ffdshow tryouts project: Discussion & Development

[tal.aloni]

Originally Posted by hoborg

Looks like there are some problems with DXVA in latest FFDShow x86 (tested on 3292 and 3297) - every 3-4s there is glitch like bad frame order, and video will freeze after few of such glitches.

fixed in rev. 3302

[albain]

Finally I have managed to import the new libswscale with mutlithreading patch and did the updates on all the ffdshow filters (resize, sharp, subtitles, colorspace converter...)

I am testing it right now but it seems (surprisely) to be going well.

Maybe it will fix the crashes with bluray subs

[Sebastiii]

Great news and work as usual
Thx,
Seb.

[STaRGaZeR]

Originally Posted by Sebastiii

Great news and work as usual
Thx,
Seb.

Indeed!

Guys, is there any way of speeding up ICL10 compiling? It's only using 1 core here and takes forever to build.

[JarrettH]

What should I do to confirm dxva is working in ffdshow? I've ticked off h264 in the dxva dialogue so far and am using ffmpeg-mt for the h264 decoder. Does a new tray icon for dxva video appear? Does it only work with EVR?

Running Windows 7, 7600 GT, ffdshow rev 3305, 196.21 drivers

[hoborg]

Originally Posted by tal.aloni

hoborg, the issue is introduced with rev. 3288,
I'm looking into it.

No problem in 3304 any longer

[albain]

Hi,

here is a test build with new libswscale and libpostproc libraries.

There are many tests to do to :
Image filters updated :
- Blur
- Deinterlace (using ffmpeg)
- Postprocessing
- Luma
- Resize
- Sharpen
- Warsharp
- Subtitles

With different colorspaces

Bluray and DVD subtitles to be tested too

Some MMX2 instructions and general optimizations have been added since the versions we kept, also we still have the multithreading patch from Haruhiko.

Libmplayer is not used anymore and its interest becomes rather limited : audio channels reordering and yadif deinterlacer
Both could be imported into FFDShow

Here is the link

Source patch (many files!)

[ikarad]

Originally Posted by albain

Hi,

here is a test build with new libswscale and libpostproc libraries.

There are many tests to do to :
Image filters updated :
- Blur
- Deinterlace (using ffmpeg)
- Postprocessing
- Luma
- Resize
- Sharpen
- Warsharp
- Subtitles

With different colorspaces

Bluray and DVD subtitles to be tested too

Some MMX2 instructions and general optimizations have been added since the versions we kept, also we still have the multithreading patch from Haruhiko.

Libmplayer is not used anymore and its interest becomes rather limited : audio channels reordering and yadif deinterlacer
Both could be imported into FFDShow

Here is the link

Source patch (many files!)

same crash with bluray subs with rgb32 or yuy 2 Even if I don't use subs from ffdshow and that I use sub renderer from mpc-hc, there is crash when I select subs. This problem only occurs since your implementation of blu rays subs support (since rev 3277).


I don't understand why. Maybe I don't select good options.
I have turn off auto load subtitle in mpc-hc and I have selected bluray subs in ffdshow. And when movie start I select subtitles in mpc-hc (navigate, subtitle langage and then select the subs)
Could you explain me how use blu ray subs from ffdshow?

[albain]

Originally Posted by ikarad

same crash with bluray subs with rgb32 or yuy 2 Even if I don't use subs from ffdshow and that I use sub renderer from mpc-hc, there is crash when I select subs. This problem only occurs since your implementation of blu rays subs support (since rev 3277).


I don't understand why. Maybe I don't select good options.
I have turn off auto load subtitle in mpc-hc and I have selected bluray subs in ffdshow. And when movie start I select subtitles in mpc-hc (navigate, subtitle langage and then select the subs)
Could you explain me how use blu ray subs from ffdshow?

I have reproduced the crash with bluray subs

[tal.aloni]

Originally Posted by albain

Libmplayer is not used anymore and its interest becomes rather limited : audio channels reordering and yadif deinterlacer
Both could be imported into FFDShow

great, we finally put things into order.
your work is much appreciated, thanks!

[ikarad]

Originally Posted by albain

I have reproduced the crash with bluray subs

Do you know where is the problem ?

[XhmikosR]

Originally Posted by albain

Hi,

here is a test build with new libswscale and libpostproc libraries.

There are many tests to do to :
Image filters updated :
- Blur
- Deinterlace (using ffmpeg)
- Postprocessing
- Luma
- Resize
- Sharpen
- Warsharp
- Subtitles

With different colorspaces

Bluray and DVD subtitles to be tested too

Some MMX2 instructions and general optimizations have been added since the versions we kept, also we still have the multithreading patch from Haruhiko.

