The AC3 encoder used to be a separate library called "Aften", which
got merged into libavcodec (literally, SVN commits and all).
The merge preserved as much features from the library as possible.
The code had two versions - a fixed point version and a floating
point version. FFmpeg had floating point DSP code used by other
codecs, the AC3 decoder including, so the floating-point DSP was
simply replaced with FFmpeg's own functions.
However, FFmpeg had no fixed-point audio code at that point. So
the encoder brought along its own fixed-point DSP functions,
including a fixed-point MDCT.
The fixed-point MDCT itself is trivially just a float MDCT with a
different type and each multiply being a fixed-point multiply.
So over time, it got refactored, and the FFT used for all other codecs
was templated.
Due to design decisions at the time, the fixed-point version of the
encoder operates at 16-bits of precision. Although convenient, this,
even at the time, was inadequate and inefficient. The encoder is noisy,
does not produce output comparable to the float encoder, and even
rings at higher frequencies due to the badly approximated winow function.
Enter MIPS (owned by Imagination Technologies at the time). They wanted
quick fixed-point decoding on their FPUless cores. So they contributed
patches to template the AC3 decoder so it had both a fixed-point
and a floating-point version. They also did the same for the AAC decoder.
They however, used 32-bit samples. Not 16-bits. And we did not have
32-bit fixed-point DSP functions, including an MDCT. But instead of
templating our MDCT to output 3 versions (float, 32-bit fixed and 16-bit fixed),
they simply copy-pasted their own MDCT into ours, and completely
ifdeffed our own MDCT code out if a 32-bit fixed point MDCT was selected.
This is also the status quo nowadays - 2 separate MDCTs, one which
produces floating point and 16-bit fixed point versions, and one
sort-of integrated which produces 32-bit MDCT.
MIPS weren't all that interested in encoding, so they left the encoder
as-is, and they didn't care much about the ifdeffery, mess or quality - it's
not their problem.
So the MDCT/FFT code has always been a thorn in anyone looking to clean up
code's eye.
Backstory over. Internally AC3 operates on 25-bit fixed-point coefficients.
So for the floating point version, the encoder simply runs the float MDCT,
and converts the resulting coefficients to 25-bit fixed-point, as AC3 is inherently
a fixed-point codec. For the fixed-point version, the input is 16-bit samples,
so to maximize precision the frame samples are analyzed and the highest set
bit is detected via ac3_max_msb_abs_int16(), and the coefficients are then
scaled up via ac3_lshift_int16(), so the input for the FFT is always at least 14 bits,
computed in normalize_samples(). After FFT, the coefficients are scaled up to 25 bits.
This patch simply changes the encoder to accept 32-bit samples, reusing
the already well-optimized 32-bit MDCT code, allowing us to clean up and drop
a large part of a very messy code of ours, as well as prepare for the future lavu/tx
conversion. The coefficients are simply scaled down to 25 bits during windowing,
skipping 2 separate scalings, as the hacks to extend precision are simply no longer
necessary. There's no point in running the MDCT always at 32 bits when you're
going to drop 6 bits off anyway, the headroom is plenty, and the MDCT rounds
properly.
This also makes the encoder even slightly more accurate over the float version,
as there's no coefficient conversion step necessary.
SIZE SAVINGS:
ARM32:
HARDCODED TABLES:
BASE - 10709590
DROP DSP - 10702872 - diff: -6.56KiB
DROP MDCT - 10667932 - diff: -34.12KiB - both: -40.68KiB
DROP FFT - 10336652 - diff: -323.52KiB - all: -364.20KiB
SOFTCODED TABLES:
BASE - 9685096
DROP DSP - 9678378 - diff: -6.56KiB
DROP MDCT - 9643466 - diff: -34.09KiB - both: -40.65KiB
DROP FFT - 9573918 - diff: -67.92KiB - all: -108.57KiB
ARM64:
HARDCODED TABLES:
BASE - 14641112
DROP DSP - 14633806 - diff: -7.13KiB
DROP MDCT - 14604812 - diff: -28.31KiB - both: -35.45KiB
DROP FFT - 14286826 - diff: -310.53KiB - all: -345.98KiB
SOFTCODED TABLES:
BASE - 13636238
DROP DSP - 13628932 - diff: -7.13KiB
DROP MDCT - 13599866 - diff: -28.38KiB - both: -35.52KiB
DROP FFT - 13542080 - diff: -56.43KiB - all: -91.95KiB
x86:
HARDCODED TABLES:
BASE - 12367336
DROP DSP - 12354698 - diff: -12.34KiB
DROP MDCT - 12331024 - diff: -23.12KiB - both: -35.46KiB
DROP FFT - 12029788 - diff: -294.18KiB - all: -329.64KiB
SOFTCODED TABLES:
BASE - 11358094
DROP DSP - 11345456 - diff: -12.34KiB
DROP MDCT - 11321742 - diff: -23.16KiB - both: -35.50KiB
DROP FFT - 11276946 - diff: -43.75KiB - all: -79.25KiB
PERFORMANCE (10min random s32le):
ARM32 - before - 39.9x - 0m15.046s
ARM32 - after - 28.2x - 0m21.525s
Speed: -30%
ARM64 - before - 36.1x - 0m16.637s
ARM64 - after - 36.0x - 0m16.727s
Speed: -0.5%
x86 - before - 184x - 0m3.277s
x86 - after - 190x - 0m3.187s
Speed: +3%
The bug it was working seems to have been fixed.
This change causes ffmpeg to use the trim filter to implement
the -t option.
FATE tests are updated due to the more accurate handling of
the last packets.
This makes -t sample-accurate for audio and will allow further
simplication in the future.
Most of the FATE changes are due to audio now being sample accurate. In
some cases a video frame was incorrectly passed with the old code, while
its was over the limit.
Some of the FATE changes are due to off-by-one different rounding being used
(lrintf vs av_rescale_q).
Some fate changes are due to 1 audio frame less being encoded (the new variant seems
matching what qatar does and according to ffprobe its closer to the requested duration)
the mapchan feature sadly is lost in this commit because it depends on resampling
being done in ffmpeg.c which is now moved completely into the av filter layer
-async is broken after this commit, this will be fixed in subsequent commits
the new filter reconfiguration system is flawed and will drop a frame on each
parameter change which is why the nelly moser checksums need updating.
Conflicts:
ffmpeg.c
tests/ref/fate/smjpeg
* qatar/master: (26 commits)
adxenc: use AVCodec.encode2()
adxenc: Use the AVFrame in ADXContext for coded_frame
indeo4: fix out-of-bounds function call.
configure: Restructure help output.
configure: Internal-only components should not be command-line selectable.
vorbisenc: use AVCodec.encode2()
libvorbis: use AVCodec.encode2()
libopencore-amrnbenc: use AVCodec.encode2()
ra144enc: use AVCodec.encode2()
nellymoserenc: use AVCodec.encode2()
roqaudioenc: use AVCodec.encode2()
libspeex: use AVCodec.encode2()
libvo_amrwbenc: use AVCodec.encode2()
libvo_aacenc: use AVCodec.encode2()
wmaenc: use AVCodec.encode2()
mpegaudioenc: use AVCodec.encode2()
libmp3lame: use AVCodec.encode2()
libgsmenc: use AVCodec.encode2()
libfaