Instead of using a combination of bitreader and -writer for copying data,
one can byte-align the (obsolete and removed) bitreader to improve performance.
With the right alignment one can even use memcpy. The right alignment
normally exists for CABAC and hence for H.265 in general.
For aligned data this reduced the time to copy the slicedata from
776520 decicycles to 33889 with 262144 runs and a 6.5mb/s H.264 video.
For unaligned data the number went down from 279196 to 97739 decicycles.
Signed-off-by: Mark Thompson <sw@jkqxz.net>
Also add SIMD which works on lines because it is faster then calculating it on
8x8 blocks using pixelutils.
Signed-off-by: Marton Balint <cus@passwd.hu>
Instead of using a combination of bitreader and -writer for copying data,
one can byte-align the (obsolete and removed) bitreader to improve performance.
One can even use memcpy in the normal case.
This improved the time needed for writing the slicedata from 33618 to
2370 decicycles when tested on a video originating from a DVD (4194394
runs).
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@googlemail.com>
Signed-off-by: Mark Thompson <sw@jkqxz.net>
Fixes: Out of memory
Fixes: 10970/clusterfuzz-testcase-minimized-ffmpeg_AV_CODEC_ID_IMM4_fuzzer-5698750043914240
Reviewed-by: Paul B Mahol <onemda@gmail.com>
Found-by: continuous fuzzing process https://github.com/google/oss-fuzz/tree/master/projects/ffmpeg
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
The first element of H264RedundantPPSContext is not a pointer to an
AVClass as required.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@googlemail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
After inspecting the source code of x265, mpv and ffmpeg, I've found that
ffmpeg mistakenly regards EVC_NAL_BLA_N_LP and HEVC_NAL_IDR_N_LP as non-
reference frames, which are acutally reference frames according to the
specification in x265, and drops them.
This patch should address the problem. I have tested it with mpv.
Signed-off-by: Mark Wu <wfwf1997@gmail.com>
Signed-off-by: James Almer <jamrial@gmail.com>
This improves the speed of decoding large patches of constant color
Fixes: Timeout
Fixes: 10967/clusterfuzz-testcase-minimized-ffmpeg_AV_CODEC_ID_QPEG_fuzzer-5630803793936384
Found-by: continuous fuzzing process https://github.com/google/oss-fuzz/tree/master/projects/ffmpeg
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
only 16b alpha is supported (not 8 bits)
following official encoder, alpha data doesn't impact
yuv plane quality.
So the alpha data encoding is done after the yuv part.
It's also avoid to loose quality in yuv part when
alpha is not uniform.
the alpha encoding funcs is mainly take from prores_ks
encoder, except for the alpha data reorganization
The earlier code used the most recent non-auxiliary slice to determine
whether an auxiliary slice has the syntax of an IDR slice, even when
the most recent slice was from a slice of a redundant frame. Now only
slices of the primary coded picture are used, as the specifications
mandate.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@googlemail.com>
Flush the buffered data in libmfx before video param reinit
in case the frames drop.
Cache the first frame causing the reinit and decode zero-size
pkt to flush the buffered pkt before reinit. After all the
buffered pkts being flushed, resume to reinit and decode.
Fix the issue in ticket #7399.
[V2]: Move the definition of zero_pkt to where it is exactly
used.
Signed-off-by: Linjie Fu <linjie.fu@intel.com>
Signed-off-by: Zhong Li <zhong.li@intel.com>
