MPC-HC tester builds for internal renderer fixes

[Hera]

Originally Posted by JanWillem32

@Hera & RGold: I'll just try again...

The latest version is better, it is just 'not there' yet.
I am not able to get it to speed up it seems.

[gilic]

I want to update my bug report. If "subpictures to buffer" is 0 mpc-hc crashes, if it's 1 or 2 the moving text on top for that specific file isn't shown/flickering and if its 3 or up everything seems fine "allow animation" is disabled. I've always struggled to find the right combination of settings for subtitles. Everything I tried so far had issues with different files.

[JanWillem32]

For basic debugging: http://forum.doom9.org/showthread.ph...34#post1553934 . Note that the DirectX debug runtimes are only used to debug the client-side of the graphics interfaces, these don't debug the program itself. To complete the text I linked; I use a very different debugger. The debugger in the compiler suite can hook the program being run and re-trace every instruction in the debug build. That's a bit 'deeper' than just executing it. This type of debugging is near-ideal, but I doubt many people here will have the compiler suite installed.

@fagoatse: Debug builds are the same as release builds, but really slow and full of extra data for error checking and tracing. There's no difference in the handling of player settings.

@Hera: That's at least something, although I don't know what part of the code fixes it (else I would add a comment to it to point out the issue). I'll try the debugger on my other machine tomorrow, maybe I'll find out some more.

@gilic: What I mostly was looking for, is a debug trace text segment, just before crashing.
If "subpictures to buffer" is 0 a different queue handler is used. I'll test that one tomorrow. I never knew it could be sensitive to this kind of thing. (The queue system doesn't create subtitles, it only stores finished subtitles for the video renderer to use.) Don't worry too much about setting the "subpictures to buffer" higher. I've re-written some of the memory functions for it. It's pretty fine now to use 12 (and maybe some more) for even the more extreme karaoke (as long as your video card has enough memory). The reference "Touhou - Bad Apple!!" sample seems to render better with more buffers than what the trunk build prefers.

[gilic]

Are there any negative side effects with a high subpicture buffer? Also would ollydebug be of any use for your purpose?

[Hera]

Sound sometimes stutters when MPC crashes (new IMO) / get stuck when seeking, offset also seems to be in 2x,xxx mark.
Seeking has less problems with smaller resolution files.

Uhm, using Haali Splitter, same behavior as seeking can be reproduced when Haali switches from one file (that is like Opening Song) to the main file (the episode itself). This doesn't happen with LAV Splitter because it still doesn't support having Opening / Ending in separate files.

Also switching to D3D FS seemed to attempt to switch instantly... and I had to kill it because it failed.

[JanWillem32]

@gilic: More subpictures means more video memory consumed. I've made the texture management render-size dependent. For a subtitle with data in every corner of the screen the memory consumption will be quite large.
Example, "Touhou - Bad Apple!!" sample on a 1080p screen: Aspect ratio correction will limit the base render area to 1440×1080. Because this sample renders in pretty much every corner, no further area reduction is used. (I should edit this to handle each piece of text separately, right now it can only blend one composed picture per video frame.) The pixel format is A8R8G8B8 (the color and alpha quality is actually even less than that). That means that there's about 1440×1080×32 bits (5.9326171875 megabyte) in each subtitle texture. That's quite a lot, but doesn't have to be a big problem if there's plenty of video memory left to buffer a bit. Normal subtitles use a smaller drawing area, and will therefore consume less memory.
If you know how to work with OllyDbg, go ahead and try. I haven't worked with it yet.

@Hera: Okay, I'll try to debug tomorrow with the Haali splitter active. I've also seen some problems with switching D3D mode on and off.

[Hera]

I think it just got stuck the same way as the seeking / haali-file-switching problem. I think... I am not a good tester :P

Also, on Windows 8, I do not get the seek bar in D3D FS mode when moving the mouse cursor down the screen. I can make it appear, by... well... selecting the video by clicking on the video a few times (clicking causes pauses)?
Mouse clicks seem to go through the video before getting the seek bar responding to the mouse - who to blame - bleep if I know

EDIT:
Another behavior I noticed - video stops playing, not during anything weird, just during normal playback.

EDIT:
Another behavior I noticed, after un-pausing, video plays, audio is no where to be found. Pausing and Un-pausing causes audio / video de-sync...

EDIT:
When going transitioning to OP/ED - sometimes the video goes back a 1/4 to 1/8 of a second. Then fast forwards to where it is supposed to be, and at that point gets stuck.

