Name and purpose are more appropriate there since the code isn't
an ideal example.
Reviewed-by: wm4 <nfxjfg@googlemail.com>
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
When support for this was added the details weren't yet finalized.
This is no longer the case.
Fixes writing of mkv/webm files with HDR.
Reported-by: Kagami Hiiragi <kagami@genshiken.org>
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Reviewed-by: James Almer <jamrial@gmail.com>
This commit replaces the current inefficient non-power-of-two FFT with a
much faster FFT based on the Prime Factor Algorithm.
Although it is already much faster than the old algorithm without SIMD,
the new algorithm makes use of the already very throughouly SIMD'd power
of two FFT, which improves performance even more across all platforms
which we have SIMD support for.
Most of the work was done by Peter Barfuss, who passed the code to me to
implement into the iMDCT and the current codebase. The code for a
5-point and 15-point FFT was derived from the previous implementation,
although it was optimized and simplified, which will make its future
SIMD easier. The 15-point FFT is currently using 6% of the current
overall decoder overhead.
The FFT can now easily be used as a forward transform by simply not
multiplying the 5-point FFT's imaginary component by -1 (which comes
from the fact that changing the complex exponential's angle by -1 also
changes the output by that) and by multiplying the "theta" angle of the
main exptab by -1. Hence the deliberately left multiplication by -1 at
the end.
FATE passes, and performance reports on other platforms/CPUs are
welcome.
Performance comparisons:
iMDCT, PFA:
101127 decicycles in speed, 32765 runs, 3 skips
iMDCT, Old:
211022 decicycles in speed, 32768 runs, 0 skips
Standalone FFT, 300000 transforms of size 960:
PFA Old FFT kiss_fft libfftw3f
3.659695s, 15.726912s, 13.300789s, 1.182222s
Being only 3x slower than libfftw3f is a big achievement by itself.
There appears to be something capping the performance in the iMDCT side
of things, possibly during the pre-stage reindexing. However, it is
certainly fast enough for now.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Prep work for the next commit, which will add a new FFT algorithm
which makes the iMDCT over 3x faster than it is currently (standalone,
the FFT is with some framesizes over 10x faster).
The new FFT algorithm uses the already thouroughly SIMD'd power of two
FFT which already has SIMD for AArch64, so users of that platform will
still see an improvement.
The previous FFT+SIMD was barely 2.5x faster than the C versions on these
platforms.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
The libopus encoder does the same thing and its better than
keeping track of when the empty flush frames appear.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Performance improvements:
quant_bands:
with: 681 decicycles in quant_bands, 8388453 runs, 155 skips
without: 1190 decicycles in quant_bands, 8388386 runs, 222 skips
Around 42% for the function
Twoloop coder:
abs_pow34:
with/without: 7.82s/8.17s
Around 4% for the entire encoder
Both:
with/without: 7.15s/8.17s
Around 12% for the entire encoder
Fast coder:
abs_pow34:
with/without: 3.40s/3.77s
Around 10% for the entire encoder
Both:
with/without: 3.02s/3.77s
Around 20% faster for the entire encoder
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Tested-by: Michael Niedermayer <michael@niedermayer.cc>
Reviewed-by: James Almer <jamrial@gmail.com>
Using lfg was an overkill in this case where the random numbers
were only used for encoder descisions. Should increase result
uniformity between different FPUs and gives a slight speedup.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This version has had much testing so there's little point in keeping it
maked as experimental.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Does nothing fancy but still sounds very decent at 128kbps.
Still room to improve by bringing in the low pass and PNS management
from the main big twoloop which should improve its quality but not
sacrifice that much speed.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Adding a check for bits == 0 would still make Coverity misdetect this,
so just revert to the normal way of setting the residue to 0.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Can also be used in future cleanups since 99% of the time the leftover
appending will just append to an already empty residue.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Shifting by more than 63 bits is undefined behavior, athough any
compiler not returning 0 after shifting by any amount would be insane.
Found by Coverity, fixes CID1363959 and CID1363960
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
All HE-AAC samples with an LFE channel make this warning get spammed on
every frame. Turning off SBR for LFE channels makes sense (since it has
much less coefficients than normal channels do), so this error print is
of no value in this case.
It makes sense to keep the error in other cases, hence why it's still
around, degraded to warning severity since the decoder will still
attempt to decode without SBR.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This code wasn't removed when the functionality was moved inside
aacenc.c with commit 8005b6de4f
Fixes CID1361962
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
The problem is that with particularly complex images and especially at
high bit depths and 5-level transforms the coefficients would overflow,
causing huge artifacts to appear. This was discovered thanks to the fate
tests, which will have to be redone as this fixes a multitude of
problems and increases PSNR.
