madVR - high quality video renderer (GPU assisted) [Archive] - Page 646

oddball

30th July 2015, 04:21

I am unable to play any videos using MPC-HC nightly and MadVR latest in Windows 10 (Yeah I know it's only just come out as final etc). I get a Direct3D creation error. I have tried fresh installs of both and set MadVR to FSE off etc. AMD R9 290 BTW with latest Catalyst drivers for Win10.

If you need more info let me know but yeah it does it on all 3 monitors.

MSL_DK

30th July 2015, 09:12

My render, upload and present queue doesn't get filled (1-3/16) when using 1080p60 no problem if i choose 1080p23 (FSE) in win 10 using AMD card no problem in windowed mode.

Ok . It's Direct3D 11 that's causing the problem. Newest driver from AMD on Windows 10 PRO x64 :(

No problem in windowed mode.

EDIT: Have to mention, no problem with win 7 on the same machine.

noee

30th July 2015, 13:00

Ok . It's Direct3D 11 that's causing the problem. Newest driver from AMD on Windows 10 PRO x64 :(

No problem in windowed mode.

EDIT: Have to mention, no problem with win 7 on the same machine.

Same issue here on Win10Pro/CCC15.7.1, but I think it's cosmetic because I've watched 2 complete movies (almost five hours) and had only one dropped frame in each and that was at playback startup.

fwiw, in Windowed mode, it would drop a frame about every 3 minutes even though queue stats were "full" reported.

oddball

30th July 2015, 13:25

http://i.imgur.com/lGhI6YU.jpg

I tried installing DX9 and DX11 in Win10 and it made no difference. I also tried an in place upgrade of Win10 to see if it fixed anything. I also tried switching to the 4770K iGPU since it works on my laptops iGPU with Win10 installed just fine. No go on my main PC. Disabling FSE and changing to DX11 presentation makes no difference either. Ideas?

EDIT: OK I tried MadVR with MPC-BE and it works!

So I think there is an issue with MPC-HC's inbuilt LAV filters. I will test with some external filters like ffdshow or LAV external. I also noticed the other renderers mess up too like EVR custom preset and sync (I get a small video in the top left corner of MPC-HC and the rest is black space). I will check the LAV thread to see if anyone else has mentioned an issue since upgrading to Win10.

EDIT2: ffdshow made no difference as a decoder so I am thinking the error is caused by LAV splitter. I will have to try Haali to test...

EDIT3 Nope! It seems to be something intrinsyc to MPC-HC. I tried ffdshow as the video decoder and Haali as the splitter and still get the same problem.

Filters currently loaded:
- Default DirectSound Device
- madVR
- Audio Switcher
- ffdshow Video Decoder
- LAV Audio Decoder (internal)
- Haali Media Splitter

So I don't really know where to go from here. Madshi can you have a look into this for me?

XMonarchY

30th July 2015, 14:37

So how many people actually contacted NVidia via their Feedback support bug submission? That is the BEST way, IMHO, to relay information directly to NVidia. GeForce or Guru3D or whichever other forums are NOT a good place to talk about things like D3D9 <-> OpenCL Interop support because those forums are for general public and problems (most of which are user errors), not HTPC madVR niche.

XRyche

30th July 2015, 14:55

balanyc

30th July 2015, 15:04

I'm using amd dual graphic laptop
if I set mpc-hc to high performance in catalyst
and set madvr to d3d11, it will crash "only" when entering windowed fullscreen mode.

d3d9 doesn't have any problem also d3d11 in windowed and FSE run fine.

edit: win 10 :D

aufkrawall

30th July 2015, 15:08

nevcairiel

30th July 2015, 16:00

D3D9 interop has been removed deliberately:
https://forums.geforce.com/default/topic/860152/geforce-drivers/official-windows-10-353-62-game-ready-display-driver-feedback-thread-7-29-15-/post/4618371/#4618371

So, chances that it will come back are not best.

