Current code is written around the "simple" decode API's limitation that
a single input packet (AU/coded frame) triggers the output of at most
one output frame. However the spec contains two cases where a coded
frame may cause multiple frames to be output (cf. C.5.2.2.2):
* start of a new sequence
* overflowing sps_max_dec_pic_buffering
The decoder currently contains rather convoluted logic to handle these
cases:
* decode/output/per-frame sequence counters,
* HEVC_FRAME_FLAG_BUMPING
* ff_hevc_bump_frame()
* special clauses in ff_hevc_output_frame()
However, with the receive_frame() API none of that is necessary, as we
can just output multiple frames at once. Previously added ContainerFifo
allows that to be done in a straightforward and efficient manner.
It provides a FIFO for "container" objects like AVFrame/AVPacket and
features an integrated FFRefStructPool-based pool to avoid allocating an
freeing them repeatedly.
Fixes: left shift of 1 by 31 places cannot be represented in type 'int'
Fixes: 70726/clusterfuzz-testcase-minimized-ffmpeg_AV_CODEC_ID_HEVC_fuzzer-6149928703819776
Found-by: continuous fuzzing process https://github.com/google/oss-fuzz/tree/master/projects/ffmpeg
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Usually the MediaCodec context will be released immediately, or it needs to stay
alive due to existing hardware buffers.
However we can free resources early in the case of
hw_buffer_count == 0 && refcount > 1, which can be reproduced by keeping frames
referenced after flushing and closing. mpv currently behaves like this.
Signed-off-by: sfan5 <sfan5@live.de>
Signed-off-by: Matthieu Bouron <matthieu.bouron@gmail.com>
In the absence of an RPU header, we can consult the colorspace tags to
make a more informed guess about whether we're looking at profile 5 or
profile 8.
This implements limited metadata compression. To be a bit more lenient,
we try and re-order the static extension blocks when testing for an
exact match.
For sanity, and to avoid producing bitstreams we couldn't ourselves
decode, we don't accept partial matches - if some extension blocks
change while others remain static, compression is disabled for the
entire frame.
This shouldn't be an issue in practice because static extension blocks
are stated to remain constant throughout the entire sequence.
This implements the limited DM metadata compression scheme described in
chapter 9 of the dolby vision bitstream specification.
The spec is a bit unclear about how to handle the presence of static
metadata inside compressed frames; in that it doesn't explicitly forbid
an encoder from repeating redundant metadata. In theory, we would need
to detect this case and then strip the corresponding duplicate metadata
from the existing set of static metadata. However, this is difficult to
implement - esspecially for the case of metadata blocks which may be
internally repeated (e.g. level 10).
That said, the spec states outright that static metadata should be
constant throughout the entire sequence, so a sane bitstream should not
have any static metadata values changing from one frame to the next (at
least up to a keyframe boundary), and therefore they should never be
present in compressed frames. As a consequence, it makes sense to treat
this as an error state regardless. (Ignoring them by default, or
erroring if either AV_EF_EXPLODE or AV_EF_AGGRESSIVE are set)
I was not able to find such samples in the wild (outside of artificially
produced test cases for this exact scenario), so I don't think we need
to worry about it until somebody produces one.
Keyframes must reset the metadata compression state, so we need to
also signal this at rpu generation time.
Default to uncompressed, because encoders cannot generally know if
a given frame will be a keyframe before they finish encoding, but also
cannot retroactively attach the RPU. (Within the confines of current
APIs)
And move the choice of desired container to `flags`. This is needed to
handle differing API requirements (e.g. libx265 requires the NAL RBSP,
but CBS BSF requires the unescaped bytes).
Limited mode can only ever maintain a single VDR RPU reference, and
furthermore requires vdr_rpu_id == 0. So in practice, it will only ever
use VDR RPU slot 0. All remaining slots get flushed in this case, to
avoid leaking partial state.
As the comment implies, DOVIContext.ext_blocks should also reflect the
current state after ff_dovi_rpu_generate().
Fluff for now, but will be needed once we start implementing metadata
compression for extension blocks as well.
Vulkan encoding was designed in a very... consolidated way.
You had to know the exact codec and profile that the image was going to
eventually be encoded as at... image creation time. Unfortunately, as good
as our code is, glimpsing into the exact future isn't what its capable of.
video_maintenance1 removed that requirement, which only then made encoding
images practically possible.
In 680d969a305c0927480573a1b455024088b51aeb, the new API was
used to find a queue family for dispatch, but the found queue
family was not used for decoding, just for dispatching.
From Jun Zhao <mypopydev@gmail.com>:
> Should we relocate this to the decoder? Other codecs typically set this
> parameter in the decoder.
Signed-off-by: Wu Jianhua <toqsxw@outlook.com>
memset tables in the main thread can become a bottleneck for the decoder.
For example, if it takes 1% of the processing time for one core, the maximum achievable FPS will be 100.
Move the memeset to worker threads will fix the issue.
For luma, qp can only change at the CU level, so the qp tab size is related to the CU.
For chroma, considering the joint CbCr, the QP tab size is related to the TU.
