The Matroska demuxer has three functions for creating packets out of
the data read: One for certain RealAudio codecs (ATRAC3, cook, sipr,
RealAudio 28.8), one for WebVTT (actually, the WebM flavour of it) and
one for all the others. Only the last function supported Matroska's
ContentCompression (e.g. it reversed zlib compression or added the
removed headers to the packets). But in Matroska, all tracks are allowed
to be compressed. This commit adds support for this.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Matroska is built around the principle that a reader does not need to
understand everything in a file in order to be able to make use of it;
it just needs to ignore the data it doesn't know about.
Our demuxer typically follows this principle, but there is one important
instance where it does not: A Block belonging to a TrackEntry with no
associated stream is treated as invalid data (i.e. the demuxer will try
to resync to the next level 1 element because it takes this as a sign
that it has lost sync). Given that we do not create streams if we don't
know or don't support the type of the TrackEntry, this impairs this
demuxer's forward compability.
Furthermore, ignoring Blocks belonging to a TrackEntry without
corresponding stream can (in future commits) also be used to ignore
TrackEntries with obviously bogus entries without affecting the other
TrackEntries (by not creating a stream for said TrackEntry).
Finally, given that matroska_find_track_by_num() already emits its own
error message in case there is no TrackEntry with a given TrackNumber,
the error message (with level AV_LOG_INFO) for this can be removed.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
A Block (meaning both a Block in a BlockGroup as well as a SimpleBlock)
must have at least three bytes after the field containing the encoded
TrackNumber. So if there are <= 3 bytes, the Matroska demuxer would
skip this block, believing it to be an empty, but valid Block.
This might discard valid nonempty Blocks, namely if the track uses header
stripping. And certain definitely spec-incompliant Blocks don't raise
errors: Those with two or less bytes left after the encoded TrackNumber
and those with three bytes left, but with flags indicating that the Block
uses lacing as then there has to be further data describing the lacing.
Furthermore, zero-sized packets were still possible because only the
size of the last entry of a lace was checked.
This commit fixes this. All spec-compliant Blocks that contain data
(even if side data only) are now returned to the caller; spec-compliant
Blocks that don't contain anything are not returned.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Some conditions which don't change and which can therefore be checked
in read_header() were instead rechecked upon parsing each block. This
has been changed.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
The Matroska demuxer splits every sequence of h Matroska Blocks into
h * w / cfs packets of size cfs; here h (sub_packet_h), w (frame_size)
and cfs (coded_framesize) are parameters from the track's CodecPrivate.
It does this by splitting the Block's data in h/2 pieces of size cfs each
and putting them into a buffer at offset m * 2 * w + n * cfs where
m (range 0..(h/2 - 1)) indicates the index of the current piece in the
current Block and n (range 0..(h - 1)) is the index of the current Block
in the current sequence of Blocks. The data in this buffer is then used
for the output packets.
The problem is that there is currently no check to actually guarantee
that no uninitialized data will be output. One instance where this is
trivially so is if h == 1; another is if cfs * h is so small that the
input pieces do not cover everything that is output. In order to
preclude this, rmdec.c checks for h * cfs == 2 * w and h >= 2. The
former requirement certainly makes much sense, as it means that for
every given m the input pieces (corresponding to the h different values
of n) form a nonoverlapping partition of the two adjacent frames of size w
corresponding to m. But precluding h == 1 is not enough, other odd
values can cause problems, too. That is because the assumption behind
the code is that h frames of size w contain data to be output, although
the real number is h/2 * 2. E.g. for h = 3, cfs = 2 and w = 3 the
current code would output four (== h * w / cfs) packets. although only
data for three (== h/2 * h) packets has been read.
(Notice that if h * cfs == 2 * w, h being even is equivalent to
cfs dividing w; the latter condition also seems very reasonable:
It means that the subframes are a partition of the frames.)
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
RealAudio 28.8 (like other RealAudio codecs) uses a special demuxing
mode in which the data of the existing Matroska Blocks is not simply
forwarded as-is. Instead data from several Blocks is recombined
together to output several packets. The parameters governing this
process are parsed from the CodecPrivate: Coded framesize (cfs), frame
size (w) and sub_packet_h (h).
