There's no need for the jump to an extra thread in most cases, especially
when relying solely on a shader to render. We can use the provided
render_to_target() functions to simplify filter writing.
Facilities are given to create fbo's and attach GL memory (renderbuffers
or textures). It also keeps track of the renderable size for use with
effective use with glViewport().
Make state changes of internal elements more reliable by locking
their state, and ensuring that they aren't blocked pushing data
downstream before trying to set their state.
Add a boolean to avoid starting tasks when the main
thread is busy trying to shut the element down.
Try harder to make switching pads work better by
making sure concurrent downloads are finished before exposing
a new set of pads.
Release the manifest lock when signalling no-more-pads, as
that can call back into adaptivedemux again
If other stream fragments are still downloading but new streams
have been scheduled, don't expose them yet - wait until the last
one finishes. Otherwise, we can cancel a partially downloaded
auxilliary stream and cause a gap.
Drop the manifest lock when performing actions that might
call back into adaptivedemux and trigger deadlocks, such
as adding/removing pads or sending in-band events (EOS).
Unlock the manifest lock when changing the child bin state to
NULL, as it might call back to acquire the manifest lock when
shutting down pads.
Drop the manifest lock while pushing events.
In the case of KEY_UNIT and TRICKMODE_KEY_UNITS seeks, we want to
"snap" to the closest fragment.
Without this, we end up pushing out a segment which does not match
the first fragment timestamp being pushed out, resulting in one or
more buffers being eventually dropped because they are out of segment.
Calling glUniformMatrix before the shader is bound is invalid and
would result in errors like:
GL_INVALID_OPERATION in glUniformMatrix(program not linked)
Move glUniformMatrix() to after the gst_gl_shader_use() call.
Rather than assuming something. e.g. zerocopy on iOS with GLES3 requires
the use of Luminance/Luminance Alpha formats and does not work with
Red/RG textures.
Take the used texture type from the memory instead.
Fixes conversion from multi-planar YUV formats with two components per plane
(NV12, NV21, YUY2, UYVY, GRAY16_*, etc) with Luminance Alpha input textures.
This is also needed for zerocopy decoding on iOS with GLES 3.x.
The intention was to assert if both maj and min were NULL (as there would be no
point calling the function). Instead if either maj or min were NULL, the assert
would occur.
Fix that.
Newer devices require using a different GLSL extension for accessing
external-oes textures in a shader using the texture() functions.
While the GL_OES_EGL_image_external_essl3 should supposedly be supported
on a any GLES3 android device, the extension was defined after a lot of the
older drivers were built so they will not know about it. Thus there are two
possible interpretations of which of texture[2D]() should be supported for
external-oes textures. Strict adherence to the GL_OES_EGL_image_external
extension spec which uses texture2D() or following GLES3's pattern, also
allowing texture() as a function for accessing external-oes textures
This adds another mangling pass to convert
#extension GL_OES_EGL_image_external : ...
into
#extension GL_OES_EGL_image_external_essl3 : ...
on GLES3 and when the GL_OES_EGL_image_external_essl3 extension is supported.
Only uses texture() when the GLES3 and the GL_OES_EGL_image_external_essl3
extension is supported for external-oes textures.
Uses GLES2 + texture2D() + GL_OES_EGL_image_external in all other external-oes
cases.
https://bugzilla.gnome.org/show_bug.cgi?id=766993
Otherwise we will leak GstGLContext's when adding the same context more than
once.
Fixes a regression caused by 5f9d10f603 in the
gstglcontext unit test that failed with:
Assertion 'tmp == NULL' failed
Until now we would start the task when the pad is activated. Part of the
activiation concist of testing if the pipeline is live or not.
Unfortunatly, this is often too soon, as it's likely that the pad get
activated before it is fully linked in dynamic pipeline.
Instead, start the task when the first serialized event arrive. This is
a safe moment as we know that the upstream chain is complete and just
like the pad activation, the pads are locked, hence cannot change.
https://bugzilla.gnome.org/show_bug.cgi?id=757548
If the input buffer is after the end of the output buffer, then waiting
for more data won't help. We will never get an input buffer for this point.
This fixes compositing of streams from rtspsrc.
https://bugzilla.gnome.org/show_bug.cgi?id=766422
Provide a function to get the affine matrix in the meta in terms of NDC
coordinates and use as a standard opengl matrix.
Also advertise support for the affine transformation meta in the allocation
query.
