madVR - high quality video renderer (GPU assisted)

[madshi]

madVR v0.85.7 released

http://madshi.net/madVR.zip

* fixed: new 2x/3x shaders sometimes got activated when they shouldn't
* renamed "speed up DXVA ..." options to "don't use copyback for DXVA ..."
* disabled option "don't use copyback for DXVA deinterlacing" by default now

I really hope that all serious bugs are fixed now.

[cyberbeing]

Originally Posted by madshi

Ehm, what? I'm confused. With v0.85.6.0 you couldn't even activate this option. That's why v0.85.6.1 was necessary in the first place. Did you really test with v0.85.6.0? Or maybe with v0.85.5? In any case, how this option works didn't change between v0.85.5 and v0.85.6.1 at all. So I wonder why your test results turned upside down!

You're correct, something is strange with my result, since after re-checking my downloads, both were indeed 0.85.6.1. I suspect the combination of "speed up DXVA deinterlacing at the cost of a little chroma blur" (which was enabled by default) + "use a separate device for DXVA processing" (which I tried enabling first) issue I mentioned in Edit3 was what I was seeing the first time.

[ryrynz]

Madshi, the blank window Overlay issue I can only get that to occur now when it's enabled before restarting, so don't worry bout it.

Regarding "use a separate device for presentation" is there a preferred setting you think for this? Reason I ask is because I have no flushing set at all and to disable it I need to flush/flush and sleep.
I can't see any obvious performance differences in display times or GPU usage between the two configs but I generally strive for the least amount of work my system has to do to get acceptable output.

Originally Posted by 6233638

No, certain videos do start dropping frames when I have that option disabled. Plays fine with it turned on.

Have you tried changing any flush settings, queues etc to help try to prevent that? Perhaps you can look into if indeed it's truly necessary for you.

[glc650]

Originally Posted by ryrynz

Have you tried changing any flush settings, queues etc to help try to prevent that? Perhaps you can look into if indeed it's truly necessary for you.

I need that option (use a separate device for presentation) enabled as well or I get slowly incrementing delayed frames. All my queue and flush settings are at their defaults. 2.4 Core 2 Quad, Radeon 6570, Win 7 x64.

[ryrynz]

Madshi, pausing playback whilst having the OSD open results in a constant almost 40% GPU usage on my HD3000, not a big deal considering but it does seem excessive.

[madshi]

Originally Posted by ryrynz

Regarding "use a separate device for presentation" is there a preferred setting you think for this? Reason I ask is because I have no flushing set at all and to disable it I need to flush/flush and sleep.

It seems several people have better performance with this option enabled. So I would guess the preferred setting is probably to have it enabled, if your GPU drivers don't have a problem with it.

Having this option enabled automatically forces a "flush & wait" after the last render step. So there's really no difference with flush settings, after all.

Originally Posted by ryrynz

Madshi, pausing playback whilst having the OSD open results in a constant almost 40% GPU usage on my HD3000, not a big deal considering but it does seem excessive.

Having the OSD open means that all render queues are emptied every second, so that the paused frame can be updated with the latest OSD information. After emptying the queues, madVR automatically starts working to fill them again. So basically in this state madVR is constantly rendering to fill the queues, then empties them again, then fills them again etc. This isn't really ideal, of course, but then having the OSD on should only be a temporary thing for debugging purposes, so I don't think GPU consumption with OSD on in paused state is very important.

[Andy o]

Hi madshi, thanks for all the updates. Just a quick one... the file name extensions (23p, 23fps, etc) don't seem to be working anymore. I'm coming from 0.55.2. If you can't reproduce it I'll post a log.

[sidi0us]

My rendering time doubled with "use a separate device for presentation" disabled, no frame drops or other problems though.

[pankov]

Andy o,
see the change log.
There was a full rewrite of the file name tags.

[Andy o]

OK thanks, has this been discussed more in depth though? I'm still adding "refreshRate=23" and nothing happens. Same with "frameRate=23". I don't know what the difference is between those two, either.

Here's the log https://docs.google.com/file/d/0B0La...NNV0hMczA/edit (to download the zip go to File-download--stupid google unzips the file right there and obscures a way to download the original zip).

[aufkrawall]

"Don't use CB for DXVA decoding" extremely lowers GPU usage for me.
Can there be a significant difference in quality with it?

[madshi]

Originally Posted by Andy o

OK thanks, has this been discussed more in depth though? I'm still adding "refreshRate=23" and nothing happens. Same with "frameRate=23". I don't know what the difference is between those two, either.

