madVR - high quality video renderer (GPU assisted)

[xv]

fullrange and height problem is libav/ffmpeg only. The height problem is also only with 4:4:4.
I found the problem for the crashes with "many files" (I tried files from my test dir that contained a lot of 4:4:4 ones): madVR/ffdshow crashes with 4:4:4 files if ffdshow output is RGB, if ffdshow output is YV12 image is a mess (just some colorful changing horizontal lines), older madVR or other renderer works fine.

Found another problem: Intel decoder cannot decode 4:4:4, but madVR accepts connection.

Example files:
4:4:4 YUV: http://www.mediafire.com/?byy5zyv46wrdmru
Internal ffmpeg has chroma plane offset, has ffdshow problem described above, other renderers work.
4:2:0 YCgCo: http://www.mediafire.com/?6mupk5zvxv7og17
Crashes internal ffmpeg (fullrange)
4:4:4 YCgCo: http://www.mediafire.com/?i57edq05b6s6e47
Crashes internal ffmpeg (fullrange), has ffdshow problem described above, other renderer works fine.
RGB: http://www.mediafire.com/?cnhaaej2p11wlub
Crashes internal ffmpeg (fullrange), has ffdshow problem described above, other renderer works fine.

Also found another problem: madVR cannot decode H.264 for some mp4 files (lavsplitter and mpc internal tested) correctly, Intel decoder seems to work (seeking is broken), ffmpeg/libav does not seem to work at all for some files. After remuxing to mkv it works.

[egandt]

Tested with Zoomplayer 8.0RC2 I can use MADVR for decoding (which seems to work fine as all as I select madvr as the render and decoder), but if I do so there are no sub-titles displayed, I noticed in a previous post it was commented upon that you might be adding them in the future, does that mean that as it is configured today, subtitles will not work when using the internal decoders? I ask because I watch a lot of softsubbed movies and Anime, normally in the form ASS, but SRT is also a possibility, I use directvobsub (currently 2.40.3093) to handle these subtitles.


Also, just because it does not get said enough, great job, and thanks for your hard work.
ERIC

[Hypernova]

Originally Posted by oddball

I'm confused. What's this about MadVR decoding? I tried with latest MPC-HC and do not see any decoding options (x264 in MKV)? All I get is MS DTV decoder if I block all my filters.

It's in madVR option. Folder "Processing"

[pankov]

madshi,
I've managed to do a quick test of the new 0.67 version and I have a bug to report:
when I start the playback directly in fullscreen there is audio and video but ZoomPlayer is not responding and I can't go out of exclusive mode ... or do anything else - windows is not responsive except for Ctrl+Alt+Del. Luckily there is a "Guard dog" in ZoomPlayer MAX and after the configured time it restarts the player. This did not happen with 0.66 and the only thing I did was to copy the new files over and re-register the filter using install.bat.

Now about the new decoders:
They work but I can't see myself using them because of lack of deinterlacing and subtitles support (no way to insert DirectVobSub or FFDShow's subtitles filter which are currently the only two ways to have subs in ZoomPlayer).

[Thunderbolt8]

thanks for the new version.

regarding the "delay playback start until render queue is full", I definately like to have it at the beginning of the movie, but I agree that there should be an option to disable it for seeking if desired. sometimes you just want to seek through the movie, because you are looking for something and then you mostly want to be able to do this as fast as possible.

as for vc-1, I'm for anything which improves it (speedwise), especially anything which improves decoding speed of interlaced vc-1, because that seems to require quite more ressources as the average AVC stuff, and often my system struggles with it (is multic core decoding actually supported here? afaik thats not the case with the other directshow decoders, so there could be some easy performance gain in there).

edit: ok, how can I actually see that these decoders are working right now? Im on windows 7. Is there any additional entry in the mph-hc > filters menu at playback?

btw. do these decoders work together with ffdshow stuff like postprocessing settings or others? so far, Ive always had ffdshow running with e.g. coreavc to accept raw video input and then apply stuff like sharpening. is this still possible? if not, would it be possible to add this then, otherwise, at least for me, I wouldnt be able to use these decoders as long as I dont have a display which can play back the content at an 1:1 pixel ratio.