I wouldn't say that his statement said "removed deliberately", it could also mean "it didn't work, hence disabled". He does state that its not clear how future support is going to look, he doesn't say its never coming back.
madshi should totally just switch everything to D3D11. ;)

aufkrawall

30th July 2015, 16:39

"not supported on Windows 10" sounds very much deliberately to me.
Well, it's a stupid statement anyway, since it works without any issues with WDDM 2.0 353.49 driver on Windows 10 (or 353.30, which even is WHQL).

I'm somewhat afraid that if madshi fully switches to D3D11, unknown nasty driver issues will show up. Even with current D3D11 mode users have more trouble than with D3D9.
With a SM5.0 NNEDI3 implementation this wouldn't be an issue.

huhn

30th July 2015, 16:40

I wouldn't say that his statement said "removed deliberately", it could also mean "it didn't work, hence disabled". He does state that its not clear how future support is going to look, he doesn't say its never coming back.
madshi should totally just switch everything to D3D11. ;)

but older WDDM 2.0 driverhad this extension.

DX9 interop is not supported on win10 so that is why the file was removed from driver 353.62. Not clear what the plan for support for this is as I am still researching this with the software team.

hard to believe. amd and intel use magic?

and about DX11. a wise men once said: "..., but D3D11 decoding API is complicated and annoying."

doesn't this count for rendering too?

Sunset1982

30th July 2015, 16:40

Hmmm... Sounds bad for the next weeks / months. Changed my gtx970 in my htpc with my r9 380. gtx is now doing his job in my gaming machine. Amd is doing his work in my htpc now.

nevcairiel

30th July 2015, 16:52

It just doesn't make any sense to assume the worst intentions from everyone, what would they have to gain by removing this deliberately, and not for technical reasons?

It's far more likely that "not supported" means it just didn't work properly, and was therefore disabled. Which could simply mean it can get fixed later if they care enough.

It's weird how on the internet everyone assumes a company is doing the worst they usually can and every statement is either a lie or legally binding, depending on what it says :p

TL;DR: don't jump to conclusions without any facts.

PS: updating a PC you rely on everydat to Win10 right away is kinda silly anyway. Everyone should expect hiccups in driver and software support for the first month or two.

madshi

30th July 2015, 16:53

They make it sound as if Windows 10 by design cannot do D3D9 interop, which is AFAIK untrue. D3D9 interop still works with AMD and Intel, no? It seems that NVidia is not aware of that pretty much every HTPC playback software (except maybe PowerDVD, I don't know) out there is still using D3D9 DXVA. Nobody I know of uses D3D11 DXVA. Of course nobody besides madVR uses OpenCL for real time video playback (except MPDN), either, so NVidia removing D3D9 interop support basically means it's only madVR which is screwed.

I'm almost tempted to "stick it to them" and refuse to do anything about it. But then, that would probably hurt users much more than NVidia. Anyway, I will still not change my time schedule. So for now NNEDI3 won't work with NVidia, when using newer drivers.

I guess that means I'll be recommending AMD GPUs from now on. I hope AMD releases next generation GPUs with HEVC DXVA, HDMI 2.0 and HDCP 2.2 soon.

P.S: Of course nevcairiel is right and I might be jumping the gun here. But my experience with GPU companies is not exactly promising...

aufkrawall

30th July 2015, 17:04

I guess that means I'll be recommending AMD GPUs from now on.
I don't think this would be a good idea, or has anyone proven that PCIe 3.0 is fast enough for 1080p60 NNEDI3 doubling?
This is not an uncommon task, since many YT videos now have 60fps and 4k displays are becoming popular.
And with PCIe 2.0, OpenCL NNEDI3 is totally useless either on AMD. And there's probably a bug in the drivers regarding PowerPlay which makes the interop performance even randomly worse for some users.
Bigger AMD GPUs (Hawaii, Fiji) are probably very fast with NNEDI3 SM5.0, I suppose Fiji might even could do already 1080p60 64 neurons.
Sorry, but that OpenCL implementation has become very useless now that NV is going nuts.