Ooops. There's a bug, making the "refreshRate" tag not work. That will be fixed in the next build. I'll not make a new build just because of this bug, though. Use "refresh=23" instead, that should do the trick, I think. Generally, "refreshRate" is a hint to the display mode changer. "frameRate" overwrites/redefines which framerate the original source has. What you really want is probably turning on film mode / IVTC for this video? In that case use "deint=film" instead of the tags mentioned above. That should also make the display mode changer pick 1080p23 in your case.

Originally Posted by aufkrawall

"Don't use CB for DXVA decoding" extremely lowers GPU usage for me.
Can there be a significant difference in quality with it?

For AMD users this option has no effect at all. For NVidia and Intel users it results in a small chroma blur.

[aufkrawall]

Originally Posted by madshi

For AMD users this option has no effect at all. For NVidia and Intel users it results in a small chroma blur.

Ok, thanks. Regarding GPU usage it makes a huge difference with 4k, but not with 1080p. CPU usage isn't really higher either, my 2500k remains in low power state.

[michkrol]

First of all, thanks for your continued hard work on MadVR. A really good piece of software.

I've encountered a rather strange bug, happens randomly and with quite specific settings. Here's how to reproduce:

1. Enable both "use separate device for presentation" and "use separate device for DXVA processing".
2. Set "DXVA2" for image up-/downscaling.
3. Play a h264 10bit file in windowed mode, while using downscaling (or upscaling).
4. The video is "jumping", as if instead of next frame, the previous frame gets rendered.
   What I mean is instead of frames 1 2 3 4, you get frames 1 2 1 4 or something.

Same file works fine with one of the "use separate device ..." options disabled or when using either other scaling algorithms or no scaling. Also other files work perfectly.

Link to sample:

www.mediafire.com/?ca88a1buue7g2dy

and log:

www.mediafire.com/?z72h4726xh4v859

Basic system info:
Intel Core i5 3570K
Intel HD 4000 + drivers v2875
Windows 7 (x64)
MadVR v0.85.7
MPC-HC Lite 1.6.6.6391 (32bit of course)
LAV Filters 0.54.1, but also happens if using MadVR's internal decoding

[hannes69]

Unfortunately I can't reproduce the problem on my PC. But I found one interesting thing. Your log contains this:

Code:
AnalyzeGpuNv12ToRgb(VideoProcessBlt): +, bitdepth: 32, yFactor: 1.00000018, yOffset: 0.00001520, range: 16.00001805..235.00005700
YCbCr -> RGB decoding matrix used by GPU:
1.00000018 0.00000000 1.73250120
1.00000018 -0.21350089 -0.51525045
1.00000018 2.11633475 0.00000000
If I play your sample on my PC, I'm getting this instead:

Code:
AnalyzeGpuNv12ToRgb(VideoProcessBlt): +, bitdepth: 32, yFactor: 1.00000024, yOffset: 0.00000380, range: 16.00000760..235.00005955
YCbCr -> RGB decoding matrix used by GPU:
1.00000024 0.00000008 1.54000108
1.00000024 -0.18300063 -0.45800027
1.00000024 1.81400121 0.00000008

Remember this problem madshi?
I have redone the tests with newest madvr build. I also noticed that I didn´t have newest VGA bios so I flashed the newest one. Don´t know if the change that now happened comes from VGA bios or madvr (don´t even know if decoding matrix has to do with settings in VGA bios...)

DXVA decoding still delivers right colors. And DXVA scaling for BT709 now seems to be the same like in your setup (very similar numbers in decoding matrix:

AnalyzeGpuNv12ToRgb(VideoProcessBlt): +, bitdepth: 17, yFactor: 0.99998474, yOffset: 0.00000000, range: 15.99975586..234.99641418
YCbCr -> RGB decoding matrix used by GPU:
0.99998474 0.00000000 1.53823853
0.99998474 -0.18417358 -0.45913696
0.99998474 1.81320190 0.00000000

Difference now: bitdepth 17 instead of 32 ???

Using DXVA scaling with BT601 content now leads to following numbers:

AnalyzeGpuNv12ToRgb(VideoProcessBlt): +, bitdepth: 17, yFactor: 0.99998474, yOffset: 0.00000000, range: 15.99975586..234.99641418
YCbCr -> RGB decoding matrix used by GPU:
0.99998474 0.00000000 1.73019409
0.99998474 -0.21400452 -0.51620483
0.99998474 2.11410522 0.00000000

The numbers have slightly changed in comparison to old test, but it´s more or less the same decoding matix. I don´t know the numbers for a correct BT601 decoding matrix, but the numbers given are the ones used.

