The return value of vaGetConfigAttributes() of VAConfigAttribRTFormat
is a bitwise flag with *all* the supported chroma types.
Previously it was assumed that the return value was a single value,
thus when returning the GST_VAAPI_CHROMA_TYPE_XXX the code was a
simple case. But it is wrong.
This patch changes the case block with a sequence of ifs testing the
bitwise. For now we assume a "priority" list in the testing sequence.
gst_vaapi_profile_get_name() returns a proper name for
GstCaps. Nonetheless, there are many profiles which don't have a name
representation for that realm.
gst_vaapi_profile_get_va_name() returns the name of the profile
according to its VAProfile name.
This new funtion is used in the encoder error message.
Now that vaapipostproc can possible handle video-direction, it
should also handle the image-orientation event from upstream if
video-direction property is set to auto.
If the video direction is unsupported by the driver, an element
warning is posted in the bus to notify the application.
gst_vaapi_enum_type_get_nick() was added in the library thus it can
be used elsewhere. It retrives the nick from an enum gtype.
This patch locks the display before querying the pipeline caps and
stores the mirror and rotation capabilities, thus they are not queried
every time the video direction is set.
Commit 0afc8131 introduced a regression and only NV12 format were
admitted, failing in any other valid color format.
This patch sets the profile to GST_VAAPI_PROFILE_VP9_0 by default.
When set multiple settings of color balance like hue, saturation,
brightness and contrast for vaapipostproc, they should be set as
parameters of color balance filter, at the same color balance
filter calling.
Otherwise, multiple color balance filter calling will cause
previous setting get reset by the last calling with default value.
Fixes#182.
Signed-off-by: Yan Wang <yan.wang@linux.intel.com>
gst_vaapi_video_pool_reserve_unlocked() hit an undocumented behavoir
because it locks twice the same mutex.
Also, n had different meanings in the current code: as an increase
value and as a new total of allocated surfaces.
This patche removes the undocumented behavoir (usually a deadlock) and
fixes the meaning of n as the new total of allocated surfaces.
Signed-off-by: Víctor Manuel Jáquez Leal <vjaquez@igalia.com>
ICQ is Intelligent Constant Quality. It will use the initial QP
vaule of icq-quality-factor to adjust QP at MB level intelligently
to improve subjective quality.
QVBR is Quality defined VBR. It will use qvbr-quality-factor to
adjust QP for each MB to get enough quality picture without waste
of bits.
The value return from GST_VAAPI_RATECONTROL_MASK will be used by
GST_VAAPI_POPCOUNT32 as its inpput. GST_VAAPI_POPCOUNT32 can only
deal with unsigned int. Otherwise there may be an error of out of
range of integer if we define few more rate-control mode.
If display size is smaller than current frame size, then the crop size
will be set as display size, which either crashes the pipeline or the
output MD5 does not match. Rather it should use the actual decoded size.
This patch removes the cropping set. For rendering we can use aspect
ratio to set display size.
Fixes#175
Signed-off-by: Wang Zhanjun <zhanjunx.wang@intel.com>
Signed-off-by: Xu Guangxin <guangxin.xu@intel.com>
For some main-10 stream, sometime the luma is 8 bits while chrome is more
than 8 bits, which cause using the wrong NV12 surface as the render target
and decoding error.
Fix#176
When create vaapi surface, it is better to use the chroma type get
from jpeg file instead of using fixed 420 format. And the correct
chroma type can be determined by horizontal_factor/vertical_factor
flags that get from jpegparse.
The old mapping values return by gst_vaapi_utils_h265_get_profile_idc is
wrong, though GST_H265_PROFILE_IDC_MAIN and GST_H265_PROFILE_IDC_MAIN_10
happened to be the correct value.
We only support Annex A profile_idc (1-4).
When input frame's formate changes, vp{8,9} encoders don't reset their frame
counter, hence the newly created frame could become a P-frame, leading to some
major troubles (sigabrt in libdrm in case of vp9). This patch adds some frame
prediction-related reset logic to the `flush' methods of GstVaapiEncoderVP8 and
GstVaapiEncoderVP9 implementations.
There are many profile levels that can support
more than 102400 kbps. Thus, increase the max
allowed bitrate property value from 102400 kbps
to 2048000 kbps (same as msdk encoder plugins).
