Attachments are not coming from the filter but from the pipeline and is
called by the filter implementation itself. Thus, this is defined by the
owner implementation and should not be in the operation callbacks.

It's particularily noticeable in the callsite given that the
get_attachments implementation is using owner.sys.

The metadata signalling the usage of a cubemap are defined in the
spatial-media repository[1] and enforce a specific layout for layout
identifier 0, which is the only one specified.

Implement this layout instead of the previous one, which was Facebook
one (using Mongolfiere 360 sample), especially given that Facebook has
mostly gone to equirectangular 360 videos since then.

It fixes the definition of each face so that they match their natual
axis with the following coordinate system (x being forward, and z being
a vector to the left, the whole in right-hand convention):

       y     x
          \ |
           \|
   z <-----(+)

It now uses the following picture layout:

1
 +-------+-------+------+
 | right | left  | up   |
 +-------+-------+------+      v
 | down  | front | back |      |
 +-------+-------+------+      +-- u
0                        1

For testing purpose, one can generate a sample for this format by using
ffmpeg, for example by converting an equirectangular video into cubemap
using the v360 filter.

   -lavfi v360=input=equirect:output=c3x2

Matroska projection can also be set, for example with:

   mkvpropedit sample --edit track:v1 --set projection-type=2

[1]: https://github.com/google/spatial-media/blob/master/docs/spherical-video-v2-rfc.md#cubemap-projection-box-cbmp

Alexandre Janniaux authored 4 years ago and Felix Paul Kühne committed 4 years ago

Meson before 0.54.2 misdetect the cross-compilation to iPhone Simulator
with architecture x86_64 and tries to execute tests, leading to
compilation failures in contrib.

Signed-off-by: Felix Paul Kühne <felix@feepk.net>

This is a manual forward-port of a8e6a381.

Unlike most UNIX calls [1] on Windows many calls expect a BOOL instead of an
int [2]. It should be the same size, but make sure this is the case when
building.

We don't seem to use any of the options with a DWORD, LINGER or Service
Provider values.

[1] https://www.man7.org/linux/man-pages/man7/socket.7.html
[2] https://docs.microsoft.com/en-us/windows/win32/api/winsock/nf-winsock-setsockopt

This is the first "official" release, see:
https://aomedia.googlesource.com/aom/+/HEAD/CHANGELOG

It's correct in there now.

It's correct in there now.

It contains a lot of things we don't need to build anymore. It's found in
recent msys2 or our Docker images.

Don't bypass steps and compile from the tarballs dir to the installation folder

The UNPACK_DIR matches the name of the tarball.

We require mingw 5.0 in configure.ac so there's no need to support contribs for
older versions.

The AWindow_ID is not used anymore since we moved the creation of
SurfaceTexture to the usage site. We also don't need to store the
SurfaceTexture in the AWindowHandler anymore.

This finally makes the SurfaceTexture independant of the AWindowHandler
and cleanup previous patches.

Signed-off-by: Alexandre Janniaux <ajanni@videolabs.io>

They are not used anymore since we create the SurfaceTexture ourselves.

Signed-off-by: Alexandre Janniaux <ajanni@videolabs.io>

Now that MediaCodec exposes a SurfaceTexture through the video context,
stop using AWindowHandler in interop and directly use the
vlc_asurfacetexture abstraction to bind the picture to the OpenGL
textures.

It means that, after dropping the hack in vlc_asurfacetexture_New,
multiple interop will finally work within the same AWindowHandler
instance and, for the future GPU Filter support, that filters can
provide their own SurfaceTexture.

Signed-off-by: Alexandre Janniaux <ajanni@videolabs.io>

Initially, the SurfaceTexture reference was created through
AWindowHandler and typically was obtained through a TextureView. By
creating the SurfaceTexture ourselves, we can avoid requesting Android
libVLC user to use libvlc_media_player_set_android_context even if the
user doesn't need any display.