[oddball]

Originally Posted by Hypernova

It's in madVR option. Folder "Processing"

Ah yeah. Found it. It's nice. But I lose ffdshow for processing video in the chain which I use for filters only so back to CUVID for me.

[pankov]

madshi,
I'm sorry but I made a mistake and my report is not entirely true. After a few more tests here are more details:
the problem happens only if I try to go out of fullscreen mode while madVR is in Exclusive mode. If madVR is in Windowed mode there is no problem to "unzoom". The problem is present in v0.66 too.
Here is a log file showing the problem with v0.67
http://www.mediafire.com/?350vxa2lsgu2ib0
I've started a movie, waited a couple of seconds and then double-clicked the video window to go out of fullscreen mode but nothing happened. Since that moment I tried different keyboard shortcuts (Ctrl+Shif+Esc; Windows button; Ctrl+J and others) to make ZoomPlayer/madVR go out of Exclusive mode but nothing helped until just after 20 seconds when ZP's Gurad dog restarted the player.

[cyberbeing]

madshi, can you update madVR [benchmark] to 0.67 so we can get an idea of how the built-in decoders perform on our systems?

Not really a functional limitation, but the Intel decoder shows significantly higher CPU load decoding h264 compared to ffmpeg/libav on an AMD X2.

Edit: madVR obviously no longer works with VSFilter either when the built-in decoders are enabled. Any further thoughts about hacking the VSFilter connection to make it work, or will we just be stuck using ISubRender for subtitles when the built-in decoders are enabled?

Edit2: There is a bug which causes madVR to display a Black Screen indefinitely when first opening a video and the built-in decoders and refresh rate changer are enabled. Re-opening the video after the refresh rate has changed is a workaround. Disabling of the Delay till Render Queue is Full option makes no difference.

Debug Log: http://www.mediafire.com/?wg8ncuhmwdgytic

Edit3: Are you able to max out the merit of the internal decoders? The 0x00800200 merit of CoreAVC (set to preferred) and the crazy 0xff800001 merit of FFDShow cause them to get used instead.

[Hypernova]

Madshi, it seems the subtitle will disappread only if the OSD (Ctrl+J) is on. Otherwise it stays.

[azaze1]

MPC-HC crashes for me now when choosing madVR renderer. If I move to EVR Custom Pres, it's fine. It was hit or miss all day, but towards the end of the day, launch any x264 mkv with madVR as output renderer crashes mpc-hc upon launch.

info: Windows 7 32bit
nvidia: GT 430 (275.33 drivers)
LAV Splitter 0.30
LAV CUVID 0.9 (tried other decoders too, mpc-hc internal, cyberlink's from pdvd 11, etc)
latest nightly MPC-HC x86, actually tried a few thinking it may have been that, no dice.

Sorry if this is not an issue with madVR, but it's brand new behavior after unregistering 0.66 and registering 0.67. I noticed the properties of the renderer options still said 0.66 after I unregistered and registered the 0.67. Only after I unregistered and re-registered for a 2nd time did it say 0.67 in properties. But the crashing occurred ever since the first upgrade attempt.

[ryrynz]

Love the new stats display.

Originally Posted by cyberbeing

madshi, can you update madVR [benchmark] to 0.67 so we can get an idea of how the built-in decoders perform on our systems?

Seconded.

[Mikey2]

Wow, thanks for this fascinating update! I did about 30 minutes worth of testing, without trying to specifically break things and find boundary conditions, I found that the core functionality seems to work for me! (Relevant info: Intel Q6600 Quad, NVidia 8600 GT, Windows 7 x64, MPC-HC 1.5.2.3329, tried both LAV Splitter and Haali (LAV worked more smoothly, but then it always has on my system,) and ReClock to a Pioneer Kuro 50" Plasma (accepting RGB 0-255 ...primary display.) I focused on ffmpeg on 1080p h.264 movies.

I think I had less problems since because of my good CPU and out-dated GPU I switched to software decoding a while back. However, this brings up a question: I switched to Software Decoding (ffdshow) because from what I understood, as long as one is not doing any post-processing, it does not matter where the decoding is done (neither the filter itself nor its processing location, such as using GPU-located Cuda in CoreAVC.) In other words, I thought "decoding" was a mathematical process that will yield the same results, simply taking a file such as H.264 file and converting to a 4CC code (such as YV12.)

