Up until now, the encoder replicated all the border pixels
for incomplete 16x16 macroblocks. In case the available width
or height is <= 8, some of the luma blocks of the MB
do not correspond to actual input, so that we should encode
them using the least amount of bits. Zeroing the block coefficients
(as this commit does) achieves this, replicating the pixels
and performing an FDCT does not.
This commit also removes the frame copying code for insufficiently
aligned dimensions.
The vsynth3-asv[12] FATE tests use a 34x34 input file and are
therefore affected by this. As the ref updates show, the size
and checksum of the encoded changes, yet the decoded output
stays the same.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
The earlier code only used the counts from the last slice.
The two FATE tests using slices show compression improvements
due to this.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
It is unnecessary, as we already have the entries sorted by
probability and therefore implicitly by length. All we need
on top of that to build the tree is the number of entries
of a given length.
Doing so gives a 3.6% speedup of ff_mjpeg_encode_huffman_close()
here; it also saves about 640B of .text here.
The new code puts values with higher probability to the left
of the tree. The old code did not and therefore
the FATE checksums needed to be updated. Due to MJPEG's
0xFF unescaping file sizes as well as file checksums
needed to be updated; the decoded picture hashes stayed
the same. Given that codes on the left of the tree have
on average fewer bits set than codes on the right, the
file sizes mostly improve (all except vsynth3-mjpeg-444).
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
this is a good compromise between speed and compression
-rw-r----- 1 michael michael 180987 Feb 6 14:29 lena-def.png
-rw-r----- 1 michael michael 128430 Feb 6 14:36 lena-pavg.png
-rw-r----- 1 michael michael 126269 Feb 6 14:36 lena-pmixed.png
-rw-r----- 1 michael michael 180987 Feb 6 14:35 lena-pnone.png
-rw-r----- 1 michael michael 127758 Feb 6 14:35 lena-ppaeth.png
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Since 45eeb1f785d6c726181568db341bbf80b43e32c0
optimal Huffman tables are the default (without slice-threading).
This made the fate-vsynth*-mjpeg-{trell-,}-huffman tests
identical to their corresponding tests without "-huffman".
This is of course wasteful, so switch the two tests with
"-huffman" counterparts back to the default tables.
Also use one of these tests to test slice threaded encoding.
It has so far been untested.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
With this modification the test would have caught the regression
introduced in 72bf3d3c12ae562fd408603483bc59058757b1a1.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Encoding was untested before this.
Notice that the filesize degradation is partially due to
mpegvideo no longer using progressive_sequence and
progressive_frame.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
When there are multiple candidates for macroblock type, the encoder
tries them all. In order to do so, it keeps several sets of states
containing the variables that get modified when encoding
the macroblock and in the end uses the best of these.
Yet one variable was set, but not included in this state:
The current macroblock's qscale value in the current picture's
qscale_table. This may currently be set multiple times in
mpv_reconstruct_mb(), yet it is read when adaptive_quant is true.
Currently, the value read can be the value set by the last attempt
to write the current macroblock and not the initial value.
Fix this by only setting the qscale_table value in one place
outside of mpv_reconstruct_mb() (where it does not belong at all).
Reviewed-by: Ramiro Polla <ramiro.polla@gmail.com>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
There is an issue with the constants used in YUV to YUV range conversion,
where the upper bound is not respected when converting to mpeg range.
With this commit, the constants are calculated at runtime, depending on
the bit depth. This approach also allows us to more easily understand how
the constants are derived.
For bit depths <= 14, the number of fixed point bits has been set to 14
for all conversions, to simplify the code.
For bit depths > 14, the number of fixed points bits has been raised and
set to 18, to allow for the conversion to be accurate enough for the mpeg
range to be respected.
The convert functions now take the conversion constants (coeff and offset)
as function arguments.
For bit depths <= 14, coeff is unsigned 16-bit and offset is 32-bit.
For bit depths > 14, coeff is unsigned 32-bit and offset is 64-bit.
