Unlike other tf.*.conv2d layers, tf.nn.conv2d does not create many
nodes (within a scope) in the graph, it just acts like other layers.
tf.nn.conv2d only creates one node in the graph, and no internal
nodes such as 'kernel' are created.
The format of native model file is also changed, a flag named
has_bias is added, so change the version number.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
Allows the creation of the sdtp atom while remuxing MP4 to MP4. This
atom is required by Apple devices (iPhone, Apple TV) in order to accept
2160p medias.
The flac parser uses a fifo to buffer its data. Consequently, when
searching for sync codes of flac packets, one needs to take care of
the possibility of wraparound. This is done by using an optimized start
code search that works on each of the continuous buffers separately and
by explicitly checking whether the last pre-wrap byte and the first
post-wrap byte constitute a valid sync code.
Moreover, the last MAX_FRAME_HEADER_SIZE - 1 bytes ought not to be searched
for (the start of) a sync code because a header that might be found in this
region might not be completely available. These bytes ought to be searched
lateron when more data is available or when flushing.
Unfortunately there was an off-by-one error in the calculation of the
length to search of the post-wrap buffer: It was too large, because the
calculation was based on the amount of bytes available in the fifo from
the last pre-wrap byte onwards. This meant that a header might be
parsed twice (once prematurely and once regularly when more data is
available); it could also mean that an invalid header will be treated as
valid (namely if the length of said invalid header is
MAX_FRAME_HEADER_SIZE and the invalid byte that will be treated as the
last byte of this potential header happens to be the right CRC-8).
Should a header be parsed twice, the second instance will be the best child
of the first instance; the first instance's score will be
FLAC_HEADER_BASE_SCORE - FLAC_HEADER_CHANGED_PENALTY ( = 3) higher than
the second instance's score. So the frame belonging to the first
instance will be output and it will be done as a zero length frame (the
difference of the header's offset and the child's offset). This has
serious consequences when flushing, as returning a zero length buffer
signals to the caller that no more data will be output; consequently the
last frames not yet output will be dropped.
Furthermore, a "sample/frame number mismatch in adjacent frames" warning
got output when returning the zero-length frame belonging to the first
header, because the child's sample/frame number of course didn't match
the expected sample frame/number given its parent.
filter/hdcd-mix.flac from the FATE-suite was affected by this (the last
frame was omitted) which is the reason why several FATE-tests needed to
be updated.
Fixes ticket #5937.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
A threshold of 1 is sufficient for simple_dump_cut.webm, 10 is used
just to be sure the next truncated file doesnt cause the same issue
Obvious alternative fixes are to simply accept that the file is broken or to
write some advanced error concealment or to
simply accept that the decoder wont stop at the end of input.
Fixes: Ticket 8069 (artifacts not the differing md5 which was there before 1afd246960)
Fixes: simple_dump_cut.webm
Fixes: regression of 1afd246960
fate-vp5 changes because the last frame is truncated and now handled
differently.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Because the lavf_container is sometimes called with only 2 arguments,
fate tests produce bash errors like this:
tests/fate-run.sh: 299: test: =: unexpected operator
This commit fixes this.
Reviewed-by: Limin Wang <lance.lmwang@gmail.com>
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Right now, the concat filter does not set the frame_rate value on any of
the out links. As a result, the default ffmpeg behaviour kicks in - to
copy the framerate from the first input to the outputs.
If a later input is higher framerate, this results in dropped frames; if
a later input is lower framerate it might cause judder.
This patch checks if all of the video inputs have the same framerate, and
if not it sets the out link to use '1/0' as the frame rate, the value
meaning "unknown/vfr".
A test is added to verify the VFR behaviour. The existing test for CFR
behaviour passes unchanged.
the info can be saved in dnn operand object without regenerating again and again,
and it is also needed for layer split/merge, and for memory reuse.
to make things step by step, this patch just focuses on c code,
the change within python script will be added later.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
This makes the code bitexact between platforms.
Intermediate timestamps between frames are preserved.
The timebase is simplified.
Rounding differs from doubles in cases where timestamps/durations
are "funny"
Suggested-by: jb
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This reverts commit a9dacdeea6.
This patch effectively made the decoder output vfr content out of samples
where cfr is expected.
Addresses ticket #7880.
Signed-off-by: James Almer <jamrial@gmail.com>
The packet counting based approach caused excessive sdt/pat/pmt for VBR, so
let's use a timestamp based approach instead similar to how we emit PCRs.
SDT/PAT/PMT period should be consistent for both VBR and CBR from now on.
Also change the type of sdt_period and pat_period to AV_OPT_TYPE_DURATION so no
floating point math is necessary.
Fixes ticket #3714.
Signed-off-by: Marton Balint <cus@passwd.hu>
background:
DNN (deep neural network) is a sub module of libavfilter, and FATE/dnn
is unit test for the DNN module, one unit test for one dnn layer.
The unit tests are not based on the APIs exported by libavfilter,
they just directly call into the functions within DNN submodule.
There is an issue when run the following command:
build$ ../ffmpeg/configure --disable-static --enable-shared
make
make fate-dnn-layer-pad
And part of error message:
tests/dnn/dnn-layer-pad-test.o: In function `test_with_mode_symmetric':
/work/media/ffmpeg/build/src/tests/dnn/dnn-layer-pad-test.c:73: undefined reference to `dnn_execute_layer_pad'
The root cause is that function dnn_execute_layer_pad is a LOCAL symbol
in libavfilter.so, and so the linker could not find it when build dnn-layer-pad-test.
