View Full Version : madVR - high quality video renderer (GPU assisted)
ryrynz
9th March 2016, 02:17
In all other doubling / downscaling scenarios, 2d works fine (as does jinc), only adds around 2-3ms to render times. When supersampling render times shoot up into the 80s ms with 2d, from around 35ms with 1d. Jinc downscaling also shows the same behaviour when supersampling, its in the osd twice and render times shoot up into the 80s.
I think you may underestimate how demanding Jinc and AR are, only madshi can comment on the performance and OSD though really. No problems have been reported about the performance of Jinc or AR, so I expect that everything is fine here.
iSeries
9th March 2016, 02:22
I think you may underestimate how demanding Jinc and AR are, only madshi can comment on the performance and OSD though really. No problems have been reported about the performance of Jinc or AR, so I expect that everything is fine here.
If it was so demanding it would be the same downscaling 4k to 1080p, but with the 4k samples i have i can use 2d ssim downscaling absolutely fine (GTX950), well below maximum render times (and its only in the osd once). This issue only happens with supersampling.
CarlosCaco
9th March 2016, 02:23
What does the "present a frame for every vsync" option have to do with this?
i was just saying that this option increase a little my render times, without this option checked i having some glitches...
ryrynz
9th March 2016, 02:30
If it was so demanding it would be the same downscaling 4k to 1080p, but with the 4k samples i have i can use 2d ssim downscaling absolutely fine (GTX950), well below maximum render times (and its only in the osd once). This issue only happens with supersampling.
No it wouldn't, you're not factoring in image doubling.
madshi, this will probably be addressed regarding the OSD, but Jinc AR shows up in the OSD when supersampling and downscaling using Jinc with no AR. It shows Image doubler = Jinc AR < Jinc.
iSeries
9th March 2016, 02:36
No it wouldn't, you're not factoring in image doubling.
madshi, this will probably be addressed regarding the OSD, but Jinc AR shows up in the OSD when supersampling and downscaling using Jinc with no AR. It shows Image doubler = Jinc AR < Jinc.
When supersampling, doubling 1080p and downscaling back to 1080p with 1d, using super-xbr to double, and using superres 3 and crispen edges to enhance, i get around 35ms render times. Downscaling with 2d it shoots up into the 80s. That can't be normal. Like i said, all other doubling scenarios 2d only adds around 2-3ms to what 1d costs, and same goes for downscaling 4k native to 1080p, no trouble at all using 2d, again only adds 2-3ms to what 1d costs.
ryrynz
9th March 2016, 03:13
When supersampling, doubling 1080p and downscaling back to 1080p with 1d, using super-xbr to double, and using superres 3 and crispen edges to enhance, i get around 35ms render times. Downscaling with 2d it shoots up into the 80s. That can't be normal. Like i said, all other doubling scenarios 2d only adds around 2-3ms to what 1d costs, and same goes for downscaling 4k native to 1080p, no trouble at all using 2d, again only adds 2-3ms to what 1d costs.
Only 2ms difference between 1D and 2D? I think something's broken, you should maybe look into resetting your settings as my GTX 960 has about 8ms render times when downscaling 4K content and 16-18ms when using 2D (with AR and LL)
tp4tissue
9th March 2016, 07:11
Hey guys.. so , with Madvr.. If I want to use it, I have to stick with the 32-bit.
What do I use to decode 10bit hevc 4K to then connect with Madvr, because my downstairs computer is not fast enough using the 32bit Lav filter to CPU decode that.
I'm a Madvr diehard.. I just thought this would be the place to post this, since you guys are also sticking with 32bit for madvr..
huhn
9th March 2016, 07:19
madVR has a 64 bit version for a very long time so no a lot of us are using 64 bit.
tp4tissue
9th March 2016, 08:09
madVR has a 64 bit version for a very long time so no a lot of us are using 64 bit.
what.. what............................ wait, is it the same download in the OP ?
ryrynz
9th March 2016, 08:37
what.. what............................ wait, is it the same download in the OP ?
