This is the first version of AV1 parser implementation in GStreamer.
A test file is also provied with several test cases. It contains a
test sequence taken from the aom testdata set, with one key and one
inter-frame. The same test sequence has been reencoded to annexb.
testdata is taken from aom testdata (and reencoded for annexb) as well
as handcrafted testcases. Once reference testdata is available, the
testing could be imporved aswell.
Co-author: He Junyan <junyan.he@hotmail.com>
Co-author: Víctor Manuel Jáquez Leal <vjaquez@igalia.com>
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/785>
the old manner does not consider the profile idc. The profile idc should
play an more important role in recognizing the profile than the other
information. And there is no need to mix profiles of different extensions
together to find the closest profile when the bits stream is not standard,
different extensions support different features and should not be mixed.
The correct way should be recognize the extension category by profile idc
firstly, and then find the closest profile.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1121>
FormatRangeExtensionProfile declares the common bits used for not
only format range extensions profiles, but also for several different
h265 extension profiles, such as high throughput, screen content
coding extensions, etc. And So the old name is not proper.
We also rename the get_h265_extension_profile function.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1121>
We should use the traget ExtensionProfile's IDC to check the
profile_compatibility_flag, rather than the profile_idc in the
stream. The old profile_compatibility_flag check always return
true. This causes that profiles with same constraint flags but
different profile_idc can't be recognized correctly. For example,
the screen-extended-main-444 profile is always be recognized as
the high-throughput-444 profile.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1121>
In case of IDR, any previously decoded pictures must be drained
before the IDR and POC of IDR should be zero. So we can output
IDR immediately. Also, when POC of current picture is expected to be
the next output POC, decoder can output the picture as well
without waiting.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1211>
... to prevent requesting decoding before the preparation.
For instance, baseclass should not request decoding a picture if there
is no parsed valid headers, since subclass is most likely
not ready to decoding it.
gst_h265_parser_parse_{vps,sps,pps} APIs were used to parse VPS/SPS/PPS and
also in order to update parser's internal state at once. Meanwhile
gst_h265_parse_{vps,sps,pps} APIs are to parse VPS/SPS/PPS without state update.
This commit introduces new APIs so that only accepted VPS/SPS/PPS by user
can be updated to be used by parser.
gst_h264_parser_parse_{sps,pps} APIs were used to parse SPS/PPS and
also in order to update parser's internal state at once. Meanwhile
gst_h264_parse_{sps,pps} APIs are to parse SPS/PPS without state update.
This commit introduces new APIs so that only accepted SPS/PPS by user
can be updated to be used by parser.
... and store all parsed values.
We are storing pic_struct_present_flag although it's not part of
this SEI message but GstH264PicTiming includes it to clarify
following syntax values.
In addition to that, by adding CpbDpbDelaysPresentFlag, we don't need to
refer to VUI anymore.
As per specification in A.3.1 h) and A.3.2 f), the maximum size of the DPB is
16. Fix the maximum in the fine and fix the formula to use MIN instead of MAX
so that we no longer always use the maximum for the profile/level.
If decoding failed because end_picture() failed, set the picture to
nonexisting, this way output_picture() will be skipped. This avoids confusing
special cases in output_picture() implementation.
gsth265parser does it already. Although corresponding API of h265parser is
gst_h265_sei_free, _clear suffix is more consistent naming for h264parser
since there are gst_h264_{sps,pps}_clear().
According to following two specs, add support for AC4 in tsdemux.
1. ETSI TS 103 190-2 V1.2.1 (2018-02) : Annex D (normative): AC-4 in MPEG-2 transport streams
2. ETSI EN 300 468 V1.16.1 (2019-08) : Annex D (normative):Service information implementation of AC-3, EnhancedAC-3, and AC-4 audio in DVB systems
That's the value of NumDeltaPocs[RefRpsIdx] and we might be able to derive
the value from given sps and slice header.
Because well known hardware implementations refer to the value, however,
storing the value makes things easier.