Libmplayer is not used anymore and its interest becomes rather limited : audio channels reordering and yadif deinterlacer
Both could be imported into FFDShow

Here is the link

Source patch (many files!)

Thanks for all your hard work. Just letting you know in case you don't, that with this patch ffdshow x64 compilation fails.

http://pastie.org/private/c21r1tymvvdfmmsmdfceg

[albain]

Originally Posted by ikarad

Do you know where is the problem ?

This is a stupid mistake : at the beginning I was using ffmpeg as the PGS parser but it is far from working so I made one from exisiting MPC's parser but I forgot to remove some code.

This problem occurred randomly (variable unitialized) which explained why it wasn't easily reproductible

Anyway, fixed in revision 3306.

Here is a test build with the ffmpeg new libraries inside

Also I have to fix the problem with the edges

@XhmikosR
I know, I am working on it, the problem comes from ffdshow, not ffmpeg which was the tough part
It is just a matter of types.

@clsid, Tal & specialists
Does someone know why we use stride_t type which maps to "__w64 int" in 32 bits and "__int64" in 64 bits ?
Why would we need to have so large variables to store the strides which rarely exceeds a few thousands ?

Compilation errors are due to this problem of mapping
I wonder if I have to
1/remap all the int types within ffmpeg to stride_t or
2/change stride_t inside ffdshow to int

I think this is approach #1 that I should use... or #2

[clsid]

I would prefer if the ffmpeg code would stay as close to the original as possible. That makes it much easier to update it on a regular basis.

[onomatopellan]

Originally Posted by albain

Anyway, fixed in revision 3306.

Here is a test build with the ffmpeg new libraries inside

Great! There is no crash anymore.
There is a problem with colored PGS though. In this image top is ffdshow rev3306 and bottom is MPC-HC subs (tried with rgb32 and nv12 outputs):


m2ts with PGS sub are made with MultiAVCHD and easySUP and it shows ok in a sony bluray player. Here is the sample:
-> SAMPLE <-

[STaRGaZeR]

Originally Posted by albain

Here is a test build with the ffmpeg new libraries inside

With YV12 video, RGB32 HQ output:

- Resize crashes the player inmediately with errors in libavcodec.dll
- ffmpeg deinterlacers like 5-tap lowpass, same error as above.
- Postprocessing, same as above.
- Sharpen, everything works OK except swscaler that produces a black screen, also msharpen doesn't do anything it seems.
- Warpsharp works OK.
- In Blur, libavcodec error again when using mplayer temporal noise reducer, black screen with swscaler gaussian blur.
- Levels, libavcodec crash.
- Luminance level fix in Picture properties, libavcodec crash.

[albain]

Originally Posted by clsid

I would prefer if the ffmpeg code would stay as close to the original as possible. That makes it much easier to update it on a regular basis.

This is the approach that I had. The ffmpeg code for those 2 libs is very close to the original code.
As for the pixel formats (moved from ffimgmt.h to where they belong)

However for these strides it is a little bit tricky because they can't be casted on the fly so one would have to copy them from the stride_t array to an int array before calling ffmpeg methods.

So for now I have changed the ffmpeg signatures from int to stride_t.
Maybe someone will find a better solution

Updated news : now the compilation works fine on 64 bits.
I'll post a 64 bits soon with the new patch

@Stargazer : thank you for testing. I can't reproduce the crashes however
What I have done (with graphstudio) :
MPC m2ts splitter -> MPC video decoder -> (yv12) -> ffdshow video -> (rgb32) -> EVR




There is some work to do indeed. I guess those crashes are all related to the same problem

[albain]

Ok, I reproduce the crashes but with MPC, not with graphstudio.

This is annoying because MPC doesn't let attach the debugger and will exit instead.

Otherwise, does someone know how to debug ffmpeg ?

I can debug it when building with MSVC but all the optimizations are disabled.

I tried with eclipse + gdb but the symbols are not loaded. I guess that an option is missing

@Ikarad : I have fixed that color problem. R and B colors were reversed as well as cut edges (fixed in revision 3307)

[onomatopellan]

Thanks albain, using rev 3307 now rgb32 output is ok but it crash again with nv12(dxva) output (using MPC-HC).

[ikarad]

Originally Posted by albain

Ok, I reproduce the crashes but with MPC, not with graphstudio.

This is annoying because MPC doesn't let attach the debugger and will exit instead.

Otherwise, does someone know how to debug ffmpeg ?

I can debug it when building with MSVC but all the optimizations are disabled.

I tried with eclipse + gdb but the symbols are not loaded. I guess that an option is missing

@Ikarad : I have fixed that color problem. R and B colors were reversed as well as cut edges (fixed in revision 3307)

Thanks there is no crash with YUY 2
Where can I download rev 3307?