The zip contains both the 32 bit version and the 64 bit version. Make sure to run install.bat after extraction.
ileile
9th March 2016, 18:08
Sometimes stereo 3d playback is not smooth, and the "dropped frames" or "presentation glitches" keep increasing. My CPU is not good enough (i3 2100), but I think it is not the reason. Because I tried some times with same movie clip, but the problem is not easy to reproduce.
another problem is that: if I connect two displays and play something using mpc-hc, when I drag the mpc-hc from one display to another, the system often crash. Not BSOD, but something like black screen, system can not respond(ctrl+alt+del usually can't work too, but I can move mouse pointer for a period). this problem may occur with only one display, when switching between window mode and exclusive mode. My GPU is AMD R7 200.
system is windows 10.
chros
9th March 2016, 20:00
what.. what............................ wait, is it the same download in the OP ?
:) yep, yep, I haven't realized it for a long time either, just kept extracting the zip file :)
James Freeman
9th March 2016, 20:19
madshi can you please add a ST.2084 EOTF transfer function under calibration tab?
aufkrawall
10th March 2016, 00:22
Sometimes my GTX 980 gets stuck in the wrong powerstate when opening videos, it doesn't run with full clock then and drops lots of frames. I suspect this is a randomly triggered bug by NNEDI3 in the OpenCL driver, a bit annoying.
Has someone else observed anything similar? Happens with the current driver (364.51) and some previous ones as well. Win 10 x64.
fedpul
10th March 2016, 00:25
Only 2ms difference between 1D and 2D? I think something's broken, you should maybe look into resetting your settings as my GTX 960 has about 8ms render times when downscaling 4K content and 16-18ms when using 2D (with AR and LL)
Hi, I can confirm that supersampling 720, the difference between SSIM 1D and SSIM 2D is between 2-3 ms. I have a GTX 970 but I can also see that when supersampling 1080 the change in rendering times are kind of abnormal, it is a jump of 30-40 ms between 1D and 2D the same haapens with Jinc. I'm not sure if this is normal or not.
I'm also trying to detect what is causing black screens in 90.15. I'm almost sure that is NNEDI in Chroma Upscaling but I need to do some more testing.
fedpul
10th March 2016, 00:27
Sometimes my GTX 980 gets stuck in the wrong powerstate when opening videos, it doesn't run with full clock then and drops lots of frames. I suspect this is a randomly triggered bug by NNEDI3 in the OpenCL driver, a bit annoying.
Has someone else observed anything similar? Happens with the current driver (364.51) and some previous ones as well. Win 10 x64.
It happens to me when I update madvr and try to open a movie without restarting the PC.
bcec
10th March 2016, 00:34
is superxbr + sr4 preferable over nnedi3@64 + sr3 when doing 1080p->2160p upscaling?
I seem to have slight preference to the latter, but I am not sure if it is placebo, or if there are indeed noticeable difference making the latter a better option?
ryrynz
10th March 2016, 01:24
is superxbr + sr4 preferable over nnedi3@64 + sr3 when doing 1080p->2160p upscaling?
I seem to have slight preference to the latter, but I am not sure if it is placebo, or if there are indeed noticeable difference making the latter a better option?
Does it really matter what the other users think? It's your opinion that matters more. Screenshot and compare to see if there's a noticeable difference. Everyone has their own preferences.
Ver Greeneyes
10th March 2016, 01:39
is superxbr + sr4 preferable over nnedi3@64 + sr3 when doing 1080p->2160p upscaling?
I actually tend to stick to sr2 - I find anything higher to be too harsh. But that's pretty much personal preference.
QBhd
10th March 2016, 01:46
madshi, has the logic behind when doubling occurs changed? I can no longer use NNEDI3 to double the luma y when I play back 1280x720 ==> 1024x768. I was really happy with the results when I was able to do this:
http://i.imgur.com/m3hTBoW.png
With 90.15 this is what is happening:
http://i.imgur.com/lxizcHQ.png
This time I have included my settings.bin which you asked for the last time.
settings.bin (https://mega.nz/#!kFAVXRpR!3PUCCyiCW4-jqVECzEECYU2v-WgNeR7ABJgihEbEXJ0)
fedpul
10th March 2016, 02:30
madshi, has the logic behind when doubling occurs changed? I can no longer use NNEDI3 to double the luma y when I play back 1280x720 ==> 1024x768. I was really happy with the results when I was able to do this:
http://i.imgur.com/m3hTBoW.png
With 90.15 this is what is happening:
http://i.imgur.com/lxizcHQ.png
This time I have included my settings.bin which you asked for the last time.
settings.bin (https://mega.nz/#!kFAVXRpR!3PUCCyiCW4-jqVECzEECYU2v-WgNeR7ABJgihEbEXJ0)
That is beacuse in the first picture you are doubling just luma y (vertical resolución if I am not wrong) and in the second you must be using supersampling so both x and y are doubled. But indeed those settings seems overkill. You should be fine with using nnedi 32 for chroma and 64 for doubling luma. Hope this helps.