x86_64:
chrRangeFromJpeg8_1920_c: 2127.4 2125.0 (1.00x)
chrRangeFromJpeg16_1920_c: 2325.2 2127.2 (1.09x)
chrRangeToJpeg8_1920_c: 3166.9 3168.7 (1.00x)
chrRangeToJpeg16_1920_c: 2152.4 3164.8 (0.68x)
lumRangeFromJpeg8_1920_c: 1263.0 1302.5 (0.97x)
lumRangeFromJpeg16_1920_c: 1080.5 1299.2 (0.83x)
lumRangeToJpeg8_1920_c: 1886.8 2112.2 (0.89x)
lumRangeToJpeg16_1920_c: 1077.0 1906.5 (0.56x)
aarch64 A55:
chrRangeFromJpeg8_1920_c: 28835.2 28835.6 (1.00x)
chrRangeFromJpeg16_1920_c: 28839.8 32680.8 (0.88x)
chrRangeToJpeg8_1920_c: 23074.7 23075.4 (1.00x)
chrRangeToJpeg16_1920_c: 17318.9 24996.0 (0.69x)
lumRangeFromJpeg8_1920_c: 15389.7 15384.5 (1.00x)
lumRangeFromJpeg16_1920_c: 15388.2 17306.7 (0.89x)
lumRangeToJpeg8_1920_c: 19227.8 19226.6 (1.00x)
lumRangeToJpeg16_1920_c: 15387.0 21146.3 (0.73x)
aarch64 A76:
chrRangeFromJpeg8_1920_c: 6324.4 6268.1 (1.01x)
chrRangeFromJpeg16_1920_c: 6339.9 11521.5 (0.55x)
chrRangeToJpeg8_1920_c: 9656.0 9612.8 (1.00x)
chrRangeToJpeg16_1920_c: 6340.4 11651.8 (0.54x)
lumRangeFromJpeg8_1920_c: 4422.0 4420.8 (1.00x)
lumRangeFromJpeg16_1920_c: 4420.9 5762.0 (0.77x)
lumRangeToJpeg8_1920_c: 5949.1 5977.5 (1.00x)
lumRangeToJpeg16_1920_c: 4446.8 5946.2 (0.75x)
NOTE: all simd optimizations for range_convert have been disabled.
they will be re-enabled when they are fixed for each architecture.
NOTE2: the same issue still exists in rgb2yuv conversions, which is not
addressed in this commit.
This made the output non binary reproducible across different versions of
zlib.
This reverts commit bce5855afb25d318e090c2e6c16117f065458356.
Signed-off-by: James Almer <jamrial@gmail.com>
stddev and PSNR values change by the removal of format conversion and because
of the added sws_flags. Either or will have the same effect even on their own.
Signed-off-by: James Almer <jamrial@gmail.com>
FATE results differ when using the original zlib and zlib-ng.
Since we don't need to test the result from zlib itself, this commit
disables compression on tests for zlib-based codecs, which ends up
giving the same results with both libraries.
Signed-off-by: James Almer <jamrial@gmail.com>
Do not pass an options dictionary to the avcodec_open2() in enc_open().
This is cleaner and more robust, as previously various bits of code
would try to interpret the contents of the options dictionary, with
varying degrees of correctness. Now they can just access the encoder
AVCodecContext directly.
Cf. 372c78dd42f2b1ca743473b9c32fad71c65919e0 - analogous change for
decoding.
A non-progressive field order is now written on the container level in
interlaced ProRes encoding tests.
These tables are supposed to contain the number of bits needed
to encode a given (run, level) pair. Yet the number of bits
for pairs needing the escape code was wrong (it only contained
the escape code and not the bits needed for run and level).
Furthermore, H.261 (a format with explicit end-of-block codes)
does not work well together with the RLTable API from rl.c:
The EOB code is the first one in ff_h261_rl_tcoeff's VLC table
and has a run value of zero. Therefore the result of get_rl_index()
is off by one for run == 0 and level values with explicit
(run, level) pair.
Fixing this necessitated changing the ref files of the
vsynth*-h261-trellis tests. Both filesizes as well as PSNR
decreased. If one used a qscale value of 11 for this test,
one would have received files with about the same size as
before this patch (with qscale 12), but with better PSNR.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
In direct_search() and ff_estimate_b_frame_motion(), penalty_factor
would be used before being initialized in estimate_motion_b(). Also,
the initialization would happen more than once unnecessarily.
The layout for the frame flags is as follow:
chroma_format u(2)
reserved u(2)
interlace_mode u(2)
reserved u(2)
chroma_format has 2 allowed values:
0: reserved
1: reserved
2: 4:2:2
3: 4:4:4
interlace_mode has 3 allowed values:
0: progressive
1: tff
2: bff
3: reserved
0x80 is what we expect for "422 not interlaced", and the extra 0x2 from
0x82 is actually writing into the reserved bits.
This byte represents 4 reserved bits followed by 4 alpha_channel_type bits.
alpha_channel_type currently has 3 differents defined values: 0 (no
alpha), 1 (8b alpha), and 2 (16b alpha), all the other values are
reserved. The 4 initial reserved bits are expected to be 0.
This byte represents 4 reserved bits followed by 4 alpha_channel_type bits.
alpha_channel_type currently has 3 differents defined values: 0 (no
alpha), 1 (8b alpha), and 2 (16b alpha), all the other values are
reserved. This part is correctly written (alpha_bits>>3 does the correct
thing), but the 4 initial bits are reserved.
Quoting SMPTE RDD 36:2015:
A decoder shall abort if it encounters a bitstream with an unsupported
bitstream_version value. If 0, the value of the chroma_format syntax
element shall be 2 (4:2:2 sampling) and the value of the
alpha_channel_type element shall be 0 (no encoded alpha); if 1, any
permissible value may be used for those syntax elements.
