Each subpicture region is merged into one for libva 0.31.0 and applied
on top of the display surface by the GPU.
.
Higher versions of the libva library can handle multiple subpicture
region associations on a surface with vaAssociateSubpicture2() function.

- subtitles are readable
- some are clipped, this needs fixing
- crash in vlc_memcpy() (line: 545 of vaapi_x11.c file)
- fix positioning of subtitles (region->i_align)
- alpha is applied when available
- chroma is applied when available

Text subtitles are not impacted by this commit.

In this case, the subpictures are rendered either at the display size or
the video size (the biggest one).

In case vaapi-x11 video output is used, then the default amount of surfaces is
not enough to guarantee smooth playback. Therefor double the amount of requested
surfaces. Currently there is no easy way of determining if vaapi-x11 video output
is going to be used, because the vaapi-codec is most likely to be constructed first.

Put shared connection to libva into a separate headerfile, so it
can be shared between vaapi.c and vaapi_x11.c. This cleansup headerfile
includes in dxva2.c and video.c, which have no interest in private
communication between vaapi.c and vaapi_x11.c.

When -V vaapi-x11 is used (and ffmpeg-vaapi-x11 option), then pass
VASurface information in picture->p_sys. This only works with direct
rendering and video filters that directly operate on the picture are
a no go.

* vaapi.c: protect surface->i_refcount
* vaapi_x11.c: use direct rendering

vlc_va_Initialize() and vlc_va_Terminate() protects vaInitialize/vaTerminate
from being called more then once. Undefining VLC_VA_SHARED_CONN forces a
private connection per each instance that calls vlc_va_Initiailize/vlc_va_Terminate.

Using memcmp results in false positives (src format and dst format is
exactly the same, but still the logic resulted in loading swscale).

The macro DefaultVisual only returns an reference and not an allocation.

Do not free pic->p_sys twice. The function picture_Delete() already
does freeing of pic->p_sys.

Use VAImageFormat information from VAAPI to fill vd->fmt.

The pictures used by vaapi-x11 vout are not valid vlc pictures, but merely
a placeholder for VASurfaceID.

Return a pointer to vlc_va_conn_t.

Use --ffmpeg-vaapi-x11 in combination with --ffmpeg-hw. It uses
the vaapi to decode pictures and display it without copying it back
to main memory.
When --no-ffmpeg-vaapi-x11 is used then the picture is copied from
the GPU to main memory and displayed later by another vout. This uses
more CPU then when the picture is displayed from the GPU contenxt directly.

Implement picture pool for VAAPI-X11 vout.

Use vlc_va_* wrapper to call vaInitialize and vaTerminate

Initialize/terminate libva only once. VAAPI calls are now being used by
2 modules which only have interaction with each other through vlc core.
Both modules (codec and vout) can be initiated independently they should
not call vaIntialize (or vaTerminate) one after each other (or simultanously).

Add option to select stream_output module to enable an es by default.

Video output VAAPI-X11 (-V vaapi-x11) depends on --ffmpeg-hw to be enabled.

Signed-off-by: Jean-Baptiste Kempf <jb@videolan.org>

This fixes the crash when Media Library tries to add "My Videos" / "My
Pictures" / etc. playlist entries to a ML when browsing them through "My
Computer" service discovery.

Signed-off-by: Laurent Aimar <fenrir@videolan.org>