QBhd
10th March 2016, 02:54
It's the exact same settings... just different versions of madVR. I'm not worried about overkill (noise and heat are non-existant and I don't pay for electricity)... The problem is with the new logic, I can't use anything that doubles the image, it's far too demanding. I don't need to double the x. 1280>2560<1024 is not something I want or need. Doubling the y is rather nice 720>1440<768 is a thing of beauty (to me at least)
QB
fedpul
10th March 2016, 03:18
Your point is good, maybe in the future could be an option for selecting to double just y, x or both.
Enviado desde mi MI 4LTE mediante Tapatalk
bcec
10th March 2016, 04:27
Does it really matter what the other users think? It's your opinion that matters more. Screenshot and compare to see if there's a noticeable difference. Everyone has their own preferences.
Great suggestion! what's the right way to take a screen shot in madvr. I see that printscreen gives me a black screen
Asmodian
10th March 2016, 04:31
You have to turn off "Windowed overlay" to take screenshots.
bcec
10th March 2016, 04:52
You have to turn off "Windowed overlay" to take screenshots.
Thank you!
bcec
10th March 2016, 04:59
Does it really matter what the other users think? It's your opinion that matters more. Screenshot and compare to see if there's a noticeable difference. Everyone has their own preferences.
Thanks for the recommendation!
I actually tend to stick to sr2 - I find anything higher to be too harsh. But that's pretty much personal preference.
After comparing screenshots, you are absolutely right, 4 is quite harsh. I like 2 as well.
[edited]: Also, super-xbr seem to resolve more aliasing than nnedi3@64. Look at the left round frame of his eyeglasses, super-xbr is doing a much better job:
Nnedi3@64:
http://s11.postimg.org/a4xbf8cur/vic_nedi_2.png
Super-xbr:
http://s11.postimg.org/aiynetwyb/vic_xbr_2.png
leandronb
10th March 2016, 05:02
It could be a problem with the video decoder, but it's hard to say for sure. Which video decoder are you using? LAV? Which decoding mode (software, DXVA native, DXVA copyback, CUVID, ...)?
The software i am using to play video:
MPC-HC + LAV filters + AC3 filter + madvr.
And the decoding mode is CUVID, using a gtx 660ti.
ryrynz
10th March 2016, 05:11
Also, superxbr seem to resolve less aliasing than nnedi3@64. Look at the left round frame of his eyeglasses:
I think what you meant to say was that suber-xbr resolves more aliasing than NNEDI3 64 neurons. Super-xbr is looking good there, although the sizes of the images aren't the same.
The top of the ear is another noteworthy mention for suber-xbr here also, NNEDI3 dominates for lower resolution animated content without question though.
bcec
10th March 2016, 05:22
I think what you meant to say was that suber-xbr resolves more aliasing than NNEDI3 64 neurons. Super-xbr is looking good there, although the sizes of the images aren't the same.
yup, that's what I meant (edited the post) :thanks:
btw, the size of the original images are the same. they are just not cropped equally (i did a quick crop of the face before posting)
Warner306
10th March 2016, 06:49
is superxbr + sr4 preferable over nnedi3@64 + sr3 when doing 1080p->2160p upscaling?
I seem to have slight preference to the latter, but I am not sure if it is placebo, or if there are indeed noticeable difference making the latter a better option?
I prefer SuperRes (3) at a 2x scaling factor. I don't like an oversharpened image, but find I can improve the image further by adding crispen edges (1.0), as well. This will make higher SuperRes values look better.
super-xbr and NNEDI3 look similar with SuperRes, but super-xbr will enhance any existing ringing, while NNEDI3 will not. If you can't see ringing, or if doesn't bother you, use super-xbr.
RyuzakiL
10th March 2016, 08:10
Has anyone tried the latest Crimson driver 16.3 on Win10x64, MadVR?
and does MadVR work properly on Win10x64? or is it still Win8.1 that MadVR works perfectly.
ryrynz
10th March 2016, 08:51
Has anyone tried the latest Crimson driver 16.3 on Win10x64, MadVR?
and does MadVR work properly on Win10x64? or is it still Win8.1 that MadVR works perfectly.