Adds vpp mirroring support to vaapipostproc. Use
property video-direction. Valid values are identity,
horiz or vert. Default is identity (no mirror).
Closes#89
v2: Use GstVideoOrientationMethod enum
v3: Don't warn for VA_MIRROR_NONE.
Use GST_TYPE_VIDEO_ORIENTATION_METHOD type.
v4: Query VAAPI caps when setting mirror value
instead of during per-frame processing.
v5: Return TRUE in warning cases when setting mirror value.
When enable low delay B, the reference list 1 will be same with
reference list 0, so need reset the num_ref_idx_l1_active_minus1
to num_ref_idx_l0_active_minus1.
Fixes: #160
g_assert will take no effect when glib's G_DISABLE_ASSERT macro is
defined. The function inside the g_assert will take no effect and
we will fail to set the correct slice number.
Use VAPictureParameterBufferHEVCExtension&
VASliceParameterBufferHEVCExtension to pass extension setting from
some extension profile clips which may include these information.
The hevc extension setting only supported after libva release 2.2.0
(API 1.2.0).
Allow users to set the target-percentage for
variable rate controls. The default value is
70 (as hard-coded prior).
v2: minimum allowed value changed from 0 to 1
v3: target-percentage unchanged if CBR used
Resolves#129
The advanced trellis algorithm is supported in VA driver. We add
its support as a property named "trellis" of encoder.
It only works for H264 now, should be more in future.
Can not find a suitable chrome_type for this GST_VIDEO_FORMAT_RGB
packed 24 format. Just use GST_VAAPI_CHROMA_TYPE_RGB32 as its chrome
type. This kind of surface will just be created by new API with fourcc
and no old style chrome based creation is available.
fixes: #151
Coverity scan bug:
If the function actually returns a null value, a null pointer
dereference will occur.
In gst_vaapi_encoder_ensure_param_roi_regions(): Return value of
function which returns null is dereferenced without checking
The decoder's surface chroma type should depend on the bit depth
of VP9's parser. For 10bits VP9 stream, we need to use P10LE kind
10 bits surface as the decoder result.
Fixes#155
This patch fixes a regression from commit 5b1fe9c6.
max_qp, in rate control configuration, appeared in libva release
2.1 (API 1.1), thus it is required to guard the VA API version.
Fixes: #150
string_of_va_chroma_format() gets a wrong string format description.
For example, the YUV420_10BPP get a string of 0x00000100 as output.
It's because VA_RT_FORMAT_xxx values are macro definitions. And
STRINGIFY(VA_RT_FORMAT_xxx) will expand to its real value
0x00000XXX.
To avoid the macro evaluation, it is changed to show only the color
format without VA_RT_FORMAT_ prefix.
The current manner can not recognize the correct level and always
set the tier to main. Need to add frame rate check to recognize
levels such as 4.1, 6.2, etc. We also add a logic to check main
and high tier based on bitrate.
Fixes: #145
media-driver currently fails to set a correct value of max_qp when
min_qp is different to zero, in CBR and VBR mode, generating full
quality frames, thus unexpected huge output.
This patch sets max_qp to an arbitrary value to avoid this output
temporary.
Fixes: #144
Currently the minimal value for either min_qp and init_qp are 1,
but VA documentation specifiy that zero is also valid and means
to ignore the quantiser.
The default value is not changed though to avoid behaivor changes
to users.
GValueArray is deprecated. Use GstValueArray instead.
gst_param_spec_array can be deserialized from command line using:
vaapih264enc view-ids="<(uint)40,(uint)100>" num-views=2
While the g_param_spec_value_array() can not, and always get
error: "gst_value_deserialize_g_value_array: unimplemented"
Also fixed an out-of-range bug.
The order in gst_vaapi_enc_picture_encode when encoding one
picture is not very correct. The misc parameters are set before
the picture parameters. Some of the misc parameters such as
ROI may change the current picture parameters. But the later
setting of picture parameter will re-init all picture related
parameters and clear the previous setting. The right order
should be picture parameter first and then misc parameters.
Signed-off-by: He Junyan <junyan.he@hotmail.com>
This code is just confused. It's asking for at least as many bits of
(z-axis) depth as the root window has bits of (color) depth. For rgb565
or rgb888 this is harmless, but at 10 bits per channel this demands a
30-bit or deeper Z buffer. While some hardware could in principle do a
32-bit Z buffer, Mesa does not expose such fbconfigs (at least on Intel
and AMD).
