps < 25.5) "Off"
else "On"

In the "On" profile, enable smooth motion with the option "only if there would be motion judder without it...".

{Basically I want to use smooth motion the way it is selected in the "On" profile, but I need to disable it specifically for 25 fps content as LAV outputs still images as 25 fps video and there is a reported bug in madVR due to which a black screen is produced for still images if smooth motion is enabled.}



I was wondering how to use "command line to execute when this profile is activated/ deactivated". Can someone please show me an example of using this functionality?

Oh well, I actually didn't get around implementing this yet... :o Will add that in the next build, too. What do you mean with an example? You mean you're wondering what the purpose of this functionality is and what it could be good for? Well, I don't know, to be honest, with the current feature set. But one of the next planned features will make the command line stuff useful for some users.

You have answered my questions. I am looking forward to these features.

[The 8472]

WHAT!?

On my PC (Windows 8.1 x64, AMD HD7770) I get the following rendering times:

1080p24 playback, 1680x1050 monitor:
- random dithering: 2.0 ms
- error diffusion: 8.3 ms

1080p60 playback, 1680x1050 monitor:
- random dithering: 2.6 ms
- error diffusion: 10.1 ms

With a 1920x1080 monitor cost of error diffusion would be about 15% higher, but that's it. Sure, error diffusion is not free. But it's not *that* demanding, either. And we're talking about a HD7770 here. There are some much more powerful beasts out there


GTX 760
Smooth motion on
60Hz
doubling 720p -> 1440p with nnedi 32 neurons
2560x1440 display


random dithering: 51-55% GPU load
error diffusion: 90+% load, occasional frame drops

I only see present times in the OSD, no rendering times.

Have you considered using void and cluster ordered dither instead of error diffusion? It would only require a lookup table instead of inspecting neighboring pixels. That might be more shader-friendly.

[omarank]

On my PC (Windows 8.1 x64, AMD HD7770) I get the following rendering times:

1080p24 playback, 1680x1050 monitor:
- random dithering: 2.0 ms
- error diffusion: 8.3 ms

1080p60 playback, 1680x1050 monitor:
- random dithering: 2.6 ms
- error diffusion: 10.1 ms

With a 1920x1080 monitor cost of error diffusion would be about 15% higher, but that's it. Sure, error diffusion is not free. But it's not *that* demanding, either.

Below are the rendering times on my system (Windows 7 32 bit, GTX 260):

Smooth Motion OFF:
- Random dithering: 8.7 ms
- Error diffusion: 18.5 ms

Smooth Motion ON:
- Random dithering: 10.3 ms
- Error diffusion: 27.8 ms

This is for 720p video on 1280x1024 monitor (no image scaling in the output).
[Settings: FSE, CPU – 24, GPU – 16, frames in advance – 12, default flush settings, Chroma – Jinc3 AR]

Is the performance hit with error diffusion due to OpenCL implementation or the algorithm itself?

[cyberbeing]

With a 1920x1080 monitor cost of error diffusion would be about 15% higher, but that's it. Sure, error diffusion is not free. But it's not *that* demanding, either. And we're talking about a HD7770 here. There are some much more powerful beasts out there.

Relatively speaking, the OpenCL error-diffusion dither has the same cost at 1080p 24hz as Jinc4|AR Luma + Jinc4|AR Chroma 1280x720 -> 1920x1080 scaling + Random Dither. Overall I'd say it's probably the second most expensive thing in madVR behind NNEDI3 now, considering it's a couple ms slower than NNEDI3|16 1280x720 -> 2560x1440 display. At least that's the case with my GTX 770.

You may want to consider implementing something similar to the 'dither fruit' method from MPV, which is essentially an ordered dither with optional temporal randomization. It would give people another low-noise dither alternative, at much lower performance cost.


So would it be fair to say that error diffusion is more true to the source compared to random dithering?

Yes, I'd say its more true to the source, but the dither quality feels somewhat questionable on less than pristine sources. From what I can tell, it only does a hard mix of existing source colors, lightness, hues. For example, I saw it create a dither comprising dark specks on a light surface. Similarly I've seen it hard mix red specks and green specks in a pattern. While this works to some extent perceptually, it seems like it would be superior if the dither created new colors to better assist in creating smooth transitions.

[The 8472]

Relatively speaking, the OpenCL error-diffusion dither has the same cost at 1080p as Jinc4|AR Luma + Jinc4|AR Chroma 1280x720 -> 1920x1080 scaling + Random Dither. You're comparing apples and oranges here. Upscaling depends on input framerate, dithering on output framerate. Both also depend on the target resolution. And one is implemented as shaders and the other as OpenCL, so implementation efficiency might also differ between vendors.

So the relative costs can vary substantially between different setups.

[leeperry]

XP was a bit faster and more seamless when doing FSE/FSW transitions. There's nothing we can do about that. Well, maybe I could find a way (or maybe not), but I consider that a cosmetical problem and as such not very important at the moment.