Why not test it yourself?
Windows 10 works perfectly for me using a GTX 960 with latest everything.
rxracer
10th March 2016, 09:27
Sorry if this has been asked before. 1840 pages was daunting, and I couldn't find an answer searching.
I have 2 displays. If I set the viewing display to secondary I get dropped frames and glitches. If I'm doing nothing on the PC it drops a frame and produces a glitch every 7 seconds on the dot. I timed it.. bizarre. If I start using the pc on the primary monitor the video drops frames and glitches rapidly.
Is any of this a known issue? I've changed and disabled every single setting in MadVR, to no avail. I use Lav Filters and Zoom Player. Thinking about it now I'll try with MPC-HC. I hope it isn''t the player cos I really don't like any others.
huhn
10th March 2016, 10:08
Why not test it yourself?
Windows 10 works perfectly for me using a GTX 960 with latest everything.
there are problem with deinterlancing with windows 10.
of cause this is not an madVR error.
retrue
10th March 2016, 11:49
I am using last development versions of MPC-HC and LAVFilters.
I updated yesterday from madVR 0.9010 to 0.9015. When I try to play videos I get sound but not image, not in window mode, neither in fullscreen (curiously, with the last version of PotPlayer+MadVR I have image in fullscreen mode but not in window mode). This happens with vids at 1080p and codec H264. It doesn't happen with vids at 480p and codec H264.
The problem appears with version 0.9011. I have returned to madVR 0.9010 and the problem has disappeared.
huhn
10th March 2016, 13:36
are you using image doubling with "always"?
you may just run out of vram.
retrue
10th March 2016, 14:04
are you using image doubling with "always"?
you may just run out of vram.
Image doubling: double luma (64 neurons) and double chroma (32 neurons), always if upscaling is needed.
But with version madvr 0.9010 I don't have problems.
adhara
10th March 2016, 14:07
Hi,
When doing 1080p to 2160p upscale with MadVr, I get some video judders (every 15/20 sec) on my UHD TV (2160p/60Hz & HDCP 2.2).
The video input signal on my TV is well seen as 2160p.
I tested several video rates: 23.976Hz, 24 Hz, ... Same issue.
No dropped/delayed/... frame on MadVr. My GC is a Nvidia 970 GTX.
How come ?
When playing back a 2160p video directly on the TV (with an usb key for example), no problem.
fedpul
10th March 2016, 14:34
Hi, I can confirm that supersampling 720, the difference between SSIM 1D and SSIM 2D is between 2-3 ms. I have a GTX 970 but I can also see that when supersampling 1080 the change in rendering times are kind of abnormal, it is a jump of 30-40 ms between 1D and 2D the same haapens with Jinc. I'm not sure if this is normal or not.
I'm also trying to detect what is causing black screens in 90.15. I'm almost sure that is NNEDI in Chroma Upscaling but I need to do some more testing.
Quoting myself, I have done more tests.
Hi madshi, I kept getting black screens with 90.15, with Chroma NNEDI enabled, Doublind disabled, Chroma Jinc and Doubling NNEDI, and obviously with both using NNEDI. I need to check without using NNEDI at all. I always use NNEDI 32 Neurons for the tests. Black screens are really random, first one occurs after more than 50 minutes of playback, so testing requires lots of time.
nevcairiel
10th March 2016, 15:13
non-framepacked 3D output is broken in 0.90.15.
SBS only has the right eye showing, TB only the bottom, line interleaved is entirely black, and column interleaved behaves just like SBS (which seems odd in itself). On top of that, if i tick the swap eye checkbox, it becomes entirely black all the time.
When I pause playback the image shows up properly, but goes away again when I play again, at least for SBS and TB, interleaved seems to always stay dead.
On the same note, it would be nice if "auto" 3D mode would not switch the resolution if the OS doesn't even offer a stereo 3d mode in the control panel.
RainyDog
10th March 2016, 15:42
Sometimes my GTX 980 gets stuck in the wrong powerstate when opening videos, it doesn't run with full clock then and drops lots of frames. I suspect this is a randomly triggered bug by NNEDI3 in the OpenCL driver, a bit annoying.
Has someone else observed anything similar? Happens with the current driver (364.51) and some previous ones as well. Win 10 x64.
