This helps for testing plane filtering.

If config.filter_planes is set on a filter, generate one output texture
for each input texture, and call draw() for each plane separately.

A "direct" sampler (i.e. a sampler without interop) only supported one
texture, because the output of the previous filter was assumed to be
RGBA in a single plane.

A deinterlace filter will output several planes, so the sampler of the
next filter must be able to receive all of them.

A sampler gives access to the input picture, in RGBA.

Add support for giving access to individual input planes, selected by
vlc_gl_sampler_SelectPlane(), in their native format.

The number of input planes and their sizes might be useful to filters
implementation.

It could be retrieved by vlc_fourcc_GetChromaDescription(), but for
convenience, expose them in sampler.

Compute the shader extensions in a separate function, to be able to
reuse it.

The type and function to use depends on the texture target.

Extract the mapping to a separate function to be able to reuse it.

Variables were generated on the fly with an index in the name. For
example:

    uniform sampler2D Texture0;
    uniform mat3 TexCoordsMap0;
    uniform sampler2D Texture1;
    uniform mat3 TexCoordsMap1;
    uniform sampler2D Texture2;
    uniform mat3 TexCoordsMap2;

Replace them by arrays:

    uniform sampler2D Textures[3];
    uniform mat3 TexCoordsMaps[3];

This will enable to expose a GLSL function providing a color for a
particular plane number passed as a parameter.

Now that both types of samplers (interop and direct) use the same
fragment shader code generation, factorize their creation.

A sampler may receive their input picture either:
 - from an interop, uploading a picture_t
 - from a previous filter, giving an OpenGL texture directly (a "direct"
   sampler)

The fragment shader for a "direct" sampler was straightforward: the
input was always an RGBA texture. On the contrary, samplers from interop
always applied chroma conversion to expose individual pixels in RGBA.

However, deinterlace filters operate on individual input planes, so
the chroma conversion must be applied afterwards.

To prepare support for this feature, use the same fragment shader code
generation for both types of samplers.

This will allow to generate chroma conversion code for "direct"
samplers, in order to convert chroma between OpenGL filters.

Swizzle will be necessary even for "direct" samplers (samplers without
interop).

In OpenGL, some values of picture planes scaling are hardcoded, and
might differ from the values provided by in a vlc_chroma_description_t.

Make the computation explicit and explain why.

Alexandre Janniaux authored 4 years ago and Romain Vimont committed 4 years ago

GBM platform allows to provide an EGL context even without display
server.

Co-authored-by: Romain Vimont <rom1v@videolabs.io>

A module egl_display must only provide EGL displays where eglTerminate()
can be called, even if other modules are using EGL (typically because it
is internally refcounted).

This is always the case on Android, so egl_display was initially only
enabled on Android:
https://android.googlesource.com/platform/frameworks/native/+/master/opengl/libs/EGL/egl_display.cpp

But this may also be the case on other platforms which support
EGL_KHR_display_reference:
https://www.khronos.org/registry/EGL/extensions/KHR/EGL_KHR_display_reference.txt



Co-authored-by: Alexandre Janniaux <ajanni@videolabs.io>

Alexandre Janniaux authored 4 years ago and Romain Vimont committed 4 years ago

If the EGL extension EGL_KHR_surfaceless_context is available, create a
surfaceless context.

Co-authored-by: Romain Vimont <rom1v@videolabs.io>

The filter egl_pbuffer was initially enabled only for Android, because
the default display is always internally refcounted.

Now that the display is managed separately, egl_pbuffer may be enabled
as soon as OpenGL and EGL are available.

Add a default EGL display module which just calls
eglGetDisplay(EGL_DEFAULT_DISPLAY).

Add a new module type "egl display" to open an EGL display.

The filter egl_pbuffer does not depend on LIBM.

The macro was checking for $with_foo to be set but only $enabled_foo was
set correctly by the PKG_WITH_MODULES macro. In addition, this patch
adds an intermediate macro for the name to be readable.

Romain Vimont authored 4 years ago and Alexandre Janniaux committed 4 years ago

If npot (non-power-of-two) textures are not supported, then the picture
visible dimensions is not the same as the texture dimensions.

When updating a picture, the interop (which calls glTexSubImage2D() or
similar) expects picture dimensions.

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

Romain Vimont authored 4 years ago and Alexandre Janniaux committed 4 years ago

Contrary to vlc_gl_Create(), vlc_gl_CreateOffscreen() did not initialize
gl->api_type, so it was always implicitly set to VLC_OPENGL, never
VLC_OPENGL_ES2.

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

Avoid flushing mediacodec when the decoder has been drained at the
VLC decoder_t level, because the picture output by the decoder are
inflight pictures that will not stay valid when the decoder will be
flushed. Instead, only flush it when every pictures have been released.

This is too pessimistic currently, the best alternative being doing the
flush when every picture have been rendered.

Avoid trying to queue picture after the closing of the decoder. Indeed,
the mediacodec loop cannot be killed yet at the decoder close function
because it is still needed for the integrity of the picture still used
in the pipeline.

These functions will be used with the decoder_t object closed, but the
p_sys still alive, and they are only using the p_sys object.

No functional changes.

Format change needs a restart, but when demuxer has been probed, we
only need to handle restart/reprobing when switching variants.