There is a slight performance drop associated with this change, making
the encoder slower by 1.15 times, however this is necessary in order to
avoid undefined behavior and overflows.
It would be worth to template the transforms to keep the performance for
8 bit images as 32 bit coefficients are unnecessary for that case, but
the primary use of the encoder is to encode video at 10 bits.
Reviewed-by: Christophe Gisquet <christophe.gisquet@gmail.com>
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Seems like clang might be miscompiling it and causing a signed integer overflow,
making a FATE test fail.
Doesn't seem to affect performance, it only runs on the ESC codebook.
Reviewed-by: Claudio Freire <klaussfreire@gmail.com>
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Until now, for formats which were in the spec but not in the encoder's
list of supported formats required the -strict -1 flag. This enables
support for all video formats which are specified, all the way from
QSIF525 to 8K.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
In some cases this caused the slice size rounding to generate invalid
slice sizes and overwrite some slices.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This was a leftover from before the slices were encoded in parallel.
Since the put_bits context is initialized per slice aligning it
aferwards is pointless.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit solves most of the crashes and issues with the encoder and
the bitrate setting. Now the encoder will always allocate the absolute
lowest amount of memory regardless of what the bitrate has been set to.
Therefore if a user inputs a very low bitrate the encoder will use the
maximum possible quantization (basically zero out all coefficients),
allocate a packet and encode it. There is no coupling between the
bitrate and the allocation size and so no crashes because the buffer
isn't large enough.
The maximum quantizer was raised to the size of the table now to both
keep the overshoot at ridiculous bitrates low and to improve quality
with higher bit depths (since the coefficients grow larger per transform
quantizing them to the same relative level requires larger quantization
indices).
Since the quantization index start follows the previous quantization
index for that slice, the quantization step was reduced to a static 1
to improve performance. Previously with quant/5 the step was usually
set to 0 upon start (and was later clipped to 1), that isn't a big change.
As the step size increases so does the amount of bits leftover and so
the redistribution algorithm has to iterate more and thus waste more
time.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This was a regression introduced by commit e7345abe05 which
enabled full use of the allocated packet but due to the overhead of
using field coding the buffer was too small and triggered warnings and
crashes.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
The fact that now all quantization indices costs are cached justifies
storing 20 more integers in a structure already allocated on heap.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit redistributes the leftover bytes amongst the top 150 slices
in terms of size (in the hopes that they'll be the ones pretty bitrate
starved).
A more perceptual method would probably need to cut bits off from slices
which don't need much, but that'll be implemented later.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Previously a global average was used. Using the previous quantizer
resulted in a fairly significant speedup as slice size selection settled
down quicker.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
16 bits were definitely not enough and caused artifacts to appear on
images at barely compressed images.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit moves the minimum bits per slice calculations outside of the
rate control function as it is identical for every slice.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Since coefficients differ only in the last bit when writing to the
bitstream it was possible to remove the sign from the tables, thus
halving them. Also now all quantization is done in the unsigned domain
as the sign is completely separate, which gets rid of the need to do
quantization on 32 bit signed integers.
Overall, this slightly speeds up the encoder depending on the machine.
The commit still generates bit-identical files as before the commit.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
The reference encoder limits it to 64, but testing revealed that there
is absolutely no difference for indices above 50 in amount of zeroed
coefficients.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit adds support for the (simple, allowed in the spec, but
inferior in quality) Haar wavelet transforms.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Similar to how the AAC encoder does it.
0 means the video's been compressed losslessly/almost losslessly
thoughout. Generally, the higher, the worse.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Results in dropping out in channels, usually on EIGHT_SHORT windows.
Will be reenabled once the cause has been investigated and a fix has
been made.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Takes into account whether there's pairing and if there's an LFE channel.
An SCE has more bits than CPE/2 since IS and M/S save quite a lot of bits
when channels are paired. And most of the SCEs we have are in surround
layouts which map it to the center channel, which usually carries all of
the dialogue and compression artifacts there are easily audiable.
Also refactors the init function a little bit and labels some parts of it.
Fixes bug #5233
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This was first reported on the mailing list in an earlier revision of this
encoder but was forgotten from the final commit.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit adds a new encoder capable of creating BBC/SMPTE Dirac/VC-2 HQ
profile files.