If this is the case, what is the advantage of doing it on your filter rather than on a separate filter where we have the option to do post-processing?

Does this have to do with the 8Bit+ color processing?

If so, can I expect to see any picture-quality improvements right now?

If not, are there any performance enhancements in doing the decoding and rendering within your filter rather than via a separate filter?

Oh, and two more observations:

1) As was mentioned earlier, I was also seeing some issues with ReClock. If the video was not previously played in ReClock (using a different decoder,) ReClock is reporting that it is refreshing based on the default frame-rate (as set in ReClock's settings panel.) However, if it was already played and the frame-rate was already in its database, then it was correctly reporting the correct fps. (Still, it was a little weird, it only reported the fps and not the FOURCC code as it does on other files.) I can show you screenshots if you would like...For now as a workaround I am going to play the file using ffdshow as the decoder first (until the framerate is confirmed by reclock and added to their database. (This is shown when the clock icon switches from blinking red/green to a solid green.) Only then will I block ffdshow and use your decoder...

2) This brings me to an annoyance that has existed for some time: Can you possibly enable the filter to make configuration changes without already playing a video? This is the only DirectShow filter (that I know-of) that does not allow this. I never minded it too much in the past, but with significant functional changes such as using a decoder or not as an option, I would like to make those selections without already playing something. (Heck, it can even be an issue to some people in the future if they only have their systems configured to use madVR as the decoder - one cannot even turn the decoder on before using some other decoder just to play a video and get into the setting screen in the first place!

Anyway, I will echo other comments here: I LOVE this sort of transparent development. As a software engineer myself, I find myself too often stifled from my most creative ideas because of the constraints placed by normal software development. (E.g. this release would never fly in a normal release world, but you have a bunch of super-intelligent nerds here that love this stuff and are happy to play with new functionality, even if it isn't perfect yet. Involving us right from the start (and actually listening to us) is a big part of why I think this is by far the best renderer out there.

Thanks again!
MikeY

[madshi]

Originally Posted by pirlouy

You're the man for audio renderer; stop saying you're not interested; think a little bit at others instead of being selfish !

Haha! I simply don't have enough time. If I did start working on audio, madVR would suffer. So no, I have no plans working on the audio side of things.

Originally Posted by Wile-E-Coyote

Madshi, if you can have MadVR decode MPEG2, can you not go all the way and have it work with DVDs?

Because making DVDs work is not as easy as it sounds. It's on my to do list, as are many other things.

Originally Posted by robpdotcom

Yes. For instance, if I set madVR to present 8 frames in advance, the OSD shows "8-8/8". If I use ffdshow, it fluctuates, but never goes higher than "7-8/8".

Ah ok, well, this difference is really minor. I'd say 7-8 is pretty much the same as 8.

Originally Posted by robpdotcom

Tried with Haali, LAVSplitter, and MPC-HC's internal splitter with no change.

Well, the only VC-1 files I have are from Blurays, and most are letter-boxed. But I did find four that are not letter-boxed, and they all showed the same corrupted image.

Ah yes, I can reproduce it. It only happens with libav/ffmpeg, but not with Intel, correct?

Originally Posted by HitomiKun

libav/ffmpeg for h264 works fluid and doesn't drain a lot CPU power but it shows animated subtitles (.ASS files) slightly laggy and also a bit timed off. It gets very confusing on fast scenes, because the subs aren't 100% timed right with libav/ffmpeg.

This is caused by the bad libav decoder timestamps in the current madVR version. This will be fixed in a future version.

Originally Posted by Hypernova

2. The combination of Intel decoder+unchecked delay playback will usually crash MPC-HC on my setup after a couple successive seekings.

Ok, will check that.

Originally Posted by Hypernova

Also, the a/v unsynced problem on h264 with avi when using multithread is presented with both libav/ffmpeg and Intel decoder. This make me wonder since isn't this supposed to me libav/ffmpeg only bug?

I don't know. Do you have a sample which shows this problem clearly and quickly?

Originally Posted by fairchild

Also the new yCMS seems to be giving better results than in the last version. (not sure if you and yesgrey changed anything around)

yCMS hasn't changed for months. But the way madVR uses yCMS has changed. 2-3 versions ago madVR/yCMS were pretty much broken. Now it should work fine again, so maybe that's the "improvement" you're seeing? Or is it better than e.g. 5 versions ago, too?

