Allow plain gst_buffer_map() interface to work with gstreamer-vaapi
video buffers, i.e. expose the underlying GstVaapiSurfaceProxy to the
caller. This is the only sensible enough thing to do in this mode as
the underlying surface pixels need to be extracted through an explicit
call to the gst_video_frame_map() function instead.
A possible use-case of this is to implement a "handoff" signal handler
to fakesink or identity element for further processing.
Fix gst_vaapi_video_allocator_new() to silently check for direct-rendering
mode support, and not trigger fatal-criticals if either test surface or
image could not be created. Typical case: pixel format mismatch, e.g. NV12
supported by most hardware vs. I420 supported by most software decoders.
On map, ensure we have GST_MAP_WRITE flags since this is only what we
support for now. Likewise, on unmap, make sure that the VA image is
unmapped for either read or write, while still committing it to the
VA surface if write was requested.
Break the circular references between GstVaapiContext and its children
GstVaapiSurfaces. Since the VA surfaces held an extra reference to the
context, which holds a reference to its VA surfaces, then none of those
were released.
How does this impact support for subpictures?
The only situation when the parent context needs to disappear is when
it is replaced with another one because of a resolution change in the
video stream for instance, or a normal destroy. In this case, it does
not really matter to apply subpictures to the peer surfaces since they
are either gone, or those that are left in the pipe can probably bear
a reinstantiation of the subpictures for it.
So, parent_context is set to NULL when the parent context is destroyed,
other VA surfaces can still get subpictures attached to them, individually
not as a whole. i.e. subpictures for surface S1 will be created from
active composition buffers and associated to S1, subpictures for S2 will
be created from the next active composition buffers, etc. We don't try
to cache the subpictures in those cases (pending surfaces until EOS
is reached, or pending surfaces until new surfaces matching new VA context
get to be used instead).
Make sure to create the GLX context once the window object has completed
its creation. Since gl_resize() relies on the newly created window size,
then we cannot simply overload the GstVaapiWindowClass::create() hook.
So, we just call into gst_vaapi_window_glx_ensure_context() once the
window object is created in the gst_vaapi_window_glx_new*() functions.
Make it possible to add extra an extra filter to most of display cache
lookup functions so that the GstVaapiDisplay instance can really match
a compatible and existing display by type, instead of relying on extra
string tags (e.g. "X11:" prefix, etc.).
Improve check for display cache infrastructure. In particular, for X11 and
GLX backends, we need to make sure that we can create a GstVaapiDisplayX11
from another GstVaapiDisplayGLX, i.e. underlying X11 and VA displays can be
shared. Besides, allocating a GstVaapiDisplayGLX while a GstVaapiDisplayX11
already exists will have to generate different VA displays.
In GStreamer 0.10 builds, gst_vaapi_uploader_get_buffer() was used
but it exhibited a memory leak because the surface generated for the
GstVaapiVideoMeta totally lost its parent video pool. So, it was not
possible to release that surface back to the parent pool when the meta
gets released, and the memory consumption kept growing.
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
Since GST_VAAPI_IS_xxx_VIDEO_POOL() was only testing for NULL and not
the underlying object type, the gst_vaapi_video_meta_new_from_pool()
was hereby totally broken. Fixed this regression by using the newly
provided gst_vaapi_video_pool_get_object_type() function.
Drop obsolete GST_VAAPI_IS_xxx() helper macros since we are no longer
deriving from GObject and so those were only checking for whether the
argument was NULL or not. This is now irrelevant, and even confusing
to some extent, because we no longer have type checking.
Note: this incurs more type checking (review) but the libgstvaapi is
rather small, so this is manageable.
The whole libgstvaapi libraries got a major refresh to get rid of GObject.
This is a fundamental change that requires a new SONAME. More changes are
underway to streamline the core libraries.
So far, the net result is a reduction of .text size (code) by 32KB, i.e. -10%.
On one particular test (sintel HD trailer), the total number of executed
instruction was reduced by 8%.
Make GstVaapiID a gsize instead of guessing an underlying integer large
enough to hold all bits of a pointer. Also drop GST_VAAPI_ID_NONE since
this is plain zero and that it is no longer passed as varargs.
Port GstVaapiDecoder and GstVaapiDecoder{MPEG2,MPEG4,JPEG,H264,VC1} to
GstVaapiMiniObject. Add gst_vaapi_decoder_set_codec_state_changed_func()
helper function to let the user add a callback to a function triggered
whenever the codec state (e.g. caps) changes.
Port GstVaapiVideoPool, GstVaapiSurfacePool and GstVaapiImagePool to
GstVaapiMiniObject. Drop gst_vaapi_video_pool_get_caps() since it was
no longer used for a long time. Make object allocators static, i.e.
local to the shared library.
Drop support for user-defined data since this capability was not used
so far and GstVaapiMiniObject represents the smallest reference counted
object type. Add missing GST_VAAPI_MINI_OBJECT_CLASS() helper macro.
Besides, since GstVaapiMiniObject is a libgstvaapi internal object, it
is also possible to further simplify the layout of the object. i.e. merge
GstVaapiMiniObjectBase into GstVaapiMiniObject.
Propagate the picture size from the bitstream to the GstVaapiDecoder,
and subsequent user who installed a signal on notify::caps. This fixes
decoding of TS streams when the demuxer failed to extract the required
information.
Add gst_vaapi_decoder_get_frame_with_timeout() helper function that will
wait for a frame to be decoded, until the specified timeout in microseconds,
prior to returning to the caller.
This is a fix to performance regression from 851cc0, whereby the vaapidecode
loop executed on the srcpad task was called to often, thus starving all CPU
resources.
Rework heuristics to detect when downstream element ran into errors,
and thus failing to release any VA surface in due time for the current
frame to get decoded. In particular, recalibrate the render time base
when the first frame gets submitted downstream, or when there is no
timestamp that could be inferred.
Rework GstVideoDecoder::handle_frame() to decode the current frame,
while possibly waiting for a free surface, and separately submit all
decoded frames from a task. This makes it possible to pop and render
decoded frames as soon as possible.
Fix reference counting bug for passthrough mode, whereby the input buffer
was propagated as is downstream through gst_pad_push() without increasing
its reference count before. The was a problem when gst_pad_push() returns
an error and we further decrease the reference count of the input buffer.