The URI downloader is creating the source element with
gst_element_factory_make() that returns a floating reference that nobody
is consuming. This is causing problems in WebKit, where the smart
pointers used to take references of the source elment get confused and
end up consuming the floating reference and then releasing the element,
which usually crashes because the URI downloader still tries to use its
src element. See https://bugs.webkit.org/show_bug.cgi?id=144040.
This commit adds two helper functions to ensure and destroy the source
element, to make the code simpler and less error prone. The ensure
method takes care of checking if we can reuse the existing one or we
need to create a new one, taking always its ownership. The destroy
method simply avoids duplicated code to set the source to NULL state and
then unref it.
https://bugzilla.gnome.org/show_bug.cgi?id=766053
The gst_adaptive_demux_wait_until() function can be woken up either
by its end_time being reached, or from other threads that want to
interrupt the waiting thread.
If the thread is interrupted, it needs to cancel its async clock callback
by unscheduling the clock callback. However, the callback task might already
have been activated, but is waiting for the mutex to become available. In this
case, the call to unschedule does not stop the callback from executing.
The solution to this second issue is to use a reference counted object that
is decremented by both the gst_adaptive_demux_wait_until() function and the
call to gst_clock_id_wait_async (). In this way, the GstAdaptiveDemuxTimer
object is only deleted when both the gst_adaptive_demux_wait_until() function
and the async callback are finished with the object.
https://bugzilla.gnome.org/show_bug.cgi?id=765728
Because current GstEGLImageMemory does not inherit GstGLMemory, GLUpload
allocates additional GLMemory and upload the decoded contents from the decoder
which uses EGLImage (e.g. gst-omx in RPi).
This work adds GstGLMemoryEGL to avoid this overhead. Decoders allocate
GstGLMemoryEGL and decode its contents to the EGLImage of GstGLMemoryEGL. And
GLUpload uses this memory without allocation of additional textures and blit
operations.
[Matthew Waters]: gst-indent the sources and fix a critical retreiving the egl
display from the memory.
https://bugzilla.gnome.org/show_bug.cgi?id=760916
Allows creating wrapped memories with GstGLAllocationParams.
The wrapped pointers will be set in the parameters before being passed
to the memory allocation function.
Some platforms provide an old version of GLES2/gl2.h and GLES2/gl2ext.h that
will fail when including GLES3/gl3.h due to missing typedef's.
Seen on the RPi.
Adds a new function to mpegts lib to create a iso639 language
descriptor from a language and use it in mpegtsmux to add
a language descriptor to audio streams that have a language set.
https://bugzilla.gnome.org/show_bug.cgi?id=763647
This fixes a race where we check if there is a clock, then it get
removed and we endup calling gst_clock_new_single_shot_id() with a NULL
pointer instead of a valid clock and also calling gst_object_unref()
with a NULL pointer later.
https://bugzilla.gnome.org/show_bug.cgi?id=757548
There are several places in adaptivedemux where it waits for
time to pass, for example to wait until it should next download
a fragment. The problem with this approach is that it means that
unit tests are forced to execute in realtime.
This commit replaces the use of g_cond_wait_until() with single
shot GstClockID that signals the condition variable. Under normal
usage, this behaves exactly as before. A unit test can replace the
system clock with a GstTestClock, allowing the test to control the
timing in adaptivedemux.
https://bugzilla.gnome.org/show_bug.cgi?id=762147
A realtime clock is used in many places, such as deciding which
fragment to select at start up and deciding how long to sleep
before a fragment becomes available. For example dashdemux needs
sample the client's estimate of UTC when selecting where to start
in a live DASH stream.
The problem with dashdemux calculating the client's idea of UTC is
that it makes it difficult to create unit tests, because the passage
of time is a factor in the test.
This commit changes dashdemux and adaptivedemux to use the
GstSystemClock, so that a unit test can replace the system clock when
it needs to be able to control the clock.
This commit makes no change to the behaviour under normal usage, as
GstSystemClock is based upon the system time.
https://bugzilla.gnome.org/show_bug.cgi?id=762147
Don't create many short lived locks/conds in gst_gl_window_send_message. This is
a micro optimization to save a bunch of pthread_* calls which are expensive on
OSX/iOS and possibly other platforms.
We weren't using the result of find_best_format at all.
Also, move the find_best_format usage to the default update_caps() to make
sure that it is also overridable.
https://bugzilla.gnome.org/show_bug.cgi?id=764363
The subsampling_x, subsampling_y, bit_depth, color_space and color_range
fileds are moved from GstVp9FrameHdr to the global GstVp9Parser structure.
