While invoking the DPB bumping process in presence of many views,
it could be necessary to output previous pictures that are ready,
in a whole. i.e. emitting all view components from the very first
view order index zero to the very last one in its original access
unit; and not starting from the view order index of the picture
that caused the DPB bumping process to be invoked.
As a reminder, the maximum number of frames in DPB for MultiView
High profile with more than 2 views is not necessarily a multiple
of the number of views.
This fixes decoding of MVCNV-4.264.
Let the utility layer handle dynamic growth of the inter-view pictures
array. By definition, setting a new size to the array will effectively
grow the array, but would also fill in the newly created elements with
empty entries (NULL), thus also increasing the reported length, which
is not correct.
When decoding Multiview High profile streams with a large number of
views, it is not possible to make the VAPictureParameterBufferH264.
ReferenceFrames[] array hold the complete DPB, with all possibly
active pictures to be used for inter-view prediction in the current
access unit.
So reduce the scope of the ReferenceFrames[] array to only include
the set of reference pictures that are going to be used for decoding
the current picture. Basically, this is a union of all RefPicListX[]
array, for all slices constituting the decoded picture.
The inter-view reference components and inter-view only reference
components that are included in the reference picture lists shall
be considered as not being marked as "used for short-term reference"
or "used for long-term reference". This means that reference flags
should all be removed from VAPictureH264.flags.
This fixes decoding of MVCNV-2.264.
If the VA driver exposes ad-hoc H.264 MVC profiles, then we have to
be careful to detect profiles changes and not reset the underlying
VA context erroneously. In MVC situations, we could indeed get a
profile_idc change for every SPS that gets activated, alternatively
(base-view -> non-base view -> base-view, etc.).
An improved fix would be to characterize the exact profile to use
once and for all when SPS NAL units are parsed. This would also
allow for fallbacks to a base-view decoding only mode.
Exclusively use VA drivers that support raw packed headers for encoding.
i.e. simply submit packed headers Subset SPS and Prefix NAL units. This
provides for better compatibility accross the various VA drivers and HW
generations since no particular API is needed beyond what readily exists.
Since we are encoding each view independently from each other, we
need a higher number of pre-allocated surfaces to be used as the
reconstructed frames. For Stereo High profile encoding, this means
to effectively double the number of frames to be stored in the DPB.
Add initial support for Subset SPS, Prefix NAL and Slice Extension NAL
for non-base-view streams encoding, and the usual SPS, PPS and Slice
NALs for base-view encoding.
The H.264 Stereo High profile encoding mode will be turned on when the
"num-views" parameter is set to 2. The source (raw) YUV frames will be
considered as Left/Right view, alternatively.
Each of the two views has its own frames reordering pool and reference
frames list management system. Inter-view references are not supported
yet, so the views are encoded independently from each other.
Signed-off-by: Li Xiaowei <xiaowei.a.li@intel.com>
[limited to Stereo High profile per the definition of MAX_NUM_VIEWS]
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
Create structures to maintain the reference frames list (RefPool) and
frames reordering (ReorderPool) logic.
This is a prerequisite for H.264 MVC support.
Signed-off-by: Li Xiaowei <xiaowei.a.li@intel.com>
Add provisions to write subset SPS headers to the bitstream in view
to supporting the H.264 MVC specification.
This assumes the libva "staging" branch is in use.
Signed-off-by: Li Xiaowei <xiaowei.a.li@intel.com>
Optimize lookups of view ids / view order indices by caching the result
of the calculatiosn right into the GstVaapiParserInfoH264 struct. This
terribly simplifies is_new_access_unit() and find_first_field() functions.
Add safe fallbacks for MVC profiles:
- all MultiView High profile streams with 2 views at most can be decoded
with a Stereo High profile compliant decoder ;
- all Stereo High profile streams with only progressive views can be
decoded with a MultiView High profile compliant decoder ;
- all drivers that support slice-level decoding could normally support
MVC profiles when the DPB holds at most 16 frames.
In order to have a stricter conforming implementation, we need to carefully
detect access unit boundaries. Additional operations could be necessary to
perform at those boundaries.
Detect the first VCL NAL unit of a picture for MVC, based on the
view_id as per H.7.4.1.2.4. Note that we only need to detect new
view components.
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
Always cache the previous NAL unit so that we could check whether
there is a Prefix NAL unit immediately preceding the current slice
or IDR NAL unit. In that case, the NAL unit metadata is copied into
the current NAL unit. Otherwise, some default values are inferred,
tentatively. e.g. view_id shall be set to 0 and inter_view_flag to 1.
[infer default values for slice if previous NAL was not a Prefix]
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
Allow decoding for base views of MVC encoded streams. For now, just skip
the slice extension and prefix NAL units, and skip non-base view frames.
Signed-off-by: Xiaowei Li <xiaowei.a.li@intel.com>
[fixed memory leak, improved check for MVC NAL units]
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
Factor out process by which the decoded picture with the lowest POC
is found, and possibly output. Likewise, the storage and marking of
a reference decoded, or non-reference decoded picture, into the DPB
could also be simplified as they mostly share the same operations.