madshi

30th July 2015, 17:29

LAV Video Decoder seems to be able to easily do 4Kp24 10bit copyback via PCIe 3.0, without breaking into a sweat, and that requires more bandwidth than 1080p60.

aufkrawall

30th July 2015, 17:33

Is the frame not already doubled to 4k when PCIe copyback happens?

madshi

30th July 2015, 17:42

Ooops, yes, that's true... :o Anyway, I think 1080p24 is the most important, and at least that should work fine. Not sure about 60p.

aufkrawall

30th July 2015, 18:03

Yes, it should work, since PCIe can transfer ~190fps 1080p.
But it's still vulnerable to driver quirks, here someone only has 115fps:
http://www.forum-3dcenter.org/vbulletin/showpost.php?p=10703918&postcount=62

And crazy AMD drivers can interfere even worse, only 36fps with PCIe 2.0:
http://www.forum-3dcenter.org/vbulletin/showpost.php?p=10701293&postcount=12

Some weird PowerPlay behavior:
http://www.forum-3dcenter.org/vbulletin/showpost.php?p=10703924&postcount=63

I don't think a copyback solution would ever be "clean" and thus stable for everybody. AMD still doesn't seem to have fully fixed flawed performance after 13.12 driver or it was never really stable.

With SM5.0, results would be totally reliable. Either GPU ist fast enough or it's not and the driver hardly has any opportunity to interfere.
Maybe peformance can be further improved for SM5.0? e.g. Prefer Vector is much faster on Maxwell than other options.

huhn

30th July 2015, 18:07

there are cards that doesn't get 24 fps.

and 13.12 is just faster in term of openCL performance the number of frames that can be copied is the same.

aufkrawall

30th July 2015, 18:19

there are cards that doesn't get 24 fps.

I'm not sure I'm getting what you mean.
Copyback performance can even be worse with smaller Radeons?
This would be bad. Then probably NNEDI3 could be used on more cards in general if done via SM5.0.

Btw: 300 Radeon series refresh introduced much improved cooling systems like already known from Maxwell cards. Very good for "highend" HTPCs with enough compute power for NNEDI3.

and 13.12 is just faster in term of openCL performance the number of frames that can be copied is the same.
This wouldn't be an issue anymore with SM5.0 either. ^^

You may already guessed it, but I'm extremely sad that I can't use NNEDI3 anymore, since I watch videos on my gaming system which profits by new drivers.
I'm really missing madshi's deband filter in MPDN...

XRyche

30th July 2015, 18:40

I'm having a similar problem as oddball with Windows 10 Pro. Only real difference is that i have an MSI AMD R7 265 OC using Catalyst 15.7.1 , Driver version 15.20.1062.1002(The latest driver specifically for Windows 10 directly from the AMD website not from Windows update.) and I'm getting a black screen, audio is fine. I tried switching to EVR custom with the same MPC-HC and the video shows perfectly fine. So I'm pretty sure it's madVR and not MPC-HC or the built-in LAV filters, for me anyway.

Here's an update: I disabled any Nnedi3, whether image doubling or chroma scaling, and use super-xbr instead. Video was just fine. I was using Nnedi3 for image doubling upto 720p and was using Nnedi3 for chroma upsampling with 1080p before going to Windows 10.

aufkrawall

30th July 2015, 18:58

NNEDI3 doesn't work for you anymore on Windows 10?

XRyche

30th July 2015, 19:11

No sir/mama. I either get a brownish image or no image at all. Also it takes about 2 to 5 seconds for the image to initialize in windowed mode, even without Nnedi3. I'm not using delay playback either.

Update: I used Nnedi3 32 neurons luma and chroma doubling and if I wait about 20 to 30 seconds I get a brownish tinted image. When I try to take a .jpg screenshot with MPC-HC's F5 the screenshot appears normal. I guess I wasn't waiting long enough. But 20 to 30 seconds seems like an inordinately long time, doesn't it?