We're not actually using the Z buffer, so just stop asking for one.
[wl_shell] is officially [deprecated], so provide support for the
XDG-shell protocol should be provided by all desktop-like compositors.
(In case they don't, we can of course fall back to wl_shell).
Note that the XML file is directly provided by the `wayland-protocols`
dependency and generates the protocol marshalling code.
[wl_shell]: https://people.freedesktop.org/~whot/wayland-doxygen/wayland/Client/group__iface__wl__shell.html
[deprecated]: 698dde1958
native-id property is problematic since the variable that stores it is
gsize, which is platform specific, and in some is bigger than unsigned
long, and there are not way to handle gsize properties.
Also, GST_VAAPI_ID_INVALID is defined in gsize terms, and we would
like to keep using it for this scope.
This patch removes the native-id property and set it manually in
gst_vaapi_window_new_internal().
gst_vaapi_encoder_put_frame() and gst_vaapi_encoder_flush() duplicates
the same code segment where the coded buffer is created, the picture
encoded on it and pushed to the async queue.
The function gst_vaapi_encoder_encode_and_queue() refactor this.
In order to flush the pending pictures, a new internal encoder vmethod
is used: get_pending_reordered()
This method follows an iterator pattern which will return the next
picture to encode and push.
The base encoder will call this function in a loop when flush() is called.
For now, only H.264 and H.265 encoders implement this flushing mechanism.
This is another step in the gobjectification of the internal library
of gstreamer-vaapi. Now it is the turn of GstVaapiWindow and its
derivates.
The idea is to minimize the changeset keeping the same design as
much as possible.
GstVaapiWindow is defined as an abstract class with two properties:
the GstVaapiDisplay and the native ID. Thus, many of the
GstVaapiObject macros were copied as GstVaapiWindow macros.
The function gst_vaapi_window_new_internal() is kept as a decorator
of for calling gst_vaapi_window_create() and the possibility of
failure.
The descendant classes, such as glx, still use the private
structures, but through the gobject mechanism.
Add 444 10bit yuv format Y410, which can be used to decode
main-444 10bit streams. Currently, this feature is only
supported by media-driver in Icelake.
Add more kinds of chrometype which will be used to describe
new video formats. Sync it with 1.4.0 version header file.
Alse delete useless GST_VAAPI_CHROMA_TYPE_YUV410 chrome type.
Signed-off-by: He Junyan <junyan.he@hotmail.com>
We lack some video format because endianness declare.
The video format should not directly relate to endianness. For example,
ARGB on big endian should not be simplely seen as BGRA on little endian
machine. We should provide endianess convert or format convert help
functions if endianness does not match.
https://gitlab.freedesktop.org/gstreamer/gstreamer-vaapi/issues/112
Signed-off-by: Junyan He <junyan.he@intel.com>
Add 422 10bit yuv format Y210, which can be used to decode
main-10-422 10bit streams. Currently, this feature is only
supported by media-driver in Icelake.
https://bugzilla.gnome.org/show_bug.cgi?id=797264
VA_ROI_RC_QP_DELTA_SUPPORT return value will be ignored when the
rate control mode is set as CQP. In CQP mode, it shouldn't check
roi_rc_qp_delta_support return value from driver backend.
https://bugzilla.gnome.org/show_bug.cgi?id=797087
In gst_vaapi_context_reset(), if the context has to be destroyed, make
sure to create it first before allocating its associated surfaces.
This patch fixes a regression introduced in commit 82872f4 because
the formats available in the current context now are ensured before
creating the context's surfaces.
https://bugzilla.gnome.org/show_bug.cgi?id=797277
Removed exposed macros GST_VAAPI_OBJECT_DISPLAY() and
GST_VAAPI_OBJECT_ID() to plugins, keeping them only for internal
library usage.
The purpose is readability.
https://bugzilla.gnome.org/show_bug.cgi?id=797139
When running several vaapi elements at the concurrently, at
initialization, there is a race condition when extractin the avaible
formats for images and subpictures.