So now DXVA decoding and BT709 DXVA scaling are right, BT601 I don´t know....

Edit: found out something more:
http://www.equasys.de/colorconversion.html

The BT709 decoding matrix is similar to "YPbPr to RGB color conversion for HDTV" and the unknown matrix which is used for BT601 DXVA scaling in my case is similar to "YCbCr to RGB color conversion for HDTV".

Confusing...

[bugmen0t]

I tried the avisnyth script below to do a 24p to 60p conversion using blending (my first script, so no idea if it's correct).
It causes the madVR render times to decrease massively. How can that be? There must be something wrong, right?

SetMemoryMax(512)
SetMTMode(2)
AssumeFPS(24)
ConvertFPS(36)
SelectEvery(3,0,0,1,2,2)
SetMTMode(1)
GetMTMode(false) > 0 ? distributor() : last

[konakona]

I had problems with hi10p 1080p playback(720 worked fine)-i would drop about 50% of frames. (i think rendering queue was almost empty)
I noticed when trying to play back hi10p 1080p anime for testing that it worked just fine untill subtitles (karaoke) started-i tried messing around with gpu and cpu buffers in madvr settings-set them to 16-cpu and 12-gpu-no improvement at all.
then,i messed with sub-pictures to buffer in subtitles options in mpc-hc. Changing it from 5 to 1 did help,althrough i dont have animated karaoke subtitles now,just static ones. Setting it to 0 made the problem to come back.
It doesnt have any impact visual impact on static subtitles but i can now play back those videos just fine.

And so,i have a question-WHY does subtitles buffer have such impact on playing back such content?

im running xp32bit,phenom II x2 3,1ghz and radeon HD4870(1gb version),and 4gb of ddr3 ram.
my screen is set to 76,416hz@1680x1050 (cant yet get it working to 71,928hz). Downscaling i use is AR filter,linear light Catmull-Rom and chroma upsampling is 3-taps Jinc with AR filter.
its mkv file so mpc used haali media splitter,lav video and madFlac. (latest versions of lav and madvr)

[shimaflarex]

Originally Posted by konakona

I had problems with hi10p 1080p playback(720 worked fine)-i would drop about 50% of frames. (i think rendering queue was almost empty)
I noticed when trying to play back hi10p 1080p anime for testing that it worked just fine untill subtitles (karaoke) started-i tried messing around with gpu and cpu buffers in madvr settings-set them to 16-cpu and 12-gpu-no improvement at all.
then,i messed with sub-pictures to buffer in subtitles options in mpc-hc. Changing it from 5 to 1 did help,althrough i dont have animated karaoke subtitles now,just static ones. Setting it to 0 made the problem to come back.
It doesnt have any impact visual impact on static subtitles but i can now play back those videos just fine.

And so,i have a question-WHY does subtitles buffer have such impact on playing back such content?

im running xp32bit,phenom II x2 3,1ghz and radeon HD4870(1gb version),and 4gb of ddr3 ram.
my screen is set to 76,416hz@1680x1050 (cant yet get it working to 71,928hz). Downscaling i use is AR filter,linear light Catmull-Rom and chroma upsampling is 3-taps Jinc with AR filter.
its mkv file so mpc used haali media splitter,lav video and madFlac. (latest versions of lav and madvr)

Your problem is the high CPU usage of the MPC-HC internal subtitle renderer.
Disable it and use xy-vsfilter.

[konakona]

Originally Posted by shimaflarex

Your problem is the high CPU usage of the MPC-HC internal subtitle renderer.
Disable it and use xy-vsfilter.

dont you think it will decrease quality of subtitles on materials with lower resolution than my desktop?

[shimaflarex]

Originally Posted by konakona

dont you think it will decrease quality of subtitles on materials with lower resolution than my desktop?

I personally don't mind too much having the subtitles rendered at the video resolution. MadVR upscaling algorithms do a fine job making them good when stretched. You should also consider that some typesetting rendered with MPC-HC looks different of what it should be, due to it using a different color matrix and not doing a proper resolution upscaling.

When xy-vsfilter devs finish the support for the new rendering interface, it will be able to render subtitles at higher resolution without those problems.

You can, of course, keep using MPC-HC subtitle renderer for SD videos, if you are not having any slowdown with those...