@cyberbeing, have you done your yCMS comparison (before/after the changes I introduced)? What did you find?

Originally Posted by Hprd

Well the issue with the ques running out with the one option checked still extists.

There's probably not much I can do about that. Seems to be an issue with the GPU drivers.

Originally Posted by xv

fullrange and height problem is libav/ffmpeg only. The height problem is also only with 4:4:4.
I found the problem for the crashes with "many files" (I tried files from my test dir that contained a lot of 4:4:4 ones): madVR/ffdshow crashes with 4:4:4 files if ffdshow output is RGB, if ffdshow output is YV12 image is a mess (just some colorful changing horizontal lines), older madVR or other renderer works fine.

Found another problem: Intel decoder cannot decode 4:4:4, but madVR accepts connection.

Example files:
4:4:4 YUV: http://www.mediafire.com/?byy5zyv46wrdmru
Internal ffmpeg has chroma plane offset, has ffdshow problem described above, other renderers work.
4:2:0 YCgCo: http://www.mediafire.com/?6mupk5zvxv7og17
Crashes internal ffmpeg (fullrange)
4:4:4 YCgCo: http://www.mediafire.com/?i57edq05b6s6e47
Crashes internal ffmpeg (fullrange), has ffdshow problem described above, other renderer works fine.
RGB: http://www.mediafire.com/?cnhaaej2p11wlub
Crashes internal ffmpeg (fullrange), has ffdshow problem described above, other renderer works fine.

Thanks!!

Originally Posted by xv

Also found another problem: madVR cannot decode H.264 for some mp4 files (lavsplitter and mpc internal tested) correctly, Intel decoder seems to work (seeking is broken), ffmpeg/libav does not seem to work at all for some files. After remuxing to mkv it works.

Can I have a sample for that, too, please?

Originally Posted by egandt

Tested with Zoomplayer 8.0RC2 I can use MADVR for decoding (which seems to work fine as all as I select madvr as the render and decoder), but if I do so there are no sub-titles displayed, I noticed in a previous post it was commented upon that you might be adding them in the future, does that mean that as it is configured today, subtitles will not work when using the internal decoders?

When using ZoomPlayer and the internal madVR decoders, subtitles will not work, because ZoomPlayer doesn't have its own subtitle rendering engine. So if you need subtitles and want to use ZP, you should use external decoders for now.

Originally Posted by pankov

Now about the new decoders:
They work but I can't see myself using them because of lack of deinterlacing and subtitles support (no way to insert DirectVobSub or FFDShow's subtitles filter which are currently the only two ways to have subs in ZoomPlayer).

Yes, in your situation the internal decoders don't make much sense at the moment.

Originally Posted by Thunderbolt8

regarding the "delay playback start until render queue is full", I definately like to have it at the beginning of the movie, but I agree that there should be an option to disable it for seeking if desired.

More opinions on that, anybody?

Originally Posted by Thunderbolt8

as for vc-1, I'm for anything which improves it (speedwise), especially anything which improves decoding speed of interlaced vc-1, because that seems to require quite more ressources as the average AVC stuff, and often my system struggles with it (is multic core decoding actually supported here? afaik thats not the case with the other directshow decoders, so there could be some easy performance gain in there).

libav/ffmpeg doesn't support interlaced VC-1 decoding at all. You could try whether the Intel decoder decodes interlaced VC-1 faster than the MS VC-1 decoder does. Does it?

Originally Posted by Thunderbolt8

edit: ok, how can I actually see that these decoders are working right now? Im on windows 7. Is there any additional entry in the mph-hc > filters menu at playback?

You can see that by the lack of any video decoder in the filters list. If you don't see any external decoder there, then it must be madVR doing the decoding.

Originally Posted by Thunderbolt8

btw. do these decoders work together with ffdshow stuff like postprocessing settings or others? so far, Ive always had ffdshow running with e.g. coreavc to accept raw video input and then apply stuff like sharpening. is this still possible? if not, would it be possible to add this then, otherwise, at least for me, I wouldnt be able to use these decoders as long as I dont have a display which can play back the content at an 1:1 pixel ratio.

