vaapiencode might hang when the pipeline is stopped without any EOS,
e.g. when <Ctrl>+C is pressed, thus causing the srcpad task to keep
running and locked. This fixes a possible deadlock on state change
from PAUSED to READY.
https://bugzilla.gnome.org/show_bug.cgi?id=720584
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
Add gst_vaapi_video_format_new_template_caps_with_features() helper
function to add the supplied caps feature string on GStreamer >= 1.2.
Add gst_vaapi_find_preferred_caps_feature() helper function to discover
the "best" caps feature to use for the supplied pad. In practice, we
will always favor memory:VASurface first, then meta:GLTextureUploadMeta,
and finally the system memory caps.
https://bugzilla.gnome.org/show_bug.cgi?id=719372
The GLTextureUpload function is not in charge of doing the overlay
composition if any.
https://bugzilla.gnome.org/show_bug.cgi?id=721859
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
Allow user to precise the largest profile to use for encoding due
to target decoder constraints. For instance, if CABAC entropy coding
mode is requested by "constrained-baseline" profile only is desired,
then an error is returned during codec configuration.
Also make sure that the suitable profile we derived actually matches
what the HW can cope with.
https://bugzilla.gnome.org/show_bug.cgi?id=719694
Various clean-ups to improve consistency and readability: rename some
variables, drop unused macro definitions, drop initialization of vars
that are zero-initialized from the base class, drop un-necessary casts.
Add interface to communicate the encoder resolution and related info
like framerate, interlaced vs. progressive, etc. This new interface
supersedes gst_vaapi_encoder_set_format() and doesn't use any GstCaps
but rather use GstVideoCodecState.
Note that gst_vaapi_encoder_set_codec_state() is also a synchronization
point for codec config. This means that the encoder is reconfigured
there to match the latest properties.
Don't try to destroy an encoder, in GstVideoEncoder::finish() handler,
if it was not created in the first place. Return "not-negotiated" error
since this means we did not even reach GstVideoEncoder::set_format(),
where the encoder could have been created.
This fixes a crash when the vaapiencode_* plug-in elements get deallocated
and that we failed to negotiate either pad.
https://bugzilla.gnome.org/show_bug.cgi?id=719704
The specified caps in gst_video_encoder_set_output_state() function
arguments should not contain any resolution, pixel-aspect-ratio,
framerate, codec-data et al. Those rather should be set through the
returned GstVideoCodecState. This means that output caps creation
could be delayed until before gst_video_encoder_finish_frame() is
called.
This greatly simplifies the GstVideoEncoder::set_format() callback
by the way.
Make base GstVaapiEncode class an abstract type so that we cannot
create an instance from it without going through any of the codec
specific derived class.
Rename a few member functions to make them more consistent:
- alloc_encoder(): now reduced to allocate the encoder object only;
- alloc_buffer(): allocate buffer from srcpad, and copy bitstream.
Update MPEG-2 and H.264 encode elements to cope with the new core
libgstvaapi properties API. i.e. all configurable properties are now
directly handled at the GstVaapiEncoder level.
Besides, this also makes sure to not use or modify the GstVaapiEncoder
private definitions directly. Private data need to remain private.
https://bugzilla.gnome.org/show_bug.cgi?id=719529
Add gst_vaapi_encoder_set_bitrate() interface to allow the user control
the bitrate for encoding. Currently, changing this parameter is only
valid before the first frame is encoded. Should the value be modified
afterwards, then GST_VAAPI_ENCODER_STATUS_ERROR_OPERATION_FAILED is
returned.
https://bugzilla.gnome.org/show_bug.cgi?id=719529
Add gst_vaapi_encoder_set_rate_control() interface to request a new
rate control mode for encoding. Changing the rate control mode is
only valid prior to encoding the very first frame. Afterwards, an
error ("operation-failed") is issued.
https://bugzilla.gnome.org/show_bug.cgi?id=719529
Fixes a crash when multiple vaapidecode elements are finalized since
the debug category is created once in the class init method.
This is a regression from git commit 7e58d60.
https://bugzilla.gnome.org/show_bug.cgi?id=721390
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
Fix formts for various GST_DEBUG et al. invocations. More precisely,
make size_t arguments use the %zu format specifier accordingly; force
XID formats to be a 32-bit unsigned integer; and fix the format used
for gst_vaapi_create_surface_with_format() error cases since we have
been using strings nowadays.