This patch add a lock when the those arrays are filled.
https://bugzilla.gnome.org/show_bug.cgi?id=797039
Low delay B frame provide the function of transforming
P frame into low delay B frame which frame type is B, but
only reference predictive frames. This can be used when P
frame unsupported. Especially for P and B both unsupported,
in this case, I and low delay B frame can be encoded in a
stream.
https://bugzilla.gnome.org/show_bug.cgi?id=796984
In some cases, the found_picture ended up being evicted and freed, which
would lead to a use after free when accessing picture->base.poc. In this
fix, we take a ref on the picture before calling dpb_evict.
https://bugzilla.gnome.org/show_bug.cgi?id=787124
Similarly to previous patch, we have no error concealment. As a side
effect, it's better to skip slices with missing references then passing
NULL pointers to the accelerator. Passing NULL pointer would lead to
major visual artifact, a behaviour that is likely undefined.
https://bugzilla.gnome.org/show_bug.cgi?id=787124
Usually in case of MVC decoding, dpb size is increasedi if subset sps.
That's why it resets context without this patch.
But for some media it doesn't increase dpb size. Even in this case we
should reset context to deal with MVC decoding.
Otherwise, it leads to assert.
https://bugzilla.gnome.org/show_bug.cgi?id=787124
If the unit could not be parsed, just skip this nal and keep parsing
what is left in the adapter. We need to flush the broken unit in the
decoder specific parser because the generic code does not know about
units boundary. This increases error resilliance.
Before this, the broken unit would stay in the adapter and EOS would be
returned. Which stopped the streaming. Just removing the EOS would have
lead to the adapter size growing indefinitely.
https://bugzilla.gnome.org/show_bug.cgi?id=796863
h264 log messages were logged in default category because a regression
in code. This patch renable the usage of vaapi logging category.
This regression was introduced in commit 7c365bdd.
This patch chains up failure to executing the modification process. The
end result is that we now fail decoding the slice if this process fails.
This avoid sending a corrupted state to the accelerator. In some special
cases, this could lead to unrecoverable errors.
https://bugzilla.gnome.org/show_bug.cgi?id=796832
gst_adapter_masked_scan_uint32_peek() asserts if size is 0. Don't
try and scan in that case. This fixes assertion that would some times
happen when the stream is corrupted.
https://bugzilla.gnome.org/show_bug.cgi?id=796832
This change is due a problem encoding JPEGs with Intel's
media-driver: green/black image when playback jpeg
This patch sets component identifier and quantization table
destination selector in frame header to support packing headers
by Intel's media-driver that does not accept packed header
in AP level.
https://bugzilla.gnome.org/show_bug.cgi?id=796705
The decoder was trying to detect earlier that a field was lost base
on guessing the parity order. This breaks in streams were the parity
order changes.
This patch reverts the field order prediction code added by commit
8dd93e9c8.
https://bugzilla.gnome.org/show_bug.cgi?id=796169
This ensure that we always have sentinels set in the reference
pictures arrays. The code wasn't unsafe, this simply improve the
tracing, so instead of printing 32 lines of zeros, va tracer
prints proper empty lists.
https://bugzilla.gnome.org/show_bug.cgi?id=796169
Gobjectification for GstVaapiDisplay was almost done by the commit 185da3d1.
But still something breaking GObject code convention remains, which is
calling gst_vaapi_display_new() in each decendants.
This patch replaces it with gst_vaapi_display_config(), defined in private
header.
https://bugzilla.gnome.org/show_bug.cgi?id=796470
The function name was gst_vaapi_display_create_unlocked(), nonetheless
it wasn't called unlocked. In order to keep the semantics this patch
renames the gst_vaapi_display_create_unlocked() as
gst_vaapi_display_create(), removing the previous function
gst_vaapi_display_create().
https://bugzilla.gnome.org/show_bug.cgi?id=796470
This is, practically, a revert of commit dcf135e2.
The parent logic is useful for the EGL display, which is a decorator
of the real windowing subsystem (X11 or Wayland). Thus it is avoided
calling vaInitialize() and vaTerminate() twice.
https://bugzilla.gnome.org/show_bug.cgi?id=795391
gst_vaapi_display_egl_new_with_native_display() has been broken since
it wasn't used.
Currently it's needed to call this API to create a display providing
the EGL display, so it could avoid duplicated calls to the native
display (eg. eglTerminate).