If you want to use ffdshow processing then the internal decoders don't make much sense for you.

Originally Posted by oddball

Ah yeah. Found it. It's nice. But I lose ffdshow for processing video in the chain which I use for filters only so back to CUVID for me.

Yes.

Originally Posted by pankov

the problem happens only if I try to go out of fullscreen mode while madVR is in Exclusive mode. If madVR is in Windowed mode there is no problem to "unzoom". The problem is present in v0.66 too.

It seems that it's ZP which is frozen in this situation? Have you asked Blight for help? It might be more of a ZP problem, but I'm not sure...

Originally Posted by cyberbeing

madshi, can you update madVR [benchmark] to 0.67 so we can get an idea of how the built-in decoders perform on our systems?

Yeah, will do that with the next build. Will need to double check that the benchmark works correctly with the internal decoders, too.

Originally Posted by cyberbeing

Not really a functional limitation, but the Intel decoder shows significantly higher CPU load decoding h264 compared to ffmpeg/libav on an AMD X2.

Yeah, that's my impression as well. How about VC-1 and MPEG2 performance of the Intel decoders, compared to other decoders?

Originally Posted by cyberbeing

Edit: madVR obviously no longer works with VSFilter either when the built-in decoders are enabled. Any further thoughts about hacking the VSFilter connection to make it work, or will we just be stuck using ISubRender for subtitles when the built-in decoders are enabled?

For now, if you want to use the internal decoders, you're stuck with the internal subtitle render engings of MPC-HC, KMPlayer, PotPlayer and J.River MC16. Which I don't consider too much of a problem because these engines produce higher quality subtitle rendering compared to VSFilter, at least for SD content. Or is there any big disadvantage of the internal MPC-HC etc subtitle engines, compared to VSFilter?

Originally Posted by cyberbeing

Edit2: There is a bug which causes madVR to display a Black Screen indefinitely when first opening a video and the built-in decoders and refresh rate changer are enabled. Re-opening the video after the refresh rate has changed is a workaround. Disabling of the Delay till Render Queue is Full option makes no difference.

Ok, will check if I can reproduce that.

Originally Posted by Hypernova

Madshi, it seems the subtitle will disappread only if the OSD (Ctrl+J) is on. Otherwise it stays.

Correct. Basically the subtitles disappear if the same frame is redrawn in paused state. That's not really a bug in madVR, it's a bug in the VSFilter implemention used by MPC-HC. I could avoid it by storing all frames in the decoder queue twice, but that would double the memory usage.

Originally Posted by azaze1

MPC-HC crashes for me now when choosing madVR renderer. If I move to EVR Custom Pres, it's fine. It was hit or miss all day, but towards the end of the day, launch any x264 mkv with madVR as output renderer crashes mpc-hc upon launch.

Can you retry v0.66, just to make sure that it's a new problem introduced by v0.67? Thanks.

Originally Posted by Mikey2

However, this brings up a question: I switched to Software Decoding (ffdshow) because from what I understood, as long as one is not doing any post-processing, it does not matter where the decoding is done (neither the filter itself nor its processing location, such as using GPU-located Cuda in CoreAVC.) In other words, I thought "decoding" was a mathematical process that will yield the same results, simply taking a file such as H.264 file and converting to a 4CC code (such as YV12.)

If this is the case, what is the advantage of doing it on your filter rather than on a separate filter where we have the option to do post-processing?

Does this have to do with the 8Bit+ color processing?

If so, can I expect to see any picture-quality improvements right now?

There's no quality difference to be expected with normal 8bit 4:2:0 content. However, right now no external decoder can decode high bitdepth or 4:2:2/4:4:4 content and pass it on to madVR without quality loss. But then, LAV Video Decoder is going to be released soon which will do all that.

Originally Posted by Mikey2

If not, are there any performance enhancements in doing the decoding and rendering within your filter rather than via a separate filter?

There might be very small performance improvements, but right now the whole chain is brand new and not very much optimized yet, so you probably won't see better performance at this point in time. Not sure if there will ever be a *measureable* performance advantage in the future.

Originally Posted by Mikey2

Still, it was a little weird, it only reported the fps and not the FOURCC code as it does on other files.