Further update: It seems the long wait time happens with super-xbr as well. The image is okay but its a long wait for the screen to show an image. I actually drop over 500 frames before I get an image on screen. Although the wait time for an image is almost instantaneous with EVR Custom.

Edit: The excessive wait times were all in FS or FSE.

aufkrawall

30th July 2015, 19:40

That sounds like a driver bug. As usually, OpenCL seems to be more broken than the rest. ;)

XRyche

30th July 2015, 19:50

That sounds like a driver bug. As usually, OpenCL seems to be more broken than the rest. ;)

I'm going to try and do a complete clean install of the driver and see if that helps.

BloodySword

30th July 2015, 20:08

For me, madVR works just fine in MPC-BE on Windows 10, upgraded in-place from Windows 8.1 Pro 54. Well, partially. As soon as I switch to full screen, the player starts to react slowly. When I resize the window of the player, it is really really slow and stuttery. I updated madVR to the latest version 0.88.21 and still no luck, but the video is played back smooth now and there is no video stutter anymore, but the UI of the player is still acting very very slow... Also moving random windows of random applications around is stuttery and has micro glitches...

I may try the exclusive full screen mode instead and see if that makes any difference.

Edit: Exclusive mode does not make ANY difference.

Edit2: As long as I don't go into fullscreen mode, everything is like it was on Windows 8.1 Pro 64. After closing MPC-BE when I was in fullscreen mode, it goes back to normal.

Edit3: The same phenomena CAN occur when opening another video while MPC-BE is still open.

Edit4: It's ShadowPlay! When active, it screws up madVR! When I disable it, then it's not a problem anymore. Everything is then solved... But it would be nice if this problem would be investigated and fixed if possible and if it is a problem on madVR side.

Edit5: Oddly enough: Problem is gone. When I activate ShadowPlay again, the problems does not occur anymore. Seems that there was something hanging from the upgrade procedure to Windows 10. :D

Edit6: Sorry it came back. Disabling ShadowPlay solved the problem again.

Machine:
ASUS X99-E WS
GeForce GTX Titan X 12 GByte with NEC MultiSync EA244WMi over DisplayPort
32 GByte Kingston ValueRAM reg ECC DDR-2133 (Micron Chips)
Intel Xeon E5-2637v3
Intel 750 Series PCIe NVMe SSD 400 GByte
Hitachi DeskStar 4TB 7200 RPM HDD
Windows 10 Pro 64 Bit
MPC-BE x64 1.4.5 beta (build 501) with madVR 0.88.21 x64 renderer.

XRyche

30th July 2015, 21:33

I'm going to try and do a complete clean install of the driver and see if that helps.

No, a clean install using Display Driver Uninstaller (in Safe Mode) didn't change a thing.

a11599

30th July 2015, 21:40

Nobody I know of uses D3D11 DXVA.

Kodi was upgraded for full-DX11 decoding and rendering just recently.

XRyche

30th July 2015, 22:06

XMonarchY

30th July 2015, 22:11

You may already guessed it, but I'm extremely sad that I can't use NNEDI3 anymore, since I watch videos on my gaming system which profits by new drivers.
I'm really missing madshi's deband filter in MPDN...

Didn't you say 353.49 Hotfix drivers work fine?

I see many people complaining about this D3D9 OpenCL interop thing, but I only see myself and 2 other people complaining to NVidia. Maybe they would listen if more people submitted this issue to them via their "Submit Bug" official Feedback/Support system instead of hoping they would read their reply on GeForce forums?

If there is enough demand, NVidia will budge!

aufkrawall

30th July 2015, 22:17

Didn't you say 353.49 Hotfix drivers work fine?

I also said I want newest drivers for games.
In fall, there'll be lots of new games which profit by new drivers.

If there is enough demand, NVidia will budge!
You don't know Nvidia, it seems. ;)

And it will never be "enough" people complaining, as madVR is a non-commercial 1 man project and probably even most of its users are hardly aware of NNEDI3 and how it looks different to other algorithms.