Use standard GstVideoInfo related functions to build the output caps,
thus directly preserving additional fields as needed, instead of
manually copying them over through gst_vaapi_append_surface_caps().
Also ensure that the input caps are fixated first.
Add new helper functions to build video template caps.
- gst_vaapi_video_format_new_template_caps():
create GstCaps with size, frame rate and PAR to full range
- gst_vaapi_video_format_new_template_caps_from_list():
try to create a "simplified" list from the supplied formats
Add new helper functions to build GValues from GstVideoFormat:
- gst_vaapi_value_set_format():
build a GValue from the supplied video format
- gst_vaapi_value_set_format_list():
build a GValue list from the supplied array of video formats
Replace gst_vaapi_display_get_{decode,encode}_caps() APIs with more
more convenient APIs that return an array of GstVaapiProfile instead
of GstCaps: gst_vaapi_display_get_{decode,encode}_profiles().
Replace gst_vaapi_display_get_{image,subpicture}_caps() APIs, that
returned GstCaps, with more convenient APIs that return an array of
GstVideoFormat: gst_vaapi_display_get_{image,subpicture}_formats().
Makes the copies of a buffer reference their own GLTextureUploadMeta
user data and prevent the original buffer accessing already freed
memory if its copies has been released and freed.
https://bugzilla.gnome.org/show_bug.cgi?id=720336
[Propagate the original meta texture to the copy too]
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
Factor out propose_allocation() hooks, creation of video buffer pool
for the sink pad, conversion from raw YUV buffers to VA surface backed
buffers. Update vaapidecode, vaapiencode and vaapipostproc to cope
with the new GstVaapiPluginBase abilities.
Fix display creation code to check that any display obtained from a
neighbour actually has the type we expect. Note: if display type is
set to "any", we can then accept any VA display type.
Move common VA display creation code to GstVaapiPluginBase, with the
default display type remaining "any". Also add a "display-changed"
hook so that subclasses could perform additional tasks when/if the
VA display changed, due to a new display type request for instance.
All plug-ins are updated to cope with the new internal APIs.
Introduce a new GstVaapiPluginBase object that will contain all common
data structures and perform all common tasks. First step is to have a
single place to hold VA displays.
While we are at it, also make sure to store and subsequently release
the appropriate debug category for the subclasses.
The GLTextureUploadMeta implementation assumed that for each upload()
sequence, the supplied texture id is always the same as the one that
was previously cached into the underlying GstVaapiTexture. Cope with
any texture id change the expense to recreate the underlying VA/GLX
resources.
https://bugzilla.gnome.org/show_bug.cgi?id=719643
Requesting the GLTextureUpload meta on buffers in the bufferpool
prevents such metas from being de-allocated when buffers are released
in the sink.
This is particulary useful in terms of performance when using the
GLTextureUploadMeta API since the GstVaapiTexture associated with
the target texture is stored in the meta.
https://bugzilla.gnome.org/show_bug.cgi?id=712558
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
Make GstVideoGLTextureUploadMeta::upload() implementation more robust
when the GstVaapiTexture associated with the supplied texture id could
not be created.
Clean public APIs up so that to better align with the decoder APIs.
Most importantly, gst_vaapi_encoder_get_buffer() is changed to only
return the VA coded buffer proxy. Also provide useful documentation
for the public APIs.
Refactor the GstVaapiCodedBuffer APIs so that to more clearly separate
public and private interfaces. Besides, the map/unmap APIs should not
be exposed as is but appropriate accessors should be provided instead.
* GstVaapiCodedBuffer: VA coded buffer abstraction
- gst_vaapi_coded_buffer_get_size(): get coded buffer size.
- gst_vaapi_coded_buffer_copy_into(): copy coded buffer into GstBuffer
* GstVaapiCodedBufferPool: pool of VA coded buffer objects
- gst_vaapi_coded_buffer_pool_new(): create a pool of coded buffers of
the specified max size, and bound to the supplied encoder
* GstVaapiCodedBufferProxy: pool-allocated VA coded buffer object proxy
- gst_vaapi_coded_buffer_proxy_new_from_pool(): create coded buf from pool
- gst_vaapi_coded_buffer_proxy_get_buffer(): get underlying coded buffer
- gst_vaapi_coded_buffer_proxy_get_buffer_size(): get coded buffer size
Rationale: more optimized transfer functions might be provided in the
future, thus rendering the map/unmap mechanism obsolete or sub-optimal.
https://bugzilla.gnome.org/show_bug.cgi?id=719775
Fix GstElement::set_context() implementation for all plug-in elements
to avoid leaking an extra reference to the VA display, thus preventing
correct cleanup of VA resources in GStreamer 1.2 builds.