These fields are only present in keyframe or intra-only frame, no need to
duplicate them for inter-frames. This is an ABI change.
https://bugzilla.gnome.org/show_bug.cgi?id=764370
while calling eglCreateImage without a GL context current in the executing
thread works on the RPi, some other implementations will return errors.
Marshall the eglCreateImage to the GL thread to appease these implementations.
There are numerous slight differences required between Desktop GL and GLES3 for
multiple render targets.
1. gl_FragData doesn't exist at all and one is required to use
'layout (location = ?) out ...' instead.
2. gl_FragColor doesn't exist, same as 1
3. texture2D() has been deprecated
Fortunately most of these have been taken care of with GL3 and the shader
mangling already exists so just expand the conditions they are used in. The
gl_FragData issue requires a new mangle pass though. We also use this new
pass on desktop GL for consistency.
By default, reading GL_RED or GL_RG us unsupported by glReadPixels unless
exposed through GL_COLOR_READ_IMPLEMENTATION_FORMAT/TYPE. This allows
downloading multiple-planar video frames where possible.
There are some cases where it's needed for binding in/out variables in shaders.
e.g. glsl 150 (gl 3.2) doesn't support the 'layout (location = ?)' specifiers in
the shader source so we have to bind them ourselves.
Happens e.g. if a RECONFIGURE event is sent from downstream while we're
switching pads at this very moment. The old pad is gone and the stream has a
new pad.
https://bugzilla.gnome.org/show_bug.cgi?id=764404
Previously, while allocating the pad number for a new pad, aggregator was
maintaining an interesting relationship between the pad count and the pad
number.
If you requested a sink pad called "sink_6", padcount (which is badly named and
actually means number-of-pads-minus-one) would be set to 6. Which means that if
you then requested a sink pad called "sink_0", it would be assigned the name
"sink_6" again, which fails the non-uniqueness test inside gstelement.c.
This can be fixed by instead setting padcount to be 7 in that case, but this
breaks manual management of pad names by the application since it then becomes
impossible to request a pad called "sink_2". Instead, we fix this by always
directly using the requested name as the sink pad name. Uniqueness of the pad
name is tested separately inside gstreamer core. If no name is requested, we use
the next available pad number.
Note that this is important since the sinkpad numbering in aggregator is not
meaningless. Videoaggregator uses it to decide the Z-order of video frames.
Without the GST_GL_API_GLES2 bit set, we will not even attempt to look
for the function pointers in the core library and will fallback to
glFlush/glFinish.
If the user uploads their own texture without setting the unpack length, then
then the result will have the appearance of stride mismanagement due to
an incorrect row length.
If we are given caps with extra features (like the overlay composition
features), we can only deal with that when we are in passthrough mode.
Previously we were bailing entirely and not allowing passthrough filter elements
with things like textoverlay.
Fixes the following pipeline (assuming glfilter supports passthrough):
gl ! textoverlay ! glfilter ! ... ! glimagesinkelement
https://bugzilla.gnome.org/show_bug.cgi?id=763756
When transforming, xplode it out into the necessary caps features both
with and without the passthough features.
Fixes negotiation in the following class of pipelines:
gl ! textoverlay ! glupload ! glimagesinkelement
https://bugzilla.gnome.org/show_bug.cgi?id=763756
GL 1.4 (with GL_ARB_shader_objects) doesn't have glIsProgram or glIsShader
equivalents. As they are simply assertions, skip them when there isn't a
valid function pointer.
This is a regression from since mpegvideoparser was switched to
use the codecparsing library.
The problem is that the high bit of the profile_and_level is used
to specify non-hierarchical profiles and levels. Unfortunately we
were discarding that information.
Expose that escape bit, and use it in the element
https://bugzilla.gnome.org/show_bug.cgi?id=763220
With e38af23044 returning the correct contexts,
gst_gl_display_add_context() was susceptible to causing infinte loops when
adding the same GstGLContext more than once. Fix and add a test for
gst_gl_display_add_context().
Fixes glvideomixer gst-validate tests.
Execute GL calls without marshalling them to the context thread. In the cocoa
and eagl backends calling gst_gl_context_activate is cheap and therefore calling
it on the current thread and serializing GL calls with a per-context lock is
more efficient (faster and has less overhead) than marshalling everything to the
context thread.
This optimization cuts a large overhead in g_poll (continuously waking up the
context thread) and in g_mutex_*/g_cond_* (waiting for results from the context
thread).
When requesting a glcontext (regardless of thread), the result was correct.
