Our context is non-persistent, and we propagate it throughout the
pipeline. This means that if we try to reuse any gstmsdk element by
removing it from the pipeline and then re-adding it, we'll clone the
mfxSession and create a new gstmsdk context as a child of the old one
inside `gst_msdk_context_new_with_parent()`.
Normally this only allocates a few KB inside the driver, but on
Windows it seems to allocate tens of MBs which leads to linearly
increasing memory usage for each PLAYING->NULL->PLAYING state cycle
for the process. The contexts will only be freed when the pipeline
itself goes to `NULL`, which would defeat the purpose of dynamic
pipelines.
Essentially, we need to optimize the case in which the element is
removed from the pipeline and re-added and the same context is re-set
on it. To detect that case, we set the context on `old_context`, and
compare it to the new one when preparing the context. If they're the
same, we don't need to do anything.
Fixes https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/issues/946
Instead of using a proxy of `is_packetized` flag this patch
replaces it with the accessor to that flag in decoder base class,
avoiding probable mismatches.
For a skipped frame in VC1, MSDK returns the mfx surface of the reference
frame, so we have to make sure the corresponding surface for the
reference frame is not freed. In this fix, we postpone surface free because
we don't know whether a surface is referenced
Before this fix, the error is like as below:
New clock: GstSystemClock
0:00:00.181793130 23098 0x55f8a9d622d0 ERROR msdkdec
gstmsdkdec.c:622:gst_msdkdec_finish_task:<msdkvc1dec0> Couldn't find the
cached MSDK surface
Sample pipeline:
gst-launch-1.0 filesrc location=input_has_skipped_frame.wmv ! asfdemux !
vc1parse ! msdkvc1dec ! glimagesink
A post_configure virtual method is added to allow
codec subclasses to adjust the initialized parameters
after MFXVideoDECODE_DecodeHeader is called from the
gstmsdkdec::gst_msdkdec_handle_frame function.
This is useful if codecs want to adjust the output
parameters based on the codec-specific decoding
options that are present in the mfxInfoMFX structure
after MFXVideoDECODE_DecodeHeader initializes them.
The workaround for https://github.com/Intel-Media-SDK/MediaSDK/issues/1139
is required for vp8 only, so move this workaround to the corresponding
postinit_decoder function
The pipeline below works with this change
gst-launch-1.0 filesrc location=SA10104.vc1 ! \
'video/x-wmv,profile=(string)advanced',width=720,height=480,framerate=14/1 ! \
msdkvc1dec ! fakesink
MFXVideoDECODE_DecodeHeader only parses the sequence layer for VC1, so
the structure is unknown for a stream with interlace flag set in the
sequence layer. If forcing the struct to progressive in this plugin,
MediaSDK will fail to decode such streams.
An output surface is returned but without sync point when when
MFXVideoDECODE_DecodeFrameAsync () returns MFX_ERR_MORE_DATA, this
surface should be released too, otherwise the surface is occupied
and it is easy to exhaust all pre-allocated mfx surfaces.
Example pipeline (input_vp8.webm contains lots of frame with show_frame
set to 0):
gst-launch-1.0 filesrc location=input_vp8.webm ! matroskademux !
msdkvp8dec ! msdkvpp ! fakesink
0:00:05.995959693 19866 0x563f30f14590 ERROR default
gstmsdkvideomemory.c:77:gst_msdk_video_allocator_get_surface: failed to
get surface available
ERROR: from element
/GstPipeline:pipeline0/GstMatroskaDemux:matroskademux0: Internal data
stream error.
The patch adds a serios of changes to support dynamic resolution
change and efficient utilization of resources.
Major changes:
-- Use MSDK's apis to retrieve the headers instead of only relying
on upsteram notification. For eg: avc decoder requires SEI header
information for dpb count calculation which we don't get from caps.
-- For all codecs other than VP9, we force the reset of decoder
if resoultion changes to fit with gstreamer flow. VP9 enfource
the hard reset only if the new resolution is bigger.
-- delay the src caps setting till msdk api's invokation in
handle_frame to avoid caching multiple configuration values
-- ensure pool negotiation is based on decoder's allocation_caps.
--dynamic resoluttion change use an explicit allocation_query
to reclaim the buffers before closing the decoder (thanks to v4l2dec)
--In case if we don't get upstream notification of res change (for eg,
this can can happen for vp9 frames with ivfheader where ivfparse
is not able to notify the dynamic changes), we handle the the case
based on MFX_ERR_INCOMPATIBLE_VIDEO_PARAM which is the return value
of MFXVideoDECODE_DecodeFrameAsync
-- calculate the minimum surfaces to be preallocated based on
msdk suggestion, downstream requirement, async depth and scratch surface
count for smooth display.
https://bugzilla.gnome.org/show_bug.cgi?id=796566
The new property "output-order" can be set to either "display" order
which is the default where frames will be outputting in display order,
or "decoded-order" which will be outputting the frames in decoded order.
The "decoded order" output is generally useful for debugging. But there
are few
customers who use it for low-latency streaming. For eg if the customer
already knows that the stream doesn't have b-frames (which means no
algorithm requires for display order calculation), then they can use
"decoded-order"
output to skip some of the DPB logic to avoid the frame accumulation at
start-up.
The root cause of the above issue is a bit of unclarity in h264 spec +
lazy implementation of many H264 encoders; This is well handled in
gstreamer-vaapi using "low-latency" property:
https://bugzilla.gnome.org/show_bug.cgi?id=762509https://bugzilla.gnome.org/show_bug.cgi?id=795783
The gst-msdk decoders prefer packetized streams as input
and in this case we can avoid unnecessary input bitstream copy
to mfxBitstream. This works fine for codecs like h264 where
we only support byte-stream with au alignment. Other format
conversions should be done thorugh parsers. But this won't work
for codecs like vc1 where we don't have an autoplugged parser.
Even the parser is not capable to do format conversions.
Packetizing through base decoders parse() routine will bring a
lot of uncecessary of complexities and codecparser libraray dependency.
So we just use an interal gst_adaper to keep track of bitstream
which is not consumed by msdk durig AsynchronusDecoding.
This adapter will get used only if subclass implementations
set the "is_packetized" to FALSE for msdk base encoder.
https://bugzilla.gnome.org/show_bug.cgi?id=792589
How to share/create GstMsdkcontext is the following:
- Search GstMsdkContext if there's in the pipeline.
- If found, check if it's decoder, encoder or vpp by job type.
- If it's same job type, it creates another instance of
GstMsdkContext
with joined-session.
- Otherwise just use the shared GstMsdkContext.
- If not found, just creates new instance of GstMsdkContext.
https://bugzilla.gnome.org/show_bug.cgi?id=790752
1\ In decide_allocation, it makes its own msdk bufferpool.
- If downstream supports video meta, it just replace it with the msdk
bufferpool.
- If not, it uses the msdk bufferpool as a side pool, which will be
decoded into.
and will copy it to downstream's bufferpool.
2\ Decide if using video memory or system memory.
- This is not completed in this patch.
- It might be decided in update_src_caps.
- But tested for both system memory and video memory cases.
https://bugzilla.gnome.org/show_bug.cgi?id=790752
Makes GstMsdkContext to be a descendant of GstObject so that
we could track the life-cycle of the session of the driver.
Also replaces MsdkContext with this one.
Keeps msdk_d3d.c alive for the future.
https://bugzilla.gnome.org/show_bug.cgi?id=790752