Return earlier if the creation of a VA display failed. Likewise, simplify
gst_vaapi_video_context_propagate() now that we are guaranteed to have a
valid VA display.
When GstVideoMeta maps were used, the supporting functions incorrectly
used gst_buffer_get_memory() instead of gst_buffer_peek_memory(), thus
always increasing the associated GstMemory reference count and giving
zero chance to actually release that, and subsequently the VA display.
Simplify GstVaapiVideoMeta to only hold a surface proxy, which is
now allocated from a surface pool. This also means that the local
reference to the VA surface is also gone, as it could be extracted
from the associated surface proxy.
Drop the following functions that are not longer used:
- gst_vaapi_video_buffer_new_with_surface()
- gst_vaapi_video_meta_new_with_surface()
- gst_vaapi_video_meta_set_surface()
- gst_vaapi_video_meta_set_surface_from_pool()
Fix gst_vaapi_video_meta_new_from_pool() to allocate VA surface proxies
from surface pools instead of plain VA surfaces. This is to simplify
allocations now that surface proxies are created from a surface pool.
Optimize gst_vaapiencode_handle_frame() to avoid extra memory allocation,
and in particular the GstVaapiEncObjUserData object. i.e. directly use
the VA surface proxy from the source buffer. This also makes the user
data attached to the GstVideoCodecFrame more consistent between both
the decoder and encoder plug-in elements.
Simplify gst_vaapiencode_push_frame(), while also removing the call
to gst_video_encoder_negotiate() since this is implicit in _finish()
if caps changed. Also fixed memory leaks that occured on error.
Constify pointers wherever possible. Drop unused variables, and use
consistent variable names. Fix gst_vaapiencode_h264_allocate_buffer()
to correctly report errors, especially when in-place conversion from
bytestream to avcC format failed.
Move "rate-control" mode and "bitrate" properties to the GstVaapiEncode
base class. The actual range of supported rate control modes is currently
implemented as a plug-in element hook. This ought to be determined from
the GstVaapiEncoder object instead, i.e. from libgstvaapi.
Align the plug-in debug category to its actual name. i.e. enable debug
logs through vaapiencode_<CODEC> where <CODEC> is mpeg2, h264, etc. Fix
the plug-in element description to make it more consistent with other
VA-API plug-ins.
Add a GST_VAAPIENCODE_CAST() helper to avoid run-time checks against
the GObject type system. We are guaranteed to only deal with the same
plug-in element object.
Allow vaapiencode plug-in elements to encode from raw YUV buffers.
The most efficient way to do so is to let the vaapiencode elements
allocate a buffer pool, and subsequently buffers from it. This means
that upstream elements are expected to honour downstream pools.
If upstream elements insist on providing their own allocated buffers
to the vaapiencode elements, then it possibly would be more efficient
to insert a vaapipostproc element before the vaapiencode element.
This is because vaapipostproc currently has better support than other
elements for "foreign" raw YUV buffers.
Add GstVaapiEncodeMPEG2 element object. The actual plug-in element
is called "vaapiencode_mpeg2".
Valid properties:
- rate-control: rate control mode (default: cqp - constant QP)
- bitrate: desired bitrate in kbps (default: auto-calculated)
- key-period: maximal distance between two key frames (default: 30)
- max-bframes: number of B-frames between I and P (default: 2)
- quantizer: constant quantizer (default: 8)
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
Add GstVaapiEncodeH264 element object. The actual plug-in element
is called "vaapiencode_h264".
Valid properties:
- rate-control: rate control mode (default: none)
- bitrate: desired bitrate in kbps (default: auto-calculated)
- key-period: maximal distance between two key frames (default: 30)
- num-slices: number of slices per frame (default: 1)
- max-bframes: number of B-frames between I and P (default: 0)
- min-qp: minimal quantizer (default: 1)
- init-qp: initial quantizer (default: 26)
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
Fix build when Wayland headers don't live in plain system include dirs
like /usr/include but rather in /usr/include/wayland for instance.
Original patch written by Dominique Leuenberger <dimstar@opensuse.org>
https://bugzilla.gnome.org/show_bug.cgi?id=712282
Destroy VPP output surface pool on exit. Also avoid a possible crash
in double-free situation caused by insufficiently reference counted
array of formats returned during initialization.