Windows never had the intention to guarantee that every thread is getting enough CPU time without any serious interruptions. When talking about 60fps, every frame is only visible for ~16.7ms. If you don't use FSE, playback can only be smooth if madVR always gets enough time in every 16.7ms slot to prepare and present the required frame. If only once Windows fails to give madVR enough CPU time for a duration of about 16.7ms, automatically you'll get a frame drop. This is where FSE helps: By presenting several frames in advance, you increase the critical time window from 16.7ms to a multiple of 16.7ms. E.g. with 8 presented frames in advance playback will still be perfect even if Windows stops giving madVR CPU time for 100ms.
OK thanks for the details, I'll stick to FSE then :)

Luckily PotP's coder was kind enough to accept a few of my feature requests regarding its D3D GUI which makes me plenty happy now, especially as it never breaks FSE. Of course the black frames flashing is still there but it wouldn't appear to be nearly as annoying if I pause the movie before selecting context menus(such as manually rolling gamuts via PS scripts).

Sometimes I don't only get black frames but previous frames from the movie huh.

My only current problem is that I've set mVR to automatically roll refresh rates when going FS and when going from 1080p60 windowed to 1080p24/50 FSE, I randomly get an "Exclusive mode failed" error message in the top left corner of mVR...and no FSE for me :scared:

It seems completely random so I'm not sure wth is going on here? Would you care for a log?

I've checked the windows event viewer and it seems that PotP asks Windows to disable Aero before going FSE, is that part of the standard FSE procedure?

I've also tried to disable Aero when going FS in the mVR settings but then FSE didn't work anymore.

If Aero helps with tearing in windowed modes but sometimes makes FSE fail, I might just disable it altogether :sly:

Does it support 3D playback via HDMI 1.4? And does it support 4:4:4 at every refresh rate, or just at 60Hz?

No 3D support, 4:4:4 only in 60Hz. 24p is dead smooth, though :)

IIRC only their "6" serie in the 2013 F lineup will allow for 3D but only the "5" serie allows for BFI without FI.......Sammy's FI hiccups on 24p and doesn't look all that good, OTOH their BFI provides the nice added value of removing motion blur and most visible flicker. I'm very sensitive to flicker BTW.

[leeperry]

Humm, I ran more tests and apparently FSE randomly fails if I tell mVR to automatically roll refresh rates when going FS but it doesn't fail if mVR was set to roll refresh rates when playback starts...is that expected behavior? I really like the FS condition, would you care for some logs?

:thanks:

[cyberbeing]

You're comparing apples and oranges here.
I only used a Jinc example, since that's something people are already familiar with as being rather costly in performance.

So the relative costs can vary substantially between different setups.

Which is true, but I can say pretty affirmatively that it is indeed the second most costly option overall in madVR with my GTX 770 setup.

Relatively speaking, it's expensive on mid-range to high-end NVIDIA Kepler compared to all other madVR options except NNEDI3.

[petran79]

with version 0.87.4, if I enable NNEDI3, video doesnt freeze, but I get green color levels all over the video image.
Probably because its Nvidia.

If it was AMD, I'd get red color levels.....

[hannes69]

Ver Greeneyes
It doesn't seem to take much juice (2-3ms) on my laptop's Radeon Mobility HD 4650, which surprises me a little (especially given that it's stuck on the legacy driver).

madshi
Do you have 13.12 drivers installed? As I wrote in the v0.87.0 announcement post, OpenCL only works with AMD GPUs if you update to the latest drivers.

hannes69
Today I played around with my AMD 4550 GPU to test all the new features of madvr. Debanding and the profile stuff is working great.
Just for info concerning the OpenCL stuff: I don´t use it (no need for the features in my setup) and it doesn´t seem to work properly. That´s no problem because this ancient video card could certainly not handle this stuff, even if the implementation of these functions would be correct.
NNEDI3 chroma upscaling: GPU load rises, but the load is independent of the number of neurons ...
image doubling: GPU load stays the same, it doesn´t matter which settings are made (tested with lowres video)
OpenCL DXVA NV12 surfaces: no effect when using DXVA native or copyback decoder (GPU load the same)
OpenCL error diffusion: GPU load stays the same.

I doubt that any of the OpenCL stuff is working with my good old card. GPU-Z reports full OpenCL support for this GPU.
Maybe the old generation cards HD2000 - HD4000 series with legacy drivers don´t support all the OpenCL features.

All no problem for me, just wanted to notify.

Beside that the recent version of madvr works stable for me.

General idea: it would be really nice in a mid-term future that all the computing of video could be freely allocated to CPU/GPU. My CPU is bored at the moment with just decoding the video. I thought the idea of OpenCL was to use CPU and GPU interchangeable for all purposes...
Of course it is too time consuming to make code implementations for GPU and CPU for every little feature, but isn´t OpenCL comparable to a sort of cross-compiler (have no better comparison) so code can rather easily be converted?
Sorry for my non-expert point of view, just interested.

I think it is not possible to install the "normal" (non-legacy) drivers for HD2000-HD4000 series ...