[JanWillem32]

I've re-written some memory management parts (which were hard to reconfigure, as the program shouldn't invoke the admin prompt of UAC). The biggest benefit was for memory allocation speed for the color management. (Which overall didn't really become much faster, but any optimization for it is a given.)
Various little bugs were squashed, but no real new features were added.
The biggest fix was for the internal MPEG2 decoder and its (badly implemented) subtitle part. I've added semi-native NV12 support for both (next to its native I420/IYUV fomat). MPEG2 videos and DVD menus should no longer be given as converted YUY2 or RGB32 to the mixer. This should solve all issues with wrong chroma coloring on DVD menus and the issue of reverting to the MS DVD video decoder. I don't have data on how my changes will affect the DXVA implementation for MPEG2, but it should be fine. I've also discarded the option for allowing planar formats. The NV12 output can handle interlaced video, but won't perform software deinterlacing (I'm really not going to add that feature). Enable "Set interlaced flag in output media type" for proper handling of interlaced NV12 output by the video renderer. If NV12 deinterlacing fails with any of the decent video renderers, please report it. For the x64 version, SSE2 math was updated (I couldn't get it into the x86 SSE2 build yet). If the x64 version specifically has visual artifact, memory or other problems, please report it. I've changed quite a lot of code to make this decoder work properly (for me, at least), some of the edited code may be less than perfect.
As a side-effect from fixing the bugs with switching D3D fullscreen mode, the worker thread feature for the filter graph has been disabled for the internal renderers. I'll analyze if it's worth it to fix (this feature was already disabled for when a video is started in D3D mode in the first place).
I've not found any bugs with the renderer's timing system yet. I'll have to do further analysis on that.

[CruNcher]

@Jan
This might be of help @ optimizing things http://software.intel.com/en-us/arti...ols-intel-gpa/ i find it really usefull some functions are comparable to Nvidias tools though it's more Generic and seems to be usable on non Intel GFX Hardware too it can show a lot of DirectX information as well including the Shader Execution status and stall situations
You can analyze Realtime via a HUD (though they keep sensors restricted into Groups to keep the overhead down) or via recording and analyzing after (lower overhead)
It doesn't seem to work though with testers Fullscreen http://img835.imageshack.us/img835/5...nalyzedxva.png
but analyzing it with the Intel GPA System Analyzer in the background works also Realtime and that is really fast with minimal overhead updating @ record speed http://img69.imageshack.us/img69/580...emanalyzer.png and more configurable also in sensor count on screen @ once see this Video Demo http://software.intel.com/en-us/vide.../1532811356001
Btw don't be surprised by the Playback Speed in the Screenshots (no MPC-HC Performance issue) the Video is VFR and the FPS is the actual 3D Engine Render speed @ recording time

Really nice http://img252.imageshack.us/img252/879/latencynice.png you can see in realtime here what caused the Power to go up in the Frame it is the Rendering of the Meteor drive by and it's Shader Effects

[fagoatse]

Thanks Jan, will give it a try today or tomorrow.
I can confirm that 4170i has improved things a bit, although the issues described by Hera are still present.

btw, I think you should stop offering SSE builds, every CPU out there supports SSE2 and above so it's no big deal. Legacy CPUs that can only handle SSE are too slow for playback these days anyway.


@Hera
Which of your setups are you testing on?
We're having similar issues, so I presume it's the ATI one with LAV filters and haali in x86 using SSE2 builds and EVR-CP+DXVA. Perhaps there's some connection between these components. Although it doesnt happen in potplayer so i guess it's either the renderer or pausing/seeking/playing functionality being broken to some extent?

[Hera]

Originally Posted by fagoatse

@Hera
Which of your setups are you testing on?
We're having similar issues, so I presume it's the ATI one with LAV filters and haali in x86 using SSE2 builds and EVR-CP+DXVA. Perhaps there's some connection between these components. Although it doesnt happen in potplayer so i guess it's either the renderer or pausing/seeking/playing functionality being broken to some extent?

Post #909 (the one before this one) was tested on my NVIDIA ION netbook (AMD 4250M rig also exhibits same symptoms, but I don't video much). I finished Clannad ~After Story~ (1080p BDRip w. Animated Subtitles) in two days on it. I get similar problems on both computers. On both computers I am using LAV for Video (Native DXVA enabled) and Audio. I have tried LAV splitter, but it doesn't support external OP/ED and thus I reverted back to Haali Splitter. I have not tried whether or not Haali Renderer works (w. LAV CUDA ofc).

I have yet to try 4210i - but judging from the description, I am not optimistic.

I have not tried the post-breakage MPC tester builds on the Athlon XP rig yet.

[fagoatse]

Originally Posted by Hera

Post #909 (the one before this one) was tested on my NVIDIA ION netbook (AMD 4250M rig also exhibits same symptoms, but I don't video much). I finished Clannad ~After Story~ (1080p BDRip w. Animated Subtitles) in two days on it. I get similar problems on both computers. On both computers I am using LAV for Video (Native DXVA enabled) and Audio. I have tried LAV splitter, but it doesn't support external OP/ED and thus I reverted back to Haali Splitter. I have not tried whether or not Haali Renderer works (w. LAV CUDA ofc).

I have yet to try 4210i - but judging from the description, I am not optimistic.

I have not tried the post-breakage MPC tester builds on the Athlon XP rig yet.