During demuxing, h/2 pieces of data of size cfs each are read from every
Matroska (Simple)Block and put at offset m * 2 * w + n * cfs of a buffer
of size h * w, where m ranges from 0 to h/2 - 1 for each Block while n
is initially zero and incremented after a Block has been parsed until it
is h, at which poin the assembled packets are output and n reset.
The highest offset is given by (h/2 - 1) * 2 * w + (h - 1) * cfs + cfs
while the destination buffer's size is given by h * w. For even h, this
leads to a buffer overflow (and potential segfault) if h * cfs > 2 * w;
for odd h, the condition is h * cfs > 3 * w.
This commit adds a check to rule this out.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
RealAudio 28.8 does not need or use sub_packet_size for its demuxing
and this field is therefore commonly set to zero. But since 18ca491b
the Real Audio specific demuxing is no longer applied if sub_packet_size
is zero because the codepath for cook and ATRAC3 divide by it; this made
these files undecodable.
Furthermore, since 569d18aa (merged in 2c8d876d) sub_packet_size being
zero is used as an indicator for invalid data, so that a file containing
such a track was completely skipped.
This commit fixes this by not checking sub_packet_size for RealAudio
28.8 at all.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
They need a special parsing mode and in order to find out whether this
mode is in use, several checks have to be performed. They can all be
combined into one: If the buffer that is only used to assemble their
packets has been allocated, use the RealAudio parsing mode.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Only flavors 0..3 seem to exist. E.g. rmdec.c treats any flavor > 3
as invalid data. Furthermore, we do not know how big the packets to
create ought to be given that for sipr these values are not read from
the bitstream, but from a table.
Furthermore, flavor is only used for sipr, so only check it for sipr;
rmdec.c does the same. (The old check for flavor being < 0 was
always wrong given that flavor is an int that is read via avio_rb16(),
so it has been removed completely.)
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
This makes decoding far more robust, since OggS, the ogg magic,
can be commonly found randomly in streams, which previously made
the demuxer think there's a new stream or a change in such.
v4l2_m2m uses device memory mapped buffers to store dequeued
frames/packets (reference counted by AVBufferRef). When the reference
count drops to zero, the buffer ownership is returned back to the
device, so that they can re-filled with frames/packets.
There are some cases when all the capture buffers are in userspace
(i.e. due to internal buffering in ffmpeg). On the s5p-mfc this causes
an infinite wait when polling to dequeue the buffers, which can be
prevented by increasing the total number of buffers. This commit adds a
warning when all the capture buffers are dequeued.
Reviewed-by: Ming Qian <ming.qian@nxp.com>
Signed-off-by: Andriy Gelman <andriy.gelman@gmail.com>
Apparently the changes from 3c9185bf3a
aren't enough; even with that in place, I got errors like this
when trying to build for iOS:
src/libavdevice/avfoundation.m:135:5: error:
'AVCaptureDeviceTransportControlsPlaybackMode' is unavailable: not
available on iOS
AVCaptureDeviceTransportControlsPlaybackMode observed_mode;
^
The data of an AVPacket may be a part of the data of an AVBufferRef;
Therefore av_grow_packet() doesn't reallocate if the available space in
the actual buffer is sufficient for the enlargement. But given that it
also zeroes the padding it also needs to make sure that the buffer is
actually writable; this commit implements this.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
hdsenc already had an explicit function to free all allocations in case
of an error, but it was not marked as deinit function, so that it was
not called automatically when the AVFormatContext for muxing gets freed.
Using an explicit deinit function also makes the code cleaner by
allowing to return immediately without "goto fail".
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Enqueue/dequeue of the capture buffers should continue while draining.
Reference: linux/Documentation/media/uapi/v4l/dev-decoder.rst
"The client must continue to handle both queues independently,
similarly to normal decode operation. This includes:
...