Following is the list of hardware implementations
* DXVA2: ucNumDeltaPocsOfRefRpsIdx
* NVDEC/VDPAU: NumDeltaPocsOfRefRpsIdx
See C.5.2.2 Output and removal of pictures from the DPB.
If the number of pictures in the DPB is greater than or equal to
sps_max_dec_pic_buffering_minus1[HighestTid] + 1, then the picture
should be outputted.
On new_segment, the decoder is expected to negotiate. The decoder may want to
pre-allocate the needed buffers. Pass the max_dpb_size as this is needed to
determin how many buffers should be allocated.
This introduce a library which contains a set of base classes which
handles the parsing and the state tracking for the purpose of decoding
different CODECs. Currently H264, H265 and VP9 are supported. These
bases classes are used to decode with low level decoding API like DXVA,
NVDEC, VDPAU, VAAPI and V4L2 State Less decoders. The new library is
named gstreamer-codecs-1.0 / libgstcodecs.
Add a missing dependency to wl_client_dep for the wayland build. Some distros
have the wayland-client headers not installed in /usr/include (which is perfectly
valid, the pkg-config .pc file gives the right feedback).
Instead of synchronising at the ICE transport, do clock sync for the
RTP stream at the DTLS transport via the dtlssrtpenc rtp-sync
property. This avoids delaying RTCP while waiting until it is time
to output an RTP packet when rtcp-mux is enabled.
https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/issues/1212
Posting any message to parent seems to be pointless. That might break
parent window.
Regardless of the posting, parent window can catch mouse event
and also any keyboard events will be handled by parent window by default.
It's currently the only sane way we can use MoltenVK functions to
integrate with Metal API.
It also removes the need to specify the VK_ICD_FILENAMES environment
variable pointing to MoltenVK_icd.json.
Includes a new GstVulkanHandlePool base class for pooling different
resources togther. The descriptor cache object is ported to
GstVulkanHandlePool with the exact same functionality.
A new GstVulkanFenceCache is also implemented for caching fences
which is used internally by GstVulkanDevice for creating or reusing
fences.
The existing GstVulkanTrashFenceList object now caches trash objects.
Part 1 is a base class (vkvideofilter) that handles instance, device,
queue retrieval and holding that has been moved to the library
Part 2 is a fullscreenrenderquad that is still in the plugin that
performs all of the previous vulkan-specific functionality.
Weak refs don't quite work here correctly as there is always a race with
taking the lock between find_view() and remove_view(). If find_view()
returns a view that is going to removed by remove_view() then we have an
interesting situation.
In theory, the number and type of views for an image are relatively
constant and should not change one they've been set up which means that
it is actually practical to perform pool-like reference counting here
where the image holds a pool of different views that it can give out
as necessary.
It broke after removal of usage of GTimeVal that was deprecated,
it requires seconds in this unix-based creation instead of microseconds.
The downside here is that it will create an extra object just to be
discarded in order to add the microsecond part to it.
It would end up segfaulting as it would return a NULL value
Exposure mode property, extra colour tone values (aqua, emboss, sketch, neon), extra scene modes (backlight, flowers, AR, HDR).
Missing vmethods for exposure mode, analog gain, lens focus, colour temperature, min & max exposure time
Contribs by Mohammed Sameer <msameer@foolab.org>, Adam Pigg <adam@piggz.co.uk>
The symbol visibility=hidden flag was only being applied to C
compilation, so plugins implemented in C++ would leak extra symbols
than the 2 _get_desc() and _register().
That also showed that the gst-libs opencv C++ lib was not marking
symbols for export correctly because the BUILDING_GST_OPENCV define
wasn't in the C++ args, so fix that too.
The major functionality gain this provides is proper reference counting
for a descriptor set. Overall this allows us to create descriptor sets
when they are needed (or reused from a cache) without violating any of
vulkan's object synchronisation requirements.
As there are a fixed number of sets available in a pool, the number of
descriptors in elements is currently hardcoded to 32. This can be extended
in a future change to create pools on the fly if that limit is ever overrun.