They may change it back, since it's on their bugtracker and madVR got popular for HTPC users, but I'd never bet on that.

nevcairiel

30th July 2015, 22:18

I hope AMD releases next generation GPUs with HEVC DXVA, HDMI 2.0 and HDCP 2.2 soon.

AMD just finished releasing a new GPU series, and it has none of these features, so good luck waiting for that. ;)

aufkrawall

30th July 2015, 22:21

Fiji has HEVC 10 bit DXVA and next year's Arctic Island GPUs will have all the other features as well (up to 100% certain).

nevcairiel

30th July 2015, 22:30

Fiji has HEVC 10 bit DXVA

It does not, only 8-bit, which is unfortunately not very useful for commercial 4K content like UHD Blu-ray (and hopefully streamed 4K as well, as BT.2020 needs 10-bit to be good).
Not that 10-bit would be that ideal on AMD, as 10-bit isn't supported through DXVA-Native, but only Copy-Back, and AMD is notoriously bad at Copy-Back, especially considering you would want all the bandwidth to remain "free" for OpenCL-CopyBack with NNEDI3.

huhn

30th July 2015, 22:39

ashlar42

30th July 2015, 22:40

Well... I was planning anyway to wait a while before upgrading to Windows 10, one reason more...

I have a generic (and probably stupid) questione regarding scaling. With 4k... scaling 1080p to 4k, why shouldn't Nearest Neighbour suffice? It's exactly double the size both horizontally and vertically. That way one should see the exact same pictures with the exact same size of a 1080p pixel (a 2x2 pixels matrix). Otherwise we try to add detail that is not in the original image... am I wrong?

aufkrawall

30th July 2015, 22:44

It does not, only 8-bit, which is unfortunately not very useful for commercial 4K content like UHD Blu-ray (and hopefully streamed 4K as well, as BT.2020 needs 10-bit to be good).
Not that 10-bit would be that ideal on AMD, as 10-bit isn't supported through DXVA-Native, but only Copy-Back, and AMD is notoriously bad at Copy-Back, especially considering you would want all the bandwidth to remain "free" for OpenCL-CopyBack with NNEDI3.
Ah, yes, you're right:
http://www.anandtech.com/show/9390/the-amd-radeon-r9-fury-x-review/8
Another AMD fail *facepalm*. Well, still AI should have it, since I suspect Fiji is just a GCN HBM showcase.

But interesting regarding copyback.
I suppose many people have never been able to try NNEDI3 because the AMD driver sabotaged copyback performance (at least this was the case for me). OpenCL for realtime video-applications has totally failed imho. I really hope it won't be christmas until we have non-OpenCL NNEDI3 in madVR.

huhn

30th July 2015, 22:45

Nearest Neighbour for x2 scaling is a viable choice.

in the end "all" source are scaled before they are pressed on a disc/encoded so scaling them with something other than Nearest Neighbour for x2 can be closer to the original.

no scaler is adding details they most of the time remove aliasing.

aufkrawall

30th July 2015, 22:46

Nearest Neighbour for x2 scaling is a viable choice.
I disagree, 720p looks terrifying with nn on WQHD.

huhn

30th July 2015, 22:50

Ah, yes, you're right:
http://www.anandtech.com/show/9390/the-amd-radeon-r9-fury-x-review/8
Another AMD fail *facepalm*. Well, still AI should have it, since I suspect Fiji is just a GCN HBM showcase.