Signed-off-by: Victor Jaquez <vjaquez@igalia.com>
https://bugzilla.gnome.org/show_bug.cgi?id=795391
This change is due a problem decoding JPEGs with Intel's media-driver:
no image was generated.
This patch relases the VA buffers after vaEndPicture() is called,
and not before (after vaRenderPicture()).
https://bugzilla.gnome.org/show_bug.cgi?id=796505
The specification says,
"log2_max_pic_order_cnt_lsb_minus4 shall be in the range of 0 to 12, inclusive."
This patch changes the upper limit from 6 to 12.
https://bugzilla.gnome.org/show_bug.cgi?id=796179
This removes hard-coded checks on the parent subsystem of potential DRM
devices. These checks were set to exlude devices that do not originate
from the PCI bus, which is only a valid approach on x86 devices.
Other devices may have a DRM device originating from the platform
subsystem, so the checks that were previously restricted to PCI are
extended to cover platform devices as well.
https://bugzilla.gnome.org/show_bug.cgi?id=794840
Signed-off-by: Paul Kocialkowski <paul.kocialkowski@bootlin.com>
Otherwise the following poll may not return for an arbitrary amount of
time. This can happen if another wayland event queue has flushed and read
our events.
https://bugzilla.gnome.org/show_bug.cgi?id=795224
pic_width_in_luma_samples/pic_height_in_luma_samples can be 16-bit rounded
instead of 32-bit.
In addition, codedbuf_size must be calculated according to this change.
https://bugzilla.gnome.org/show_bug.cgi?id=753229
Even if it is the h264parse fault or bad video file, vaapih264dec
should set a proper value for VAPictureParameterBufferH264.num_ref_frames
as the driver might use it.
Also see "info.ref_frames = dpb_size;" in
gstvaapidecoder_h264.c::ensure_context
https://bugzilla.gnome.org/show_bug.cgi?id=793836
Check input buffers for ROI metas and pass them to VA. Also added a
new "default-roi-delta-qp" property in order to tell the encoder what
delta QP should be applied to ROI by default.
Enabled it for H264 and H265 encoders.
https://bugzilla.gnome.org/show_bug.cgi?id=768248
According to Debian package auto-building, uintptr_t is not an
unsigned long in i386 arch, raising an "incompatible pointer type"
error.
This patch adds a casting for compiler's satisfaction in i386.
Increased max values of periodic key frame for h26x codecs.
This allow more fine tunning of encoder that in certian scenario
want higher periodic key frame.
For example: it doesn't want a key frame each 10 seconds but
each 120 seconds.
https://bugzilla.gnome.org/show_bug.cgi?id=786320
eglGetDisplay() is currently broken in Mesa for Wayland. Also using
eglGetDisplay() is rather fragile, and it is recommended to use
eglGetPlatformDisplay() when possible.
In order to do that, this patch uses the helper in GstGL. If
gstreamer-vaapi is not compiled with GstGL support, eglGetDisplay()
will be used.
https://bugzilla.gnome.org/show_bug.cgi?id=790493
Frames are encoded as different layers. Frame in a particular
layer will use pictures in lower or same layer as references.
Which means decoder can drop the frames in upper layer but still
decode lower layer frames.
B-frames, except the one in top most layer, are reference frames.
All the base layer frames are I or P.
eg: with 3 temporal layers
T3: B1 B3 B5 B7
T2: B2 B6
T1: I0 P4 P8
T1, T2, T3: Temporal Layers
P1...Pn: P-Frames:
B1...Bn: B-frames:
T1: I0->P4 , P4->P8 etc..
T2: I0--> B2 <-- P4
T3: I0--> B1 <-- B2, B2 --> B3 <-- P4
Signed-off-by: Sreerenj Balachandran <sreerenj.balachandran@intel.com>
https://bugzilla.gnome.org/show_bug.cgi?id=788918
Frames are encoded as different layers. A frame in a particular
layer will use pictures in lower or same layer as references.
Which means decoder can drop the frames in upper layer but still
decode lower layer frames.
eg: with 3 temporal layers
T3: P1 P3 P5 P7
T2: P2 P6
T1: P0 P4 P8
T1, T2, T3: Temporal Layers
P1...pn: P-Frames:
P0->P1 , P0->P2, P2->P3, P0->P4......repeat
Signed-off-by: Sreerenj Balachandran <sreerenj.balachandran@intel.com>
https://bugzilla.gnome.org/show_bug.cgi?id=788918
The frame_num generation was not correctly implemented.