- queuing and dequeuing CAPTURE buffers, until a buffer marked with
the V4L2_BUF_FLAG_LAST flag is dequeued"
Signed-off-by: Ming Qian <ming.qian@nxp.com>
Signed-off-by: Andriy Gelman <andriy.gelman@gmail.com>
The DNG code hardcodes plane 0 at some places, so its better to disallow cases
that have more planes.
Fixes: eg_crash
Found-by: 黄宁 <tsukimurarin@163.com>
Reviewed-by: Nick Renieris <velocityra@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
dng spec 1.5.0.0
"BitsPerSample
Supported values are from 8 to 32 bits/sample. The depth must be the same for each sample if
SamplesPerPixel is not equal to 1."
Fixes: eg_crash
Found-by: 黄宁 <tsukimurarin@163.com>
Reviewed-by: Nick Renieris <velocityra@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
"slice mode" option seems to be unnecessary since it could be
determined by -slices/max_nal_size.
default: SM_FIXEDSLCNUM_SLICE mode with cpu-number slices.
-slices N: SM_FIXEDSLCNUM_SLICE mode with N slices.
-max_nal_size: SM_SIZELIMITED_SLICE mode with limited size slices.
Add FF_API_OPENH264_SLICE_MODE macro to remove this option after
LIBAVCODEC_VERSION_MAJOR = 59.
Signed-off-by: Linjie Fu <linjie.fu@intel.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
Libopenh264enc would set the slice according to the number of cpu cores
if uiSliceNum equals to 0 (auto) in SM_FIXEDSLCNUM_SLICE mode.
Prompt a warning for user to catch this.
Signed-off-by: Linjie Fu <linjie.fu@intel.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
RC_BITRATE_MODE:
set BITS_EXCEEDED to iCurrentBitsLevel and allows QP adjust
in RcCalculatePictureQp().
RC_BUFFERBASED_MODE:
use buffer status to adjust the video quality.
RC_TIMESTAMP_MODE:
bit rate control based on timestamp, introduced in release 1.4.
Default to use RC_QUALITY_MODE.
Signed-off-by: Linjie Fu <linjie.fu@intel.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
It would be 200kbps bitrate with gop size = 12 by default
which generated too many IDR frames in rather low bit rate.
The quality would be poor.
Set these default values to -1 to check whether it's specified
by user explicitly.
Use 2Mbps bitrate as nvenc sugguested, and leave gop size
untouched in libopenh264.
Signed-off-by: Linjie Fu <linjie.fu@intel.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
Clip iMinQp/iMaxQp to (1, 51) for user specified qp range.
If not set, leave iMinQp/iMaxQp untouched and use the values (0, 51)
initialized in FillDefault(), and the QP range would be adjusted to the
defaults inside libopenh264 library according to the iUsageType, (12, 42)
for iUsageType == CAMERA_VIDEO_REAL_TIME which is default.
<https://github.com/cisco/openh264/blob/master/codec/encoder/core/src/encoder_ext.cpp#L375>
Signed-off-by: Linjie Fu <linjie.fu@intel.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
Because not every user know about in_pad and out_pad reasonable value range
so maybe try to set 1.0, but setting 1.0 is so hugh to get an fatal error.
Suggested-by: Paul B Mahol <onemda@gmail.com>
Signed-off-by: Steven Liu <lq@chinaffmpeg.org>
The current implementation allows passing levels to libavcodec as
integers (such as "31" instead of "3.1").
However, in this case, the maximum reference frame value per level was
ignored because libavcodec converted the string to 310 instead of 31.
Since libx264 has correctly parsed the level to int
(x4->params.i_level_idc), we should rely on this value instead of
attempting to parse the level string on our own.
Signed-off-by: Josh Brewster <josh.brewster@protonmail.com>
Signed-off-by: Josh de Kock <josh@itanimul.li>
Andreas Rheinhardt
Limin Wang
Lynne
James Almer
Andriy Gelman
Martin Storsjö
Paul B Mahol
Josh de Kock
Ming Qian
Michael Niedermayer
Linjie Fu
Steven Liu
Josh Brewster
Jun Zhao