However, when requesting current glcontext on a specific thread, it could
come up with a glcontext active on another thread.
https://bugzilla.gnome.org/show_bug.cgi?id=763168
Maxsize is initialized once and should never change. Allocating data
should have no impact on the selected max size for this memory. This
causing memory map failure as the maxsize would become smaller then
size. This happened when using direct rendering in avviddec on GL that
does not support PBO transfer.
https://bugzilla.gnome.org/show_bug.cgi?id=763045
When the start_type is GST_SEEK_TYPE_NONE for a forward seek
(or stop_type for a reverse) is not set on a snap seeking operation,
the element should use the current position and then snap as requested.
Also fixes uninitialized variable complaint by clang about
'ts' variable.
Intended for use with wrapped contexts that are created shared with gst's
gl contexts in order to manage the internal sharegroup state correctly.
e.g. with caopengllayer (which is used in glimagesink and caopengllayersink
on OS X), we create a CGL context from the gst context and the sharing state
was not being correctly set on either GL context and gst_gl_context_is_shared()
was always returning FALSE.
With 11fb4fff80 only flushing with multiple
shared contexts, the required flush was not occuring causing screen
corruption or stuttering.
Note: this didn't affect GST_GL_API=opengl pipelines
https://bugzilla.gnome.org/show_bug.cgi?id=762620
When caps are already negotiated it should be possible to
select formats other than the one that was negotiated. If downstream
allows alpha video caps and it has already negotiated to a non-alpha
format, caps queries should still return the alpha caps as a possible
format as caps renegotiation can happen.
Includes tests (for compositor) to check that caps queries done after
a caps has been negotiated returns complete results
https://bugzilla.gnome.org/show_bug.cgi?id=757610
Expose the expose() and set_render_rectangle() methods. These are useful for
proper functioning of the video overlay in various situations and toolkits.
H.265 7.4.7.1 says:
> When slice_deblocking_filter_disabled_flag is not present, it is
> inferred to be equal to pps_deblocking_filter_disabled_flag.
https://bugzilla.gnome.org/show_bug.cgi?id=762351
Usually gl debug is initialized in gst_gl_context_create_thread.
But this function is not used when using the GstGLContextGPUProcess
from ChromiumGStreamerBackend.
Received signal 11 SEGV_MAPERR 000000000000
gst_debug_category_get_threshold
gst_gl_insert_debug_marker
gst_gl_base_filter_gl_start
CC libgstgl_x11_la-gstglcontext_glx.lo
In file included from gstglcontext_glx.c:39:0:
../utils/opengl_versions.h:52:43: error: ‘gles2_versions’ defined but not used [-Werror=unused-const-variable]
static const struct { int major, minor; } gles2_versions[] = {
^~~~~~~~~~~~~~
Invoke the callback right away when called on the context thread. Removes
overhead when nesting libgstgl calls (for example when working with the sync
meta).
CPU waits are more expensive and are only required if the CPU is ever going to
access the data. GPU waits perform inter-context synchronisation and are cheaper
as they don't require CPU intervention.
Adaptive demuxers need to start downloading from specific positions
(fragments) for every stream, this means that all streams can snap-seek
to a different position when requested. Snap seeking in this case will
be done in 2 steps:
1) do the snap seeking on the pad that received the seek event and
get the final position
2) use this position to do a regular seek on the other streams to
make sure they all start from the same position
More arguments were added to the stream_seek function, allowing better control
of how seeking is done. Knowing the flags and the playback direction allows
subclasses to handle snap-seeking.
And also adds a new return parameter to inform of the final
selected seeking position that is used to align the other streams.
https://bugzilla.gnome.org/show_bug.cgi?id=759158
All uses of query->context in gstglquery.c assume it exists. We can assume
this as well before unrefing it. Furthermore, gst_object_unref() will just
silently return if it ever were to not exist.
Improves the responsiveness of the pipeline when resources are close/above the
limitations of the hardware.
Any subclass that wishes not to enable qos can do so themselves.
https://bugzilla.gnome.org/show_bug.cgi?id=761519
This reverts commit 96b9666d59.
This reverts commit d11385d167.
This breaks the texture sharing with the applemedia elements as
CVOpenGLESTextureCache seems to have an arbitrary restriction on GLES2 only.
We may need them to transform into a different set of formats.
Fixes YUV->YUV with two glcolorconverts, e.g:
format=I420 ! glcolorconvert ! glcolorconvert ! format=NV12
1.0 / width does not offset by one pixel in rectangular textures (which use
unnormalized coordinates).
Provide the actual pixel offset as a uniform to the shader.
