In HEVC, P and B definitions are different from AVC: P frames have
just one reference list and so 1 MV, while B frames have two reference
lists and so 2 MVs. No matter B or P, ist reference lists can contain
forward/backward reference. So P and B can both have bi-directions
dependency, the difference is just their reference list
number (i.e. MV number). This is different from the AVC.
The *low delay b mode* refers to a special HEVC mode, in which the
stream just contain I and B frames, without P frames, and all B frames
only have forward direction dependencies (i.e. all inter frames have 2
reference lists but no backward reference in both lists). This is
similar to AVC I/P mode, but changing the P to the forward dependent
B.
The `low-delay-b` property is now just used to simply convert all P
frames to B frames when driver does not support P frames (so both
reference lists have the same references frames). This is a little
different from the meaning of low delay b mode (the two ref lists may
have the different reference frames). And the driver now can report
whether it supports P frames correctly, so there is no need to use
this property and deprecate it.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer-vaapi/-/merge_requests/284>
The vaCreateContext do not need to specify the surfaces for the
context creation now. So we do not need to bind any surface to the
context anymore. Surfaces should be the resource belong to display
and just be used in encoder/decoder context.
The previous manner has big limitation for decoder. The context's
surface number is decided by dpb size. All the surfaces in dpb will
be attached to a gstbuffer and be pushed to down stream, and the
decoder need to wait down stream free the surface and go on if not
enough surface available. For more and more use cases, this causes
deadlock. For example,
gst-launch-1.0 filesrc location=a.h264 ! h264parse ! vaapih264dec
! x264enc ! filesink location=./output.h264
will cause deadlock and make the whole pipeline hang.
the x264enc encoder need to cache more than dpb size surfaces.
The best solution is seperating the surfaces number and the dpb size.
dpb and dpb size shoule be virtual concepts maintained by the decoder.
And let the surfaces_pool in context maintain the re-use of all surfaces.
For encoder, the situation is better, all the surfaces are just used
as reference frame and no need to be pushed to down stream. We can
just reserve and set the capacity of the surfaces_pool to meet the
request.
Fix: #147Fix: #88
Co-Author: Víctor Manuel Jáquez Leal <vjaquez@igalia.com>
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer-vaapi/-/merge_requests/353>
Registering only stream's DBP size number of surfaces for decoding VA
surfaces brings issues for certain streams. This change register all
possible number of reference surfaces in a stream, which is 16.
Fixes: #94
The h265 encoder just support tune mode:
(0): none - None
(3): low-power - Low power mode
So, no need to check and set the high compression parameters.
And by the way, the current ensure_tuning_high_compression manner
of choosing the hightest profile idc as the best compression profile
is not correct. Unlike h264, in h265 the higher profile idc number
does not mean it has more compression tools, and so it has better
compression performance. It may even be un-compatible with the lower
profile idc. For example, the SCREEN_CONTENT_CODING profile with idc
9 is not compatible with 3D_MAIN profile with idc 8.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer-vaapi/-/merge_requests/348>
Right now, all outputs are handled. The means that the registry object for
all but the last are leaked. As a result the sizes are not used correctly.
With two outputs, at first the mode and physical size of the second output
are used. If the first output changes the mode, then the physical size of
the second output is used in combination with the resolution of the first
output. The resulting pixel aspect ratio is incorrect.
There seems to be no way to determine on which output the window is shown,
so just use the first one to get consistent results.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer-vaapi/-/merge_requests/341>
These properties are used for support of tile encoding. We just
support uniform mode of tile encoding, that is, separating picture
equally by (num-tile-cols X num-tile-rows).
According to HEVC spec A1, the max number of tiles in column is 20
and in rows is 22, so add two constant definitions.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer-vaapi/-/merge_requests/294>
We need to use the chroma_format_idc as the index for getting the
SubWidthC and SubHeightC values as the spec 6.1(table 6-1) defines.
The wrong SubWidthC or SubHeightC make us calculate a wrong right
or bottom offset for crop size and generate garbage in output.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer-vaapi/-/merge_requests/334>
Since bufferproxy and surface are not referenced circularly, there's
no need to keep, in the buffer proxy, a reference to the GstMemory
where it is held. This patch removes that handling.
The bufferproxy may reference the surface and the surface may also
reference the bufferproxy, producing a circular reference, which might
lead to serious resource leak problems.
Now make the relationship clearer, the bufferproxy's references is
transfered to surface, while bufferproxy just keeps the surface's
address without increasing its reference count.
The surface can be created through a bufferproxy like in
gst_vaapi_surface_new_with_dma_buf_handle(), and the surface might
get its bufferproxy via gst_vaapi_surface_get_dma_buf_handle(). In
both cases the surface holds a bufferproxy's reference.
This fixed segfault when running the pipeline below with iHD driver
(commit efe5e9a) on ICL
gst-launch-1.0 videotestsrc ! vaapivp9enc tune=low-power ! vaapivp9dec ! \
fakesink
In hevc, we can consider the -intra profile a subset of the none
-intra profile. The -intra profiles just contain I frames and we
definitely can use the none -intra profiles's context to decode
them.
Signed-off-by: Víctor Manuel Jáquez Leal <vjaquez@igalia.com>
This is a workaround for intel-media-driver bug
https://github.com/intel/media-driver/issues/865
The driver will force the RC method to CBR for HEVCe
when it parses the HRD param. Thus, any RC method
param submitted "prior" to the HRD param will be lost.
