Often, users will need to scale inputs (e.g.
with vaapipostproc) before they are submitted
to the vaapioverlay. However, this results in
multiple VPP passes/operations in the pipeline
which creates unnecessary process overhead.
This change allows for inputs to be submitted
at original scale to vaapioverlay with per-sinkpad
scale dimensions specified so they can be scaled
and blended/composited in a single VPP pass/operation
to avoid the unnecessary process overhead.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/1380>
Don't set VAAPI vpp blend flags if alpha == 1.0,
i.e. fully opaque. This can avoid extra processing
overhead on some drivers that apply blending
unconditionally when flags are present, even if the
end result is the same without blend flags (i.e. all
opaque alpha channels).
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/1380>
Note that AYUV and AYUV64 formats will be used to expand format
support, especially some packed YUV formats (e.g., Y410, YUY2)
are common DXGI formats used for hardware decoder/encoder on Windows
but those formats cannot be used as a render target. We need to handle
them differently without pixel shader help, using compute shader
for example.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/1699>
The libjpeg-turbo internal state might not be correctly initialized for
the first frame in a stream, pull the frame stride from gstreamer frame
metadata instead, which is correct even for the first frame, and which
makes this code consistent with the surrounding lines.
Fixes: e6d83d8f96 ("jpegdec: Support libjpeg-turbo colorspace conversion")
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/1687>
It is imperative that the libjpeg-turbo state is properly initialized
before jpeg_start_decompress() is called. Make sure cinfo.out_color_space
and cinfo.raw_data_out are set to their final values matching their peer
caps before calling jpeg_start_decompress().
Fixes: e6d83d8f96 ("jpegdec: Support libjpeg-turbo colorspace conversion")
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/1687>
Pull out peer caps checking code into gst_jpeg_turbo_parse_ext_fmt_convert().
This code is used by libjpeg-turbo extras to determine whether peer is capable
of handling buffers into which libjpeg-turbo can directly decode data. This
kind of check must be performed before jpeg_start_decompress() is called in
gst_jpeg_dec_prepare_decode() as well as in gst_jpeg_dec_negotiate(), hence
the common code.
This commit does modify the code a little to make it easier to call from both
call sites without much duplication, hence the extra `if (*clrspc)` test.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/1687>
This reverts commit 2aa2477208.
The commit is completely wrong, libjpeg-turbo is perfectly capable
of decoding I420 (YUV) to RGB. The test case provided alongside the
aforementioned commit passes without this revert because it decodes
image of JCS_YCrCb color space, so the new `if (clrspc == JCS_RGB)`
condition is false on that image, and the libjpeg-turbo decoding
does not get used. The real bug is hidden by that commit.
The real problem is in the call order of gst_jpeg_dec_prepare_decode()
and gst_jpeg_dec_negotiate(). The gst_jpeg_dec_prepare_decode() calls
jpeg_start_decompress() which sets up internal state of the libjpeg,
however, neither cinfo.out_color_space nor cinfo.raw_data_out are
set correctly yet. Those two are set up in gst_jpeg_dec_negotiate()
which is called a bit later. Therefore, the real fix is the set up
cinfo.out_color_space and cinfo.raw_data_out before calling
jpeg_start_decompress(). This is however a separate patch.
Fixes: 2aa2477208 ("jpegdec: only allow conversions from RGB")
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/1687>
Remove all the d3d11 and dxgi header version dependent ifdef
and bump the minimum requirement to d3d11_4.h and dxgi1_6.h.
We are already failing support old Visual Studio (Windows SDK actually)
such as Visual Studio 2015. Note that our MinGW toolchain satisfies
the requirement.
From runtime point of view, this change should be fine since
we are checking OS version with IUnknown::QueryInterface()
everywhere in order to check API availability
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/1684>
We make all MSDK encoders declare "memory:VAMemory" feature. Then
the pipeline such as:
gst-launch-1.0 -vf filesrc location=xxx.h264 ! h264parse ! \
vah264dec ! msdkh265enc ! fakesink
will choose VA memory caps between the VA decoder and MSDK encoder.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/1087>
The MSDK encoder's query function is not set and it just forwards
all query to its base class. We now need to answer the context
query correctly. Other VA plugins need to query the VA display.
By the way, the current query of "gst.msdk.Context" is also missing.
The other MSDK elements must depend on the bin's context message(
sent in context_propagate()) to set their MsdkContext correctly.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/1087>
The MSDK VPP's query function is not set and it just forwards
all query to its base class. We now need to answer the context
query correctly. Other VA plugins need to query the VA display.
By the way, the current query of "gst.msdk.Context" is also missing.
The other MSDK elements must depend on the bin's context message(
sent in context_propagate()) to set their MsdkContext correctly.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/1087>
The MSDK decoder's query function is not set and it just forwards
all query to its base class. We now need to answer the context
query correctly. Other VA plugins need to query the VA display.
By the way, the current query of "gst.msdk.Context" is also missing.
The other MSDK elements must depend on the bin's context message(
sent in context_propagate()) to set their MsdkContext correctly.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/1087>
We now can use the gst va lib's display to create our MSDK context,
and use its helper functions to simplify our code. The improved logic
is like this:
1. Every MSDK element should use gst_msdk_context_find() to find a MSDK
context from neighbour. If valid, reuse it.
2. Use gst_msdk_ensure_new_context(). It will first query neighbours
about the GstVaDisplay, if found(e.g. some VA element is connected),
use gst_msdk_context_from_external_display() to create a MSDK context.
3. Then, creating the MSDK context from scratch. It creates both the
display and MSDK context.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/1087>
The VA display object from VA lib is a common defined object. which
contain the whole display things. It is easier to use, and more important,
we can share it with the other VA plugins and keep all the VA related
plugins working on the same GPU device.
We also delete the useless gst_msdk_context_get_fd() API.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/1087>
The current manner for deciding the new temporal unit is based on
temporal delimiter(TD) OBU. We only start a new temporal unit when
the TD comes.
But some streams do not have TD at all, which makes the output "TU"
alignment fail to work. We now add check based on the relationship
between the different layers and it can successfully judge the TU edge.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/1634>
Some streams may have problematic OBUs at the beginning, which causes
the parse fail to detect the alignment and return error. For example,
there may be verbose OBUs before a valid sequence, which should be
discarded until we meet a valid sequence. We should let the parse
continue when we meet such cases, rather than just return error.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/1634>