Fix advanced deinterlacing modes with VPP to track only up to 2 past
reference buffers. This used to be 3 past reference buffers but this
doesn't fit with the existing decode pipeline that only has 4 extra
scratch surfaces.
Also optimize references tracking to be only enabled when needed, i.e.
when advanced deinterlacing mode is used. This means that we don't
need to track past references for basic bob or weave deinterlacing.
In "mixed" interlaced streams, the buffer contains additional flags that
specify whether the frame contained herein is interlaced or not. This means
that we can alternatively get progressive or interlaced frames. Make sure
to disable deinterlacing at the VPP level when the source buffer is no longer
interlaced.
Fix memory leaks with advanced deinterlacing, i.e. when we keep track
of past buffers. Completely reset the deinterlace state, thus destroying
any buffer currently held, on _start(), _stop() and _destroy().
Port vaapipostproc element to GStreamer 1.2. Support is quite minimal
right now so that to cope with auto-plugging issues/regressions. e.g.
this happens when the correct set of expected caps are being exposed.
This means that, currently, the proposed caps are not fully accurate.
Fix basic deinterlacing flags provided to gst_vaapi_set_deinterlacing()
for the first field. Render flags were supplied instead of the actual
deinterlacing flags (deint_flags).
Fix GstBaseTransform::transform_caps() implementation to always return
the complete set of allowed sink pad caps (unfixated) even if the src
pad caps we are getting are fixated. Rationale: there are just so many
possible combinations, and it was wrong to provide a unique set anyway.
As a side effect, this greatly simplifies the ability to derive src pad
caps from fixated sink pad caps.
Fix deinterlacing flags to make more sense. The TFF (top-field-first)
flag is meant to specify the organization of reference frames used in
advanced deinterlacing modes. Introduce the more explicit flag TOPFIELD
to specify that the top-field of the supplied input surface is to be
used for deinterlacing. Conversely, if not set, this means that the
bottom field of the supplied input surface will be used instead.
There are situations where gst_video_decoder_flush() is called, and
this subsequently produces a gst_video_decoder_reset() that kills the
currently active GstVideoCodecFrame. This means that it no longer
exists by the time we reach GstVideoDecoder::finish() callback, thus
possibly resulting in a crash if we assumed spare data was still
available for decode (current_frame_size > 0).
Try to honour GstVideoDecoder::reset() behaviour from GStreamer 1.0
that means a flush, thus performing the actual operations there like
calling gst_video_decoder_have_frame() if pending data is available.
Review all interactions between the main video decoder stream thread
and the decode task to derive a correct sequence of operations for
decoding. Also avoid extra atomic operations that become implicit under
the GstVideoDecoder stream lock.
Fix hard reset for seek cases by flushing the GstVaapiDecoder queue
and completely purge any decoded output frame that may come out from
it. At this stage, the GstVaapiDecoder shall be in a complete clean
state to start decoding over new buffers.
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
vaapidecode used to wait up to one second past the expected time of
presentation for the last decoded frame. This is not realistic in
practice when it comes to video pause/resume. Changed behaviour to
unconditionnally wait for a free VA surface prior to continuing the
decoding. The decode task will continue pushing the output frames to
the downstream element while also reporting errors at the same time
to the main thread.
https://bugzilla.gnome.org/show_bug.cgi?id=707108
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
The srcpad caps exposed for GStreamer 1.2 were missing any useful info
like framerate, pixel-aspect-ratio, interlace-mode et al. Not to mention
that it relied on possibly un-initialized data. Fix srcpad caps to be
initialized from a sanitized copy of GstVideoDecoder output state caps.
Note: the correct way to expose the srcpad caps triggers an additional
issue in core GStreamer auto-plugging capabilities as the correct caps
to be exposed should be format=ENCODED with memory:VASurface caps feature
at the minimum. In some situations, we could determine the underlying
VA surface format, but this is not always possible. e.g. cases where it
is not allowed to expose the underlying VA surface data, or when the
VA driver implementation cannot actually provide such information.
Currently, the decoder only supports YUV 4:2:0 output. So, expose the
output formats for GStreamer 1.2 in caps to a realistic subset. This
means NV12, I420 or YV12 but also ENCODED if we cannot determine the
underlying VA surface format, or if it is actually not allowed to get
access to the surface contents.