In addition, it means that we don't rely on code that is in the JNI
android binding instead of vlc core when using MediaCodec, which helps
other bindings like C# one to provide support for Android.

Signed-off-by: Alexandre Janniaux <ajanni@videolabs.io>

Expose public functions to create and destroy the vlc_asurfacetexture
objects from an AWindowHandler object. It will be used in particular by
Mediacodec to output to OpenGL textures.

Signed-off-by: Alexandre Janniaux <ajanni@videolabs.io>

Add a JNI variant of SurfaceTexture, which supports any Android API
version starting from API 11.

It effectively removes the need for SurfaceTexture function wrapping in
the AWindow object.

Signed-off-by: Alexandre Janniaux <ajanni@videolabs.io>

The new function use a `struct vlc_asurfacetexture` object and is
written to be extended to the case in which we don't have the NDK
ASurfaceTexture API.

Signed-off-by: Alexandre Janniaux <ajanni@videolabs.io>

The API will be used when NDK API is not available.

Signed-off-by: Alexandre Janniaux <ajanni@videolabs.io>

When creating a SurfaceTexture JNI object, we can, with recent API,
transform the SurfaceTexture into an ASurfaceTexture NDK object.
However, with previous Android version, the only way is to create the
Surface JNI object directly from the SurfaceTexture JNI object, and then
convert the Surface into a ANativeWindow NDK object for the producer
side.

This patch loads the class and the constructor taking a reference to
SurfaceTexture.

Signed-off-by: Alexandre Janniaux <ajanni@videolabs.io>

init_z needs API 26 and is the most suitable since it allows creating
SurfaceTexture in a detached state, while init_iz needs API 19 and
cannot be created without an EGL context currently bounded.

init_i is the most compatible version (API 11) but won't allow
single-buffering.

Signed-off-by: Alexandre Janniaux <ajanni@videolabs.io>

Instead of passing the full path in the jfields structure, pass the
jfields first-order structure and the method field within it
separately, allowing to use it for future additional other classes too.

Signed-off-by: Alexandre Janniaux <ajanni@videolabs.io>

Refactor SurfaceTexture related API into a structure with vtable, so as
to implement both NDK and JNI API as separate vtable.

It will be exposed to modules and will allow creating SurfaceTexture
directly in the module needing them, instead of relying on an external
SurfaceTexture provided by the binding.

jsurface and ANativeWindow are also both exposed to the user since we
don't expose the ANativeWindow/JNI functions.

The object is made so that SurfaceTexture are always created in the
detached state. The current error path is also relatively defensive
in order to simplify the transition in the future commits, but is only
a temporary path.

Signed-off-by: Alexandre Janniaux <ajanni@videolabs.io>

Avoid forward-declaring SurfaceTexture implementation. In future
refactor we will change their prototype and put them into vtable.

Signed-off-by: Alexandre Janniaux <ajanni@videolabs.io>

Those helpers are used everywhere else in the file so having them at the
beginning is easier.

The cube coordinates were defined one by one, giving little possibility
for readable refactoring or little explanation for those not knowing how
the cube is built.

Instead, provide a helper macro generating the coordinates and other
helper macros in order to select the correct face to generate. It will
ease future refactoring, in particular to rotate the cube.

The cube coordinates were defined one by one, giving little possibility
for readable refactoring or little explanation for those not knowing how
the cube is built.

Instead, provide a helper macro generating the coordinates and other
helper macros in order to select the correct face to generate. It will
ease future refactoring, in particular to rotate the cube.

Otherwise codepoint == 0 always sets default font

Currently double free is only prevented by
FT_Set_Pixel_Sizes error on pure bitmaps.

FT_LOAD_NO_BITMAP does not exclude bitmap only fonts
and FT_Glyph_To_Bitmap is no-op for bitmap fonts,
shadow glyph pointer can then end pointing to same
glyph as main glyph.

spu_Attach could occur before lock