But interesting regarding copyback.
I suppose many people have never been able to try NNEDI3 because the AMD driver sabotaged copyback performance (at least this was the case for me). OpenCL for realtime video-applications has totally failed imho. I really hope it won't be christmas until we have non-OpenCL NNEDI3 in madVR.

i can use nnedi3 with my old 6770 without problem so not sure what your problem is.

and the r9 270x is the most recommended card for madVR thanks to this very good nnedi3 performance.

huhn

30th July 2015, 22:51

I disagree, 720p looks terrifying with nn on WQHD.

the same as a 720p TV.

nevcairiel

30th July 2015, 22:52

I have a generic (and probably stupid) questione regarding scaling. With 4k... scaling 1080p to 4k, why shouldn't Nearest Neighbour suffice? It's exactly double the size both horizontally and vertically. That way one should see the exact same pictures with the exact same size of a 1080p pixel (a 2x2 pixels matrix). Otherwise we try to add detail that is not in the original image... am I wrong?

That results in terrible aliasing. Basically you are wasting the entire potential of a 4K screen, might as well buy a 1080p screen then.

Good scaling algorithms will provide a much better image.

aufkrawall

30th July 2015, 23:04

i can use nnedi3 with my old 6770 without problem so not sure what your problem is.

It's probably WQHD, I wouldn't bet the AMD copyback performance is scaling nicely with increasing frame size. Or only newer GPUs are affected by a certain bug etc.
And to what resolution would you scale with NNEDI3 on a 6770? It's an extremely slow GPU, judging from today.

and the r9 270x is the most recommended card for madVR thanks to this very good nnedi3 performance.
When scaling to 1080p, the copyback performance is probably less of a problem than the net computing power, where Pitcairn XT is not bad. It's roughly the PS4 GPU.

huhn

30th July 2015, 23:14

It's probably WQHD, I wouldn't bet the AMD copyback performance is scaling nicely with increasing frame size. Or only newer GPUs are affected by a certain bug etc.
And to what resolution would you scale with NNEDI3 on a 6770? It's an extremely slow GPU, judging from today.

When scaling to 1080p, the copyback performance is probably less of a problem than the net computing power, where Pitcairn XT is not bad. It's roughly the PS4 GPU.

1080p like 90%+ of all people. low CR QHD IPS it not for everyone.

noee

30th July 2015, 23:19

I suppose many people have never been able to try NNEDI3 because the AMD driver sabotaged copyback performance (at least this was the case for me). OpenCL for realtime video-applications has totally failed imho. I really hope it won't be christmas until we have non-OpenCL NNEDI3 in madVR.

Granted, my needs are somewhat trivial, but I use NNEDI3/32n on all my transcoded SD 24p sources on a shitty old HD6570 with some minor overclock and other than an occasional dropped frame, seems to work fine, gpu load up around 93%.

If I add h/w deint or smoothmotion (60Hz) or ED and then it's a slideshow.

aufkrawall

31st July 2015, 02:18

1080p like 90%+ of all people. low CR QHD IPS it not for everyone.
I always see 4k TVs in the brochures...

ryrynz

31st July 2015, 02:27

That results in terrible aliasing. Basically you are wasting the entire potential of a 4K screen, might as well buy a 1080p screen then.

Good scaling algorithms will provide a much better image.

Seen a lot of 4K TV's with almost NNEDI3 style upscaling. Just have the media player set to 1920 x 1080 on playback and enjoy, no PC upscaling required. Better off doing it with one of MADVR's scalers I'd say though.

aufkrawall

31st July 2015, 02:37

Have you tested really difficult pictures with it?
It's easy to look just crispy, especially when viewing from larger distance. Also four times higher pixel density helps to hide a lot of artifacts.

Btw: You can use madVR with DVB Viewer beta, and european TV is 720p50. I really doubt you want to use copyback for NNEDI3 -> WQHD (720p x 2) and DXVA2 at the same time, if even just NNEDI3 CB for itself works. It's not 2012 anymore. NNEDI3 was integrated early enough, but now it got borked at an already inconvenient time.

alx000

31st July 2015, 03:59

I also notice when playing 1080p (my display is 1080p) there is no wait times at all. This is going from Windowed player to FS and/or FSE. So I guess the excessive wait times are only when upscaling.

any news?

having the same problem, no image on FSE 1080p ... windowed works fine...

EDIT: just marked to use Direct3D 11 for presentation, and it solved the problem.