Add MPC-HC, or whichever video player you use, to 3D applications in the Nvidia control panel and change its power status to prefer maximum power.
aufkrawall
10th March 2016, 15:53
I don't think that this would help since the use of OpenCL/CUDA forces the card into a specific powerstate anyway.
leeperry
10th March 2016, 17:44
Enhance details: 1.0 is amazing for resolution
Kinda late to the party and I dunno what voodoo magic is behind "enhance details" but it really looks outstanding @0.7 with NEDI+2X SR@1+FRC+SSIM 2D 25% on SD :eek:
Werewolfy
10th March 2016, 17:45
Are you using DXVA deinterlacing or forced film mode? Does the problem occur with both? And which queue is the first one which is getting into trouble (less than half full)?
I use forced film mode. It seems that decoder queue and ivtc queue limit the other queues, it never fills completely. But the moment I have dropped frames is when render queue drops down.
I tested DXVA and it's not really better. Decoder queue go higher than before but I also have dropped frames when render queue drops down.
Here some videos of my tests, I hope it will be usefull and easier for you to understand.
http://www.filedropper.com/testdeinterlacing1_1
(It's the first time I'm using this file hosting website, I don't know what it worths).
Would be great if the new NVidia drivers improved the situation there, of course!!
Unfortunately, on my side it still doesn't work as intended even with the new Nvidia drivers, glitch handlling mode enabled or not.
Maybe try upping the GPU queue size even more?
I need to use a GPU queue of at least 20 to avoid dropped frames with interlaced NTSC DVDs. Any less gives me occasional and "unexpected" dropped frames. After extensive testing, 20-20-6 works for me perfectly.
(My setup is Win10, factory-overclocked GTX 960 and a dual core Ivy Bridge Pentium, using D3D11 windowed mode.)
Higher GPU queue doesn't help because it's limited by the decoder that's never filled. On my videos, you can see that I even used 128-24-15 without success.
Georgel
10th March 2016, 18:22
Kinda late to the party and I dunno what voodoo magic is behind "enhance details" but it really looks outstanding @0.7 with NEDI+2X SR@1+FRC+SSIM 2D 25% on SD :eek:
It does work nicely with real life videos.
It does not work well with animated videos, inducing grain, and what could be perceived as noise.
Uoppi
10th March 2016, 21:03
there are problem with deinterlancing with windows 10.
of cause this is not an madVR error.
I was just about to say Windows 10 + GTX 960 has worked quite nicely for me too.
Are you referring to some DXVA deinterlacing quality issues? At least deinterlacing seems to work glitch-free to my eyes, so I'm just curious what problems you are referring to.
iSeries
10th March 2016, 22:49
Only 2ms difference between 1D and 2D? I think something's broken, you should maybe look into resetting your settings as my GTX 960 has about 8ms render times when downscaling 4K content and 16-18ms when using 2D (with AR and LL)
It depends on how much downscaling is required but yes there is always some performance cost over 1d. However, specifically when supersampling, 2d adds 40ms to rendering times over 1d. Using 'Everest' as a test:
Example 1 (1080p supersampling):
Chroma - jinc
Doubling - superxbr 100
Enhancements - crispen edges 1, enhance detail 1, superres 3
Downscaling - ssim 1d
=30ms rendering time
Example 2 (1080p supersampling):
Chroma - jinc
Doubling - superxbr 100
Enhancements - crispen edges 1, enhance detail 1, superres 3
Downscaling - ssim 2d
=70ms rendering time
As you can see, the only difference between the two examples is 1d / 2d, with 2d adding 40ms to rendering times. Surely, that can't be right? Like I said, in every other downscaling or doubling and then downscaling scenario 2d is perfectly usable.
huhn
10th March 2016, 23:13
I was just about to say Windows 10 + GTX 960 has worked quite nicely for me too.
Are you referring to some DXVA deinterlacing quality issues? At least deinterlacing seems to work glitch-free to my eyes, so I'm just curious what problems you are referring to.
amd crimson deint is simply broken it is using a low quality bob. i haven't tested the newest beta yet.
and nvidia deint was kind of "broken" the last time i check and had issue with telecine/hybrid content. the high quality CUVID deint is not usable at all under windows 10.
huhn
10th March 2016, 23:14
I don't think that this would help since the use of OpenCL/CUDA forces the card into a specific powerstate anyway.
and that powerstate is the highes possible so it can help here.
vBulletin® v3.8.11, Copyright ©2000-2026, vBulletin Solutions Inc.