Dirac is a wavelet based codec created by the BBC a little more than 10
years ago. Since then, wavelets have mostly gone out of style as they
did not provide adequate encoding gains at lower bitrates. Dirac was a
fully featured video codec equipped with perceptual masking, support for
most popular pixel formats, interlacing, overlapped-block motion
compensation, and other features. It found new life after being stripped
of various features and standardized as the VC-2 codec by the SMPTE with
an extra profile, the HQ profile that this encoder supports, added.
The HQ profile was based off of the Low-Delay profile previously
existing in Dirac. The profile forbids DC prediction and arithmetic
coding to focus on high performance and low delay at higher bitrates.
The standard bitrates for this profile vary but generally 1:4
compression is expected (~525 Mbps vs the 2200 Mbps for uncompressed
1080p50). The codec only supports I-frames, hence the high bitrates.
The structure of this encoder is simple: do a DWT transform on the
entire image, split it into multiple slices (specified by the user) and
encode them in parallel. All of the slices are of the same size, making
rate control and threading very trivial. Although only in C, this encoder
is capable of 30 frames per second on an 4 core 8 threads Ivy Bridge.
A lookup table is used to encode most of the coefficients.
No code was used from the GSoC encoder from 2007 except for the 2
transform functions in diracenc_transforms.c. All other code was written
from scratch.
This encoder outperforms any other encoders in quality, usability and in
features. Other existing implementations do not support 4 level
transforms or 64x64 blocks (slices), which greatly increase compression.
As previously said, the codec is meant for broadcasting, hence support
for non-broadcasting image widths, heights, bit depths, aspect ratios,
etc. are limited by the "level". Although this codec supports a few
chroma subsamplings (420, 422, 444), signalling those is generally
outside the specifications of the level used (3) and the reference
decoder will outright refuse to read any image with such a flag
signalled (it only supports 1920x1080 yuv422p10). However, most
implementations will happily read files with alternate dimensions,
framerates and formats signalled.
Therefore, in order to encode files other than 1080p50 yuv422p10le, you
need to provide an "-strict -2" argument to the command line. The FFmpeg
decoder will happily read any files made with non-standard parameters,
dimensions and subsamplings, and so will other implementations. IMO this
should be "-strict -1", but I'll leave that up for discussion.
There are still plenty of stuff to implement, for instance 5 more
wavelet transforms are still in the specs and supported by the decoder.
The encoder can be lossless, given a high enough bitrate.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
The type of the option has been changed but the limit was apparently forgotten.
Some video codes can handle bitrates of over ~2.2 Gbps (like VC-2).
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Full range video had been broken by the introduction of the < 2U check
at the following line. The bitstream format kind of implies that the full
and limited ranges use different pix_fmts but that's incorrect since we
have the avctx->color_range flag. So adjust the pixel range index to
be mapped to the same pix_fmts as limited range index.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
The version structure in the main decoder context was not (and
apparently has never been) populated since it was added.
Still, having VC-2 break the existing Dirac Low Delay mode was odd and
easily avoidable had the specifications authors noticed/cared.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
The parser scans for "BBCD" to appear in the bitstream which indicate a
parse info header and once that happens, checks if the parse offsets are
sane. Since random BBCD strings might appear in the bitstream the parser
will emit a pointless warning if that happens.
This commit improves parsing by checking for a valid parse code as well
as keeping the original checks for valid parse offsets. The warnings
were removed as they serve no real purpose.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
The DSP lacked a function needed to convert signed to unsigned. This was
ignored when originally adding support and templating for bit depths
greater than 8. The 10 bit function was used for 12 bit pictures and
resulted in an improper conversion.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Has been marked for removal for over a month and has not been improved
or touched at all since it was implemented.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Too many crashes observed. Can't be helped until the autocorrelation
function is massively checked for sanity.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
'erf' is far from the best name for a variable and is not very
descriptive since the actual variable points to the comparitively best
IS phase. Therefore rename it to 'best'.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Since the parser was merged back almost 2 months ago this is the first
time the bitstream of the container has been updated.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
The type of last_frame_pb_count was chosen to be an int since overflow
is impossible (the spec says the maximum bits per frame is 6144 per
channel and the encoder checks for that).
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Reviewed-by: Paul B Mahol <onemda@gmail.com>
PSNR doesn't change as expected. The AAC spec doesn't really say
anything about how exactly to generate noise.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
With only 7 coefficients per short window at most the extra precision
makes a difference and seems to reduce crackling and stddev even
further.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