Make init_picture_ref_lists() more consistent with other functions
related to the reference marking process by supplying the current
picture as argument.
Add gst_vaapi_display_get_vendor_string() helper function to query
the underlying VA driver name. The display object owns the resulting
string, so it shall not be deallocated.
That function is thread-safe. It could be used for debugging purposes,
for instance.
Make sure to initialize one GstVaapiDisplay at a time, even in threaded
environments. This makes sure the display cache is also consistent
during the whole display creation process. In the former implementation,
there were risks that display cache got updated in another thread.
Add support for dynamic growth of the VA surfaces pool. For decoding,
this implies the recreation of the underlying VA context, as per the
requirement from VA-API. Besides, only increases are supported, not
shrinks.
It is a requirement from VA-API specification that the VA context got
from vaCreateContext(), for decoding purposes, binds the supplied set
of VA surfaces. This means that if the set of VA surfaces is to be
changed for the current decode session, then the VA context needs to
be recreated with the new set of VA surfaces.
Complement fix committed as e95a42e.
The H.264 AVC standard has to say: if the field is part of a reference
frame or a complementary reference field pair, and the other field of
the same reference frame or complementary reference field pair is also
marked as "used for long-term reference", the reference frame or
complementary reference field pair is also marked as "used for long-term
reference" and assigned LongTermFrameIdx equal to long_term_frame_idx.
This fixes decoding of MR9_BT_B in strict mode.
https://bugs.freedesktop.org/show_bug.cgi?id=64624https://bugzilla.gnome.org/show_bug.cgi?id=724518
Request the correct chroma format for decoding grayscale streams.
i.e. make lookups of the VA chroma format more generic, thus possibly
supporting more formats in the future.
This means that, if a VA driver doesn't support grayscale formats,
it is now going to fail. We cannot safely assume that maybe grayscale
was implemented on top of some YUV 4:2:0 with the chroma components
all set to 0x80.
Fix reference picture marking process with memory_management_control_op
set to 3 and 6, i.e. assign LongTermFrameIdx to a short-term reference
picture, or the current picture.
This fixes decoding of FRExt_MMCO4_Sony_B.
https://bugs.freedesktop.org/show_bug.cgi?id=64624https://bugzilla.gnome.org/show_bug.cgi?id=724518
[squashed, edited to use GST_VAAPI_PICTURE_IS_COMPLETE() macro]
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
The initialization of reference picture lists (8.2.4.2) applies to all
slices. So, the RefPicList0/1 lists need to be constructed prior to
each slice submission to the HW decoder.
This fixes decoding of video sequences where frames are encoded with
multiple slices of different types, e.g. 4 slices in this order I, P,
I, and P. More precisely, CABAST3_Sony_E and CABASTBR3_Sony_B.
https://bugzilla.gnome.org/show_bug.cgi?id=724518
When NAL units of type 13 (SPS extension) or type 19 (auxiliary slice)
are present in a video, decoders shall perform the (optional) decoding
process specified for these NAL units or shall ignore them (7.4.1).
Implement option 2 (skip) for now, as alpha composition is not
supported yet during the decoding process.
This fixes decoding of the primary coded video in alphaconformanceG.
https://bugzilla.gnome.org/show_bug.cgi?id=703928https://bugzilla.gnome.org/show_bug.cgi?id=728869https://bugzilla.gnome.org/show_bug.cgi?id=724518
[skip NAL units earlier, i.e. at parsing time]
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
When MVC slice NAL units (coded slice extension and prefix NAL) are
present, the number of NAL header bytes is 3, not 1 as usual.
Signed-off-by: Li Xiaowei <xiaowei.a.li@intel.com>
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
At the time the first VCL NAL unit of a primary coded picture is found,
and if that NAL unit was parsed to be an SPS or PPS, then the entries
in the parser may have been overriden. This means that, when the picture
is to be decoded, slice_hdr->pps could point to an invalid (the next)
PPS entry.
So, one way to solve this problem is to not use the parser PPS and
SPS info but rather maintain our own activation chain in the decoder.
https://bugzilla.gnome.org/show_bug.cgi?id=724519https://bugzilla.gnome.org/show_bug.cgi?id=724518
Retain the SEI messages that were parsed from the access unit until we
have completely decoded the current frame. This is done so that we can
peek at that data whenever necessary during decoding. e.g. for exposing
3D stereoscopic information at a later stage.
Fix support for grayscale encoded video clips, and possibly others if
the underlying driver supports the non-YUV 4:2:0 formats. i.e. defer
the decision that a surface with the desired chroma format is not
supported to the actual VA driver implementation.
https://bugzilla.gnome.org/show_bug.cgi?id=728144
Don't force allocation of VA surfaces in YUV 4:2:0 format. Rather, allow
for the upper layer to specify the desired chroma type. If the chroma
type field is not set (or yields zero), then YUV 4:2:0 format is used
by default.