According to h264 spec, frame_num should get incremented
for each frame if previous frame is a referece frame.
For eg: IPBPB sequece should have the frame numbers 0,1,2,2,3
Signed-off-by: Sreerenj Balachandran <sreerenj.balachandran@intel.com>
https://bugzilla.gnome.org/show_bug.cgi?id=788918
Adds new property "prediction-type" to select different reference
picture selection modes like hierarchical-p, hierarchical-b etc.
Signed-off-by: Sreerenj Balachandran <sreerenj.balachandran@intel.com>
https://bugzilla.gnome.org/show_bug.cgi?id=788918
Adds some basic building blocks to ease the implementation
of hierarchical prediction modes.
-- add an utility method to find temporal level of each frame
-- define max_ref_frame count based on temporal level count
-- add temporal_level_div[] for finding temporal level each frame
to be encoded.
-- find ip_period based on temporal level count
Signed-off-by: Sreerenj Balachandran <sreerenj.balachandran@intel.com>
https://bugzilla.gnome.org/show_bug.cgi?id=788918
Adds new property "temporal-levels" to select the number of
temporal levels to be included in the encoded stream.
Signed-off-by: Sreerenj Balachandran <sreerenj.balachandran@intel.com>
https://bugzilla.gnome.org/show_bug.cgi?id=788918
We can have p-frame as non-ref and also b-frame as ref
which are not supported yet. Reference flag
is the first machinery needed for more advanced
reference picture selection modes.
Signed-off-by: Sreerenj Balachandran <sreerenj.balachandran@intel.com>
https://bugzilla.gnome.org/show_bug.cgi?id=788918
Adds a new function that changes the internal EGL display to the
current one (eglGetCurrentDisplay()) and sets the current context
too (eglGetCurrentContext()).
This new function is called by gst_vaapi_texture_egl_create() updating
the GstVaapiDisplayEGL with the current EGL display.
https://bugzilla.gnome.org/show_bug.cgi?id=773453
It is required to use the context of the calling thread when wrapping
a foreign texture. According the documentation of
GstVideoGLTextureUploadMeta:
"The caller of gst_video_gl_texture_upload_meta_upload() must
have OpenGL set up and call this from a thread where it is valid
to upload something to an OpenGL texture."
This patch updates the EGL display and context in GstVaapiDisplay
instance to the one used by te renderer that uploads the texture.
Original-patch-by: Daniel van Vugt <daniel.van.vugt@canonical.com>
https://bugzilla.gnome.org/show_bug.cgi?id=773453
Instead of silently fail to export the image if there is not available
the EGL_MESA_drm_image, log an error message. Also a code refactoring
was done.
https://bugzilla.gnome.org/show_bug.cgi?id=773453
In case of streaming, controlling bitrate dynamically for encoder might be
important to manage quality of the streaming.
This patch is to support such a scenario.
https://bugzilla.gnome.org/show_bug.cgi?id=786321
gstreamer-vaapi initializes the display by trial-and-error, thus
logging an error message if the display initialisation fails the user
may be weary of the error message in the screen, if using VA-API 1.0
This commit set the VA error log handler to GStreamer warning level
while calling vaInitialize() and set it to error after that.
https://bugzilla.gnome.org/show_bug.cgi?id=783169
GstVaapiDisplayEGL is a wrapper of another GstVaapiDisplay, either X11
or Wayland. Nonetheless it created another vaDisplay for it, instead
of using the wrapped one.
This patch enables the reuse of the wrapped vaDisplay avoiding
instantiating two.
https://bugzilla.gnome.org/show_bug.cgi?id=782212
This is a difference between the GObject API and the GstVaapi one: the
GValue passed to get a property value, in GObject has to be
initialized with g_value_init(), but in GstVaapi is has not.
In order to overcome this mismatch, this patch call g_value_init()
internally only in the passed one is not already initialized.
https://bugzilla.gnome.org/show_bug.cgi?id=788058
Shuffled some code to avoid to find the properties descriptor in the
array twice, adding the internal functions _set_property() and
_get_property().
https://bugzilla.gnome.org/show_bug.cgi?id=788058