To check it, just run: readelf -s libavfilter/libavfilter.so | grep dnn
So, add dependency in fate/dnn Makefile with ffmpeg static libraries.
This is the same method used in fate/checkasm
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
This fixes make fate issue for frame thread scale in my local testing
Signed-off-by: Limin Wang <lance.lmwang@gmail.com>
Signed-off-by: Marton Balint <cus@passwd.hu>
At the moment scene change detection score uses all planes to detect scene
changes. In this regard this is similar how the frozen frames detection works.
However, in classic encoding scene change detection typically only uses the Y
plane.
We might get more resonable scores for scene change if we also use only
the Y plane for calculating the score if the pixel format is YUV. Although
this will require additional work once packed YUV formats are added,
because for the moment the generic scene sad score calculation has no way
to ignore some components in a packed format.
Signed-off-by: Limin Wang <lance.lmwang@gmail.com>
Signed-off-by: Marton Balint <cus@passwd.hu>
cuda_runtime.h as well as dynlink_loader.h used nonstandard inclusion
guards with an AV_ prefix, although these files are not in an libav*/
path. So change the inclusion guards and adapt the ref file of the
source fate test accordingly.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: Timo Rothenpieler <timo@rothenpieler.org>
why change .4 to .25, it's for:
one scenecut(pkt_pts=20040) isn't detected by 0.4 threshold
why not change to 0.3 instead of 0.25:
it will miss the scenecut(pkt_pts=20040) after applying the next
patch which enables yuvj420
for fate testing, it's better to catch all scenecut scenes.
Reviewed-by: Marton Balint <cus@passwd.hu>
Signed-off-by: Limin Wang <lance.lmwang@gmail.com>
The tests previously rounded the timestamps. Its better in a fate test to preserve
the data from the demuxer and decoder.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Commit cd48318035 added support for NV24 and NV42, including several
fate tests for these formats, but did not include the reference files
for the tests filter-pixdesc-nv24 and filter-pixdesc-nv42. As a result,
these two tests were broken.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
The implementation is pretty straight-forward. Most of the existing
NV12 codepaths work regardless of subsampling and are re-used as is.
Where necessary I wrote the slightly different NV24 versions.
Finally, the one thing that confused me for a long time was the
asm specific x86 path that did an explicit exclusion check for NV12.
I replaced that with a semi-planar check and also updated the
equivalent PPC code, which Lauri kindly checked.
These are the 4:4:4 variants of the semi-planar NV12/NV21 formats.
These formats are not used much, so we've never had a reason to add
them until now. VDPAU recently added support HEVC 4:4:4 content
and when you use the OpenGL interop, the returned surfaces are in
NV24 format, so we need the pixel format for media players, even
if there's no direct use within ffmpeg.
Separately, there are apparently webcams that use NV24, but I've
never seen one.
Up until now, the length field of most level 1 elements has been written
using eight bytes, although it is known in advance how much space the
content of said elements will take up so that it would be possible to
determine the minimal amount of bytes for the length field. This
commit changes this.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: James Almer <jamrial@gmail.com>
Given that in both the seekable as well as the non-seekable mode dynamic
buffers are used to write level 1 elements and that now no seeks are
used in the seekable case any more, the two modes can be combined; as a
consequence, the non-seekable mode automatically inherits the ability to
write CRC-32 elements.
There are no differences in case the output is seekable; when it is not
and writing CRC-32 elements is disabled, there can still be minor
differences because before this commit, the EBML ID and length field
were counted towards the cluster size limit; now they no longer are.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: James Almer <jamrial@gmail.com>
Up until now the EBML Header length field has been written with eight
bytes, although the EBML Header is always so small that only one byte
is needed for it. This patch saves seven bytes for every Matroska/Webm
file.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: James Almer <jamrial@gmail.com>
The spec in https://xiph.org/vorbis/doc/v-comment.html states that
the metadata keys are case-insensitive, so don't change the case
and update the fate test case.
Fix#7784
Reviewed-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
Signed-off-by: Jun Zhao <barryjzhao@tencent.com>
The transcode() helper function will already prepend the TARGET_PATH to
the sample path, if its a relative path. This avoids an issue on
Windows, where the relative path check could fail.
write_tmcd allows tmcd track to be created with any mode but in
mov_write_header, index for first tmcd track is only set for modes
MP4 or MOV, causing a crash if tmcd creation is attempted with other
modes.
* commit 'f8df5e2f31a5ba7b30a0e1caaaf5a03c753b3f9b':
tests: Add a convenience function for video-only lavf tests
Merged-by: James Almer <jamrial@gmail.com>
* commit 'a70eac7a9b193e8434b5bed90bd72aa4cb688363':
tests: Convert image2pipe tests to non-legacy test scripts
Merged-by: James Almer <jamrial@gmail.com>
When a JACOsub subtitle has two timestamps, they represent its start and
end times (http://unicorn.us.com/jacosub/jscripts.html#l_times); the
duration is the difference between the two, not the sum of the two.
The subtitle end times in the FATE test for this were wrong as a result;
fix them too. (This test is based on JACOsub's demo.txt, and the end
time computed for the last line using @ now matches what the comments
there say it should be.)
Also tested in practice using MPV, a LaserDisc, and some authentic 1993
JACOsub files.
Signed-off-by: Adam Sampson <ats@offog.org>
Reviewed-by: Paul B Mahol <onemda@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
If we fill with black then the generated palette will have one color more
than what the user requested. This also resulted in unwanted black specks in
the output of paletteuse, especially when generating small palettes.