Allows a cleaner control flow when there is no fence available for use
with the trash list. An always signalled fence type will always return
TRUE for gst_vulkan_fence_is_signalled.
Serve two purposes:
1. refcounting of vulkan handles with associated destruction. When
combined with the trash list, the user can ensure destruction at
the correct time according to the vulkan rules.
2. avoids polluting our API with 32-bit vs 64-bit integer/pointers
differences as exposed through the vulkan API. on 32-bit, vulkan
non-dispatchable handles are 64-bit integers and on 64-bit, they
are pointers.
By passing NULL to `g_signal_new` instead of a marshaller, GLib will
actually internally optimize the signal (if the marshaller is available
in GLib itself) by also setting the valist marshaller. This makes the
signal emission a bit more performant than the regular marshalling,
which still needs to box into `GValue` and call libffi in case of a
generic marshaller.
Note that for custom marshallers, one would use
`g_signal_set_va_marshaller()` with the valist marshaller instead.
This adds two properties:
* scte-35-pid: If not 0, enables the SCTE-35 support for the current
program. This will write the proper PMT and send SCTE-35 NULL
commands (i.e. heartbeats) at a regular interval
* scte-35-null-interval: This specifies the interval at which the
NULL commands should be sent
Sending SCTE-35 commands is done by creating the appropriate SCTE-35
GstMpegtsSection and then sending them on the muxer. See the
associated example
Commit 8a56a7de6d (camerabin2: preview:
Appsink doesn't need to sync) add a line that set the "sync" property on
the appsink. However, the author seems to forget that there's another
property setting on appsink a few lines below.
It's very likely that the added line is required because the original
line doesn't take effect (maybe because it's too late). But for whatever
reason, the original line is now redundant. So, I remove it in this
commit.
GstTranscoder adds extra_args for gir which call gst_init() during
introspection. These extra arguments are the same as the standard
ones defined in the top level meson.build as "git_init_section",
However, the top level definition also ensures an empty plugin
repository is used.
Because GstTranscoder does not use the standard args, plugins get
loaded when it is introspected. Since some of the plugins fail
without specific hardware, this causes #1100.
This patch makes it use gir_init_section.
Fixes#1100.
Some hardware decoders, for example Hantro G1, have to be told the
size of the pic_order_cnt related syntax elements pic_order_cnt_lsb,
delta_pic_order_cnt_bottom, delta_pic_order_cnt[0], and
delta_pic_order_cnt[1] in bits.
Some hardware decoders, for example Hantro G1, have to be told the size
of the dec_ref_pic_marking() syntax element in bits. Record the size so
it can be passed on to the hardware.
The calculated size of short_term_ref_pic_set is not a part of
HEVC syntax but the value is used by some stateless decoders
(e.g., vaapi, dxva, vdpau and nvdec) for the purpose of skipping
parsing the syntax by the accelerator.
Add num_ref_idx_active_override_flag and sp_for_switch_flag to
member of GstH264SliceHdr. No reason to hiding them and
some decoder implementations (e.g., DXVA) rely on externally parsed header
data which can be provided by h264parser.
* Fix meson build script for Windows. Since the Vulkan dependency
object was declared by us in case of Windows, the dependency object
shouldn't be used for finding header
* Fix build error by including Windows specific header
gstvkdisplay.c(563): error C2065: 'VK_KHR_WIN32_SURFACE_EXTENSION_NAME': undeclared identifier
1. The iOS create_surface implementation needs to call out to the main
thread to create the window (UIKit requirement)
2. get_surface() can be called and will attempt to create the VkSurface
from an invalid view/layer.
Also pass the layer for MoltenVK so we don't get pesky 'UIView function
not called on main thread' warnings.
Fix gst_event_new_seek call in gst-libs/gst/player/gstplayer.c
If rate >= 0.0, then previous code doesn't set end of segment. So, the end of segment
will be in place where previous seek put it. This is not neccesary end of media file
(in case of reverse playback). So if we play video backward for some time and then
switched to forward playing, we will get EOS somewhere in the middle of media file.