Fix vaapidecode srcpad caps to only expose RGBA video format for the
meta:GstVideoGLTextureUploadMeta feature. That's only what is supported
so far. Besides, drop this meta from the vaapisink sinkpad caps since
we really don't support that for rendering.
https://bugzilla.gnome.org/show_bug.cgi?id=711828
Fix raw YUV data uploaded as in the following pipeline:
$ gst-launch-1.0 filesrc video.yuv ! videoparse ! vaapipostproc ! vaapisink
The main reason why it failed was that the videoparse element simply
allocates GstBuffer with raw data chunk'ed off the sink pad without
any prior knowledge of the actual frame info. i.e. it basically just
calls gst_adapter_take_buffer().
We could avoid the extra copy performed in vaapipostproc if the videoparse
element was aware of the downstream pool and bothers copying line by
line, for each plane. This means that, for a single frame per buffer,
the optimizatin will be to allocate the video buffer downstream, map
it, and copy each line that is coming through until we need to fills
in the successive planes.
Still, optimized raw YUV uploads already worked with the following:
$ gst-launch-1.0 videotestsrc ! vaapipostproc ! vaapisink
https://bugzilla.gnome.org/show_bug.cgi?id=711250
[clean-ups, fixed error cases to unmap and unref outbuf]
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
If the currently selected deinterlacing method is not supported by the
underlying hardware, then try to downgrade the method to a supported one.
At the minimum, basic bob-deinterlacing shall always be supported.
Allow basic bob-deinterlacing to work when VPP is enabled. Currently,
this only covers bob-deinterlacing when the output pixel format is
explicitly set.
Add initial support for basic scaling with size specified through the
"width" and "height" properties. If either user-provided dimension is
zero and "force-aspect-ratio" is set to true (the default), then the
other dimension is scaled to preserve the aspect ratio.
If VPP is available, we always try to implicitly convert the source
buffer to the "native" surface format for the underlying accelerator.
This means that no optimization is performed yet to propagate raw YUV
buffers to the downstream element as is, if VPP is available. i.e. it
will always cause a color conversion.
Even if we only support deinterlacing for now, use flags to specify
which filters are to be applied to each frame we receive in transform().
This is preparatory work for integrating new filters.
Add support for "mixed" interlace-mode, whereby the video frame buffer
shall be deinterlaced only if its flags mention that's actually an
interlaced frame buffer.
Reset the buffer pool allocator only if the config caps changed in a
sensible way: format or resolution change. i.e. don't bother with
other caps like colorimetry et al. as this doesn't affect the way to
allocate VA surfaces or images.
Enable read and write mappings only if direct-rendering is supported.
Otherwise, this means that we may need to download data from the VA
surface first for correctness, even if the VA surface doesn't need to
be read at all. i.e. sometimes, READWRITE mappings are meant for
surfaces that are written to first, and read afterwards for further
processing.
https://bugzilla.gnome.org/show_bug.cgi?id=704078
Fix check for direct-rendering if the creation of VA surfaces with
an explicit pixel format is not support, e.g. VA-API < 0.34.0, and
that we tried to allocate a VA surface based on the corresponding
chroma type. i.e. in that particular case, we have to make sure that
the derived image has actually the expected format.
Fix GstVaapiVideoBufferPool::reset_buffer() to reset the underlying
memory resources, and more particularly the VA surface proxy. Most
importantly, the GstVaapiVideoMeta is retained. Cached surface in
memory are released, thus triggering a new allocation the next time
we need to map the buffer.
Make sure GstVaapiVideoMemory allocates VA surface proxies from a
pool stored in the parent VA memory allocator.
This fixes the following scenario:
- VA video buffer 1 is allocated from a buffer pool
- Another video buffer is created, and inherits info from buffer 1
- Buffer 1 is released, thus pushing it back to the buffer pool
- New buffer alloc request comes it, this yields buffer 1 back
- At this stage, buffers 1 and 2 still share the same underlying VA
surface, but buffer 2 was already submitted downstream for further
processing, thus conflicting with additional processing we were
about to perform on buffer 1.
Maybe the core GstBufferPool implementation should have been fixed
instead to actually make sure that the returned GstBuffer memory we
found from the pool is writable?
Always make sure to allocate a VA surface proxy for GstVaapiUploader
allocated buffers, i.e. make gst_vaapi_uploader_get_buffer() allocate
a proxy surface.
This fixes cases where we want to retain the underlying surface longer,
instead of releasing it back to the surface pool right away.