So if we're not in 4:2:2 or if there is alpha, we are not allowed to use
version 0.
Should fix integer overflows, and improve encoding results.
Reviewed-by: Paul B Mahol <onemda@gmail.com>
Signed-off-by: James Almer <jamrial@gmail.com>
It is done later in ff_mpv_frame_start() (and nobody uses
current_picture_ptr between setting it in ff_mpv_frame_start()).
(The reason the vsynth*-h263-obmc ref files change is because
the call to ff_find_unused_picture() now happens after the older
pictures have been unreferenced in ff_mpv_frame_start(),
so that their slots in the picture array can be immediately
reused; the obmc code is somehow buggy and changes its output
depending on the earlier contents of the motion_val buffer.)
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
These tables are only used by encoders and only for the current picture;
ergo they need not be put into the picture at all, but rather into
the encoder's context. They also don't need to be refcounted,
because there is only one owner.
In contrast to this, the earlier code refcounts them which
incurs unnecessary overhead. These references are not unreferenced
in ff_mpeg_unref_picture() (they are kept in order to have something
like a buffer pool), so that several buffers are kept at the same
time, although only one is needed, thereby wasting memory.
The code also propagates references to other pictures not part of
the pictures array (namely the copy of the current/next/last picture
in the MpegEncContext which get references of their own). These
references are not unreferenced in ff_mpeg_unref_picture() (the
buffers are probably kept in order to have something like a pool),
yet if the current picture is a B-frame, it gets unreferenced
at the end of ff_mpv_encode_picture() and its slot in the picture
array will therefore be reused the next time; but the copy of the
current picture also still has its references and therefore
these buffers will be made duplicated in order to make them writable
in the next call to ff_mpv_encode_picture(). This is of course
unnecessary.
Finally, ff_find_unused_picture() is supposed to just return
any unused picture and the code is supposed to work with it;
yet for the vsynth*-mpeg4-adap tests the result depends upon
the content of these buffers; given that this patchset
changes the content of these buffers (the initial content is now
the state of these buffers after encoding the last frame;
before this patch the buffers used came from the last picture
that occupied the same slot in the picture array) their ref-files
needed to be changed. This points to a bug somewhere (if one removes
the initialization, one gets uninitialized reads in
adaptive_quantization in ratecontrol.c).
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
For 422 frames we should not use hard coded 8 to calculate mb size for
uv plane. Chroma shift should be taken into consideration to be
compatiple with different sampling format.
The error is reported by fate test when av_cpu_max_align() return 64
on the platform supporting AVX512. This is a hidden error and it is
exposed after commit 17a59a634c39b00a680c6ebbaea58db95594d13d.
mpeg2enc has a mechanism to reuse frames. When it computes SSE (sum of
squared error) on current mb, reconstructed mb will be wrote to the
previous mb space, so that the memory can be saved. However if the align
is 64, the frame is shared in somewhere else, so the frame cannot be
reused and a new frame to store reconstrued data is created. Because the
height of mb is wrong when compute sse on 422 frame, starting from the
second line of macro block, changed data is read when frame is reused
(we need to read row 16 rather than row 8 if frame is 422), and unchanged
data is read when frame is not reused (a new frame is created so the
original frame will not be changed).
That is why commit 17a59a634c39b00a680c6ebbaea58db95594d13d exposes this
issue, because it add av_cpu_max_align() and this function return 64 on
platform supporting AVX512 which lead to creating a frame in mpeg2enc,
and this lead to the different outputs.
Signed-off-by: Wenbin Chen <wenbin.chen@intel.com>
Signed-off-by: Marton Balint <cus@passwd.hu>
log2() remains, this can either be replaced by a integer implementation or the table
hardcoded if needed
Tested-by: Anton Khirnov <anton@khirnov.net>
Tested-by: Martin Storsjö <martin@martin.st>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Accidentally resurrected in fc49f22c3b735db5aaac5f98e40b7124a2be13b8
and 7711f19eda40a7fd1c8a327f1700ffdd115e1667,
forgotten in 6ebc71847e21a8abaf7ff3d902cb1cf63511e7b4 and
1a6a088f7c7b164042ad16d43d05543ce1bacfa4 or never needed
(filter-aemphasis).
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
These conversion appears to be exhibiting the same rounding error as the rgbf32 formats where.
I seperated the rounding value from the 16 and 128 offsets, I think it makes it a little more clear.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This is utilized by various media ingests to figure out the bit
rate of the content you are pushing towards it, so write it for
video, audio and subtitle tracks in case at least one nonzero value
is available. It is only mentioned for timed metadata sample
descriptions in QTFF, so limit it only to ISOBMFF (MODE_MP4) mode.
Updates the FATE tests which have their results changed due to the
20 extra bytes being written per track.