1. Correctly describe what we can caps we can transform to/from.
i.e. no YUV->YUV or GRAY->YUV or YUV->GRAY (except for passthrough).
2. Prefer similar formats and ignore incompatible formats on fixation.
Useful when gst_gl_window.c::gst_gl_window_new is not used.
This is the case when using a custom GstGLWindow.
(ex: GstGLWindowGPUProcess from Chromium)
Allows the subclass to completely override the chosen src caps.
This is needed as videoaggregator generally has no idea exactly
what operation is being performed.
- Adds a fixate_caps vfunc for fixation
- Merges gst_video_aggregator_update_converters() into
gst_videoaggregator_update_src_caps() as we need some of its info
for proper caps handling.
- Pass the downstream caps to the update_caps vfunc
https://bugzilla.gnome.org/show_bug.cgi?id=756207
The function gst_adaptive_demux_stream_update_source() function creates
a new GstPad called internal_pad. This pad is not freed when releasing
the stream.
The solution is to set GST_PAD_FLAG_NEED_PARENT so that the chain
functions do not get called when the pad has no parent and then
remove the parent in the gst_adaptive_demux_stream_free() function. This
causes the refcount of the pad to be set to zero.
https://bugzilla.gnome.org/show_bug.cgi?id=760982
It's useful enough already to be used in other elements for audio aggregation,
let's give people the opportunity to use it and give it some API testing.
https://bugzilla.gnome.org/show_bug.cgi?id=760733
Handling the ghostpad and its internal pad was causing more issues
than helping because of their coupled activation/deactivation
actions.
As we have to install custom chain,event and query functions it is
better to use a floating sink pad internally in the demuxer and just
use those pad functions to push through a standard pad in the demuxer
https://bugzilla.gnome.org/show_bug.cgi?id=757951
The messages are stored by gst_gl_async_debug_store_log_msg() and output later
by a corresponding store(), output() or an unset()/free().
Some wrapper macros are provided to avoid callers explicitly using __FILE__,
GST_FUNCTION and __LINE__
This makes a pipeline like:
... ! video/x-raw(memory:GLMemory),format=UYVY ! glcolorconvert !
video/x-raw(memory:GLMemory),format={UYVY, NV12} ! ...
passthrough instead of converting UYVY => NV12. The conversion would happen
before this change since the element (and basetransform) transform the src caps
to format={NV12, UYVY} (since NV12 comes first in the glcolorconvert:src
template) and then the default caps fixate func would fixate to NV12. Blah.
Also there's no need to intersect against the template caps in ::transform_caps
since basetransform does that right after calling the vfunc.
The optimistic download_transfer was not setting the required flag to not
perform glReadPixels on subsequent map (READ). resulting in glReadPixels
happening twice.
Fixed adaptivedemux seeking without flushing that just wants
to update stop position. This required protecting the segment
variables with a new mutex so that the seeking thread and the
download threads could safely manipulate the segment and
events related to it.
This contention is only locked/unlocked when starting a new
download, when the first fragment of a segment is received and
when seeking so, hopefully, it won't damage performance.
Some operations are unnecessary when running with only a single GL
context.
e.g. glFlush when setting a fence object as the flush happens on wait.
API: gst_gl_context_is_shared
Avoids downloading and pushing a full segment just to get 1 nanosecond
of data. This happens frequently when seeking is done with flags
that adjust to boundaries or when the start is aligned with segment
starts. The later is common when segment durations is a multiple of
a second.
For reverse, set position to segment.stop when starting and also
don't set the position to fragment end timestamp when it finishes,
just leave it at the fragment start.
1. Various elements/base classes only perform a subset check on accept-caps
2. Some GL elements have texture-target in their pad template
3. When checking subsets, only the caps to check are allowed to contain extra
fields. If the 'template' caps have extra fields, the subset fails.
Thus without texture-target on the caps, various accept-caps implementations
were failing.
Also, add some convenience functions for setting and retrieving
texture targets to/from GValue.
https://bugzilla.gnome.org/show_bug.cgi?id=759860
In very few cases the simple version was actually needed and having the
parameters hidden by a _full() version caused application that actually needed
it to not use it.
GL 2.1 only supports pbo upload.
The wrapped data pointer was only being set on the pbo memory and on the
glmemory so when a download was requested (in GL 2.1), glmemory was
allocating a new data pointer and thus not returning the wrapped data.
Exposing the navigation thread's main context, GSourceFuncs and structs called
key_event and mouse_event is exposing a bit too much of the internals. Let's
just go with two functions to asynchronously send navigation events on the
window with the same API as the synchronous ones.