I tried 4210i and while performance is really impressive I can confirm two bugs you'd reported earlier, namely:

Another behavior I noticed - video stops playing, not during anything weird, just during normal playback.

Another behavior I noticed, after un-pausing, video plays, audio is no where to be found. Pausing and Un-pausing causes audio / video de-sync...

And of course the issue I mentioned about a week ago regarding video stall caused by heavy karaoke subs with animations(tested using Doki's release of toradora BD in 10bit). I'm pretty sure it's ISR's fault. In potplayer or regular mpchc when the CPU hits the wall and the video decoder cannot keep up it starts dropping frames, but in this case it just freezes, though audio is still being decoded. Well, basically any video decoding and rendering is dead and cpu sits at ~0%(just audio being processed but thats not very taxing). It seems to only happen on low-end CPUs(1.6 GHz dual core in this case). I've also noticed that after updating to 4120i and LAV filters to 0.50 the hang happens a few seconds later than it always used to(it's 100% reproducible at exactly the same moment).

[Hera]

Originally Posted by fagoatse

And of course the issue I mentioned about a week ago regarding video stall caused by heavy karaoke subs with animations(tested using Doki's release of toradora BD in 10bit). I'm pretty sure it's ISR's fault. In potplayer or regular mpchc when the CPU hits the wall and the video decoder cannot keep up it starts dropping frames, but in this case it just freezes, though audio is still being decoded. Well, basically any video decoding and rendering is dead and cpu sits at ~0%(just audio being processed but thats not very taxing). It seems to only happen on low-end CPUs(1.6 GHz dual core in this case). I've also noticed that after updating to 4120i and LAV filters to 0.50 the hang happens a few seconds later than it always used to(it's 100% reproducible at exactly the same moment).

I noticed video just getting stuck not during heavy subtitles - just normal speech subtitles (DXVA).

But, yet seeking getting stuck IS related to CPU power IMO: smaller video sizes do get stuck LESS when seeking. Some thread probably times out because it is not getting enough CPU time.

[fagoatse]

Originally Posted by Hera

I noticed video just getting stuck not during heavy subtitles - just normal speech subtitles (DXVA).

But, yet seeking getting stuck IS related to CPU power IMO: smaller video sizes do get stuck LESS when seeking. Some thread probably times out because it is not getting enough CPU time.

Yeah, although im not getting any hangs while playing 8bit stuff in dxva2 native. TBH HD6310 is okay with anything I throw at it, thanks to UVD3 I guess.
I've had quite a few random freezes in HW mode as well, but the hangs I've described above occur every time in SW mode when playing demanding 10bit content. It's probably as you're saying, not enough cpu time and inability to recover from it as it should - frame dropping. I can bear with occasional frame drops but not with being forced to restart playback.

[CruNcher]

I experience a reproducible crash it's not their with dfr4170i i'll look more into it later, it happens with specific files on EVR could be maybe the Resolution crash Problem regressed back in ?

[RGold]

Using latest build, player will freeze as follows:

1. Playback will end and and the player will exit full screen.
2. Drag and drop another video file into the player window.
3. Player will freeze with white window.

[JanWillem32]

@CruNcher: That's indeed a very nice HUD (and a few other tools), thanks for the links.
What exactly do you mean by "Resolution crash Problem"?

@fagoatse: Unfortunately, the trunk x86 build is set at an SSE level, so I can't drop testing and legacy support for those parts.

For the little discussion in between: it's indeed the subtitle renderer. When seeking, the subtitle renderer's queue is emptied. After that, the subtitle rendering thread starts to render new frames. In doing so, it blocks other parts of the player until the subtitle picture queue is full again. I've been re-writing the queue handler, with some success. It's just very hard to get it "just right" that it will work nicely for even low-power systems, while still being capable of rendering a subtitle frame for every video frame on faster systems. (I can post a build here, but I'm not 100% sure it's stable enough for general usage. It performs pretty well, though.)

@RGold: Is it with specific settings, or the defaults? I can't seem to replicate the issue. Do automatic transitions of the playlist work properly? Does the same thing happen in bordered windowed and fullscreen exclusive modes?

[Hera]

I also noticed the video get stuck after a resume from sleep.

[heavy voice]
JanWillem32 is credit to team!
[/heavy voice]

[RGold]

Originally Posted by JanWillem32

@RGold: Is it with specific settings, or the defaults? I can't seem to replicate the issue. Do automatic transitions of the playlist work properly? Does the same thing happen in bordered windowed and fullscreen exclusive modes?

1. For 24Hz video material, using autocahnge fullscreen monitor mode for 23.981 @ 1600X1200.
2. Default render settings with Aero is disabled.
3. It's only happening in bordered windowed mode.
4. Does not happen with the trunk version.

[burfadel]

Current build of this is 200 builds behind the mainstream MPC-HC...

Not that the current version doesn't work great, just keen for an update!