Therefore, VBR, ICQ and QVBR for HEVCe can't be
effectively enabled if the RC method param "precedes"
the HRD param.
To work around this issue, set the HRD param before
the RC method param so the driver will parse the RC
method param "after" the HRD param.
Afaict, other codecs in the driver (and other drivers)
do not appear to be dependent on the order of HRD and
RC param submission.
GstVaapiPixmap is an abstract base class which only implementation
were GstVaapiPixmapX11. This class were used for a special type of
rendering in the tests apps, utterly unrelated in GStreamer.
Since gstreamer-vaapi is no longer a general-user wrapper for VA-API
we should remove this unused API.
This removal drops libxrender dependency.
If the dependent_slice_segment_flag is true, most slice info derived from last slice.
So we need check the slice type after we call populate_dependent_slice_hdr
Since commit 32bf6f1e GLTextureUpload is broken because i965
doesn't report properly RGBA support. It could be possible to use RGBx
but GLTextureUpload only regotiates RGBA.
The simplest fix to this regression is adding synthetically the RGBA
format in the internal format map.
Instead of break at the fist foud quirk in the table, iterate all over
so it would be feasible to add several quirks for one driver per
element in array.
The intel-media-driver (iHD) can't convert output color
primaries when doing YUV to/from RGB CSC. Thus, we must
keep the output color primaries the same as the input
color primaries for this case.
fixes#238
When creating surfaces it is possible to pass to VA hints of its usage,
so the driver may do some optimizations.
This commit adds the handling of encoding/decoding hints.
I've just discovered iHD driver in Skylake doesn't have VideoProc
entry point, hence, in this platform, when vaapioverlay is tried to be
registered, critical warnings are raised because blend doesn't have a
display assigned.
As it is possible to have drivers without EntryPointVideoProc it is
required to handle it gracefully. This patch does that: only tries to
register vaapioverlay if the testing display has VPP and finalize()
vmethods, in filter and blend, bail out if display is NULL.
Commit 1168d6d5 showed up a regression: decode_sps() stores the unit's
parser info in sps array. If that parser info comes from decoding
codec data, that parser info will have an undefined state which might
break ensure_sps().
This patch sets the parser info state, at decoding codec data, with
the internal parser state. This is similar with h264 decoder apprach.
Original-patch-by: Xu Guangxin <guangxin.xu@intel.com>
VAProcColorStandardExplicit and associated VAProcColorProperties
(primaries, transfer and matrix) are not supported until
VA-API 1.2.0.
Use VAProcColorStandardNone instead of VAProcColorStandardExplicit
if VA-API < 1.2.0.
Fixes#231
Addresses #228 on iHD side. It seems iHD can't handle
VAProcColorStandardSRGB in all situations for vpp. But
it has no problem when we specify the sRGB parameters
via VAProcColorStandardExplicit parameters.
We've always sent VA_SOURCE_RANGE_UNKNOWN to the driver.
And, the [iHD] driver essentially computes the same color
range as gstreamer when we send VA_SOURCE_RANGE_UNKNOWN for
cases were gstreamer computes it automatically. But,
if the user wants to make it explicit, we should try
to honor it.
This mechanism comes from ffmpeg vaapi implementation, where they have
their own quirks.
A specific driver is identified by a substring present in the vendor
string. If that substring is found, a set of bitwise flags are store.
These flags can be accessed through the function
gst_vaapi_display_has_driver_quirks().
The purpose for this first quirks is to disable the put image try for
AMD Gallium driver (see [1]).
1. https://gitlab.freedesktop.org/gstreamer/gstreamer-vaapi/merge_requests/72
This commit tries to centralize the selection of vaCreateSurfaces
version, instead of having fallbacks everywhere.
These fallbacks are hacks, added because new drivers use the latest
version of vaCreateSurfaces (with surface attributes) [1], meanwhile
old drivers (or profiles as JPEG decoder in i965) might rather use the
old version.
In order to select which method, there's detected hack: each config
context has a list of valid formats, in the case of JPEG decoder the
list only contains "rare" 4:2:2 formats (ICM3, GRAY8) which aren't
handled correctly by the current gstreamer-vaapi code [2].
The hack consist in identify if the format list contains an arbitrary
preferred format (which is suposedly well supported by
gstreamer-vaapi, mostly NV12). If no prefered colour format is found,
the the old version of vaCreateSurfaces is used, and the surfaces wil
be mapped into a image with their own color format.
1. https://bugzilla.gnome.org/show_bug.cgi?id=797143
2. https://bugzilla.gnome.org/show_bug.cgi?id=797222
When baseline-as-constrained is set, the decoder will expose support
for baseline decoding and assume that the baseline content is
constrained-baseline. This can be handy to decode streams in hardware
that would otherwise not be possible to decode. A lot of baseline
content is in fact constrained.
The formats array is always created, in order to keep the logic and
to avoid broken caps, if this formats array doesn't contain any
elements, it has to be unref and the function should return NULL.
When the tune is NONE, we now can choose entrypoint freely. So the
GST_VAAPI_ENCODER_TUNE macro may not return the correct current
entrypoint.
We also delay CTU size calculation after entrypoint has been decided.
FEI encoders are not actively mantained neither tested, and it is
using infrastructure that is changing and FEI is stopping this
effort.
Also it is required to rethink how FEI can be used in GStreamer.
Instead of using a parent structure that has to be derived by API
consumers, this change propse a simplification by using the common
pattern of GTK of passing a function pointer and user data which will
be passed as its parameter. That user data contains the state and the
function will be called to update that state.