This commit always sets end of segment, thus fixing this bug
During resizes, the VkQueuePresent can return OUT_OF_DATE and if a buffer
is displayed returning OUT_OF_DATE it would error out and stop the pipeline.
We already have a explicit check for OUT_OF_DATE and the same general
error check in the statements following so just use that code.
NULL checking the main_context does not help as we've just destroyed the
GMainContext and set that field to NULL, not to mention it's unnecessary.
Fixes a leak of display's GSource.
Add a separate epb_cache variable to the codecparser NalReader to
detect Emulation Prevention Bytes separately from the main bit cache.
This fixes problems where the existing logic can mistakenly detect
multiple EPB with a sequence like: 0x00 0x00 0x03 0x00 0x03. In that
case, the 5th byte should not be regarded as an EPB.
3-byte emulation bytes can confuse the current code that skips
bits at the end of an SEI. Use a simpler method that's also
quicker because it skips all remaining bits in one go instead
of 1 bit at a time.
WebKit's websrc depends on the main-thread for download completion
rendezvous. This exposed a number of deadlocks in adaptivedemux due to
it holding the MANIFEST_LOCK during network requests, and also needing
to hold it to change_state and resolve queries, which frequently occur
during these download windows.
Make demux->running MT-safe so that it can be accessed without using the
MANIFEST_LOCK. In case a source is downloading and requires a MT-thread
notification for completion of the fragment download, a state change
during this download window will deadlock unless we cancel the downloads
and ensure they are not restarted before we finish the state-change.
Also make demux->priv->have_manifest MT-safe. A duration query happening
in the window described above can deadlock for the same reason. Other
src queries (like SEEKING) that happen in this window also could
deadlock, but I haven't hit this scenario.
Increase granularity of API_LOCK'ing in change_state as well. We need to
cancel downloads before trying to take this lock, since sink events
(EOS) will hold it before starting a fragment download.
The SPS parsing functions take a parse_vui_param flag
to skip VUI parsing, but there's no indication in the output
SPS struct that the VUI was skipped.
The only caller that ever passed FALSE seems to be the
important gst_h264_parser_parse_nal() function, meaning - so the
cached SPS were always silently invalid. That needs changing
anyway, meaning noone ever passes FALSE.
I don't see any use for saving a few microseconds in
order to silently produce garbage, and since this is still
unstable API, let's remove the parse_vui_param.
The spec calls for pic_timing SEI to be absent unless
there's either a CpbDpbDelaysPresentFlag or
pic_struct_present_flag in the SPS VUI data. If
both those flags are missing, warn.
If parsing an SEI errors out, it might not consume
all bits, leaving extra unparsed data in the reader
that the outer loop then tries to parse as a new
appended SEI.
Skip all the bits if any are left over to avoid
'finding' extra garbage SEI in the parsing.
When parsing SEI that require an SPS, return
GST_H264_PARSER_BROKEN_LINK instead of a generic
parsing error to let callers distinguish
bitstream errors from (expected) missing packets
when resuming decode.
"High Throughput", "Multiview", "Scalable", "3D", "Screen Content Coding",
and "Scalable format range extensions" profiles can be supported
via h265parser APIs now.
The same timing_info will be present at vps or vui.
When the timeing_info is present in the VPS, vui_timing_info
, when present, shall be equal to vps_timing_info, and when
not present, is inferred to be equal to vps_timing_info.
It's almost a fork of glwindow_win32 implementation.
To build on Windows, Vulkan SDK (at https://vulkan.lunarg.com/sdk/home)
and VK_SDK_PATH environment are required. Note that VK_SDK_PATH environment
setting is a part of the SDK installation.
Limitations:
- No transport changes at all (ICE, DTLS)
- Codec changes are untested and probably don't work
- Stream removal doesn't remove transports (i.e. non-bundled transports
will stay around until webrtcbin is shutdown)
- Unified Plan SDP only. No Plan-B support.