The decoded FOURCC is not known to Reclock because the FOURCC conversion stuff is all done internally in madVR.

Originally Posted by Mikey2

This brings me to an annoyance that has existed for some time: Can you possibly enable the filter to make configuration changes without already playing a video? This is the only DirectShow filter (that I know-of) that does not allow this.

It annoys me, too, sometimes. It's just a question of where to spend my programming time.

[Hypernova]

About the subtitle: I see. Thank you for clarification

About a sample, I'll quote my post in ffdshow thread:

Originally Posted by Hypernova

I don't know if I did any wrong but only the first part seem to playable after splited by DGSplit. Please use the second half where the karaoke appear to see the lag.

http://www.mediafire.com/?9jf19139i6e0uh2

It supposed to be libav/ffmpeg only, so if Intel decoder has the same problem maybe it give some clues for ffmpeg developers?

[ForceX]

Originally Posted by Thunderbolt8

thanks for the new version.

regarding the "delay playback start until render queue is full", I definitely like to have it at the beginning of the movie, but I agree that there should be an option to disable it for seeking if desired. sometimes you just want to seek through the movie, because you are looking for something and then you mostly want to be able to do this as fast as possible.

I also agree with this.

So far my only real issue has been with subtitles blinking while using the internal libav decoder, but it's still nice to have a good 10bit decoder. I expect the fansubbers to move to Hi10P. I hope to see a libass-based sub renderer from you in the future.

[madshi]

Originally Posted by Hypernova

It supposed to be libav/ffmpeg only, so if Intel decoder has the same problem maybe it give some clues for ffmpeg developers?

Don't know, I'll see what I'll find out.

Originally Posted by ForceX

So far my only real issue has been with subtitles blinking while using the internal libav decoder

That should be fixed in a future version.

[Hprd]

Yeah, i was thinking of saying it might be my end (as now that i think of it, ever since updating to 275 quadro drivers i've seen this issue, which was about when .63/64 came out), but just kinda wanted a response from you.

Anyhow, keep up the good work then!

[cyberbeing]

Originally Posted by madshi

Or is there any big disadvantage of the internal MPC-HC etc subtitle engines, compared to VSFilter?

The internal MPC-HC subtitle filter is much slower, which can make using pre-buffering cause lots of flickering with complex subs on slower computers. It seems it can flicker even when pre-buffering is disabled as well. Other than general slowness/flickering, I'm not the best person to ask, but I remember it having issues in the past causing it to produce non-identical output compared to VSFilter with certain subtitles. It could be better now, I don't really know. If you go on IRC #cccp @ irc.rizon.net or #aegisub @ irc.rizon.net someone may be able to give you a better idea of the current issues it has (another option would be making a post on AnimeSuki in the Fansub Groups board).

Bug: Just now I discovered a bug in madVR 0.67 which causes stuttering with moving subtitles which doesn't exist in VMR9. It only happens when the madVR ffmpeg decoder is enabled, so it may have to do with the bad timestamp problem you spoke of?

Another Bug: madVR crashes with the Intel Decoder and 10-bit H264.

[madshi]

Originally Posted by cyberbeing

The internal MPC-HC subtitle filter is much slower, which make using prebuffering cause lots of flickering with complex subs. It can flicker even when prebuffering is disabled. Other than general slowness/flickering, I'm not the best person to ask, but I remember it having issues in the past causing it to produce non-identical output compared to VSFilter with certain subtitles. It could be better now, I don't really know. If you go on IRC #cccp @ irc.rizon.net or #aegisub @ irc.rizon.net someone may be able to give you a better idea of the current issues it has (another option would be making a post on AnimeSuki in the Fansub Groups board).

Well, I'll keep that in mind for later.

Originally Posted by cyberbeing

Bug: Just now I discovered a bug in madVR 0.67 which causes stuttering with moving subtitles which doesn't exist in VMR9. It only happens when the madVR ffmpeg decoder is enabled, so it may have to do with bad timestamp problem you spoke of?

Another Bug: madVR crashes with the Intel Decoder and 10-bit H264.

Yes, the stuttering should hopefully be fixed once I implement proper timestamp handling. And the 10bit Intel crash will be fixed in the next build, too.

[cyberbeing]

madshi, would it be possible to make it so madVR configuration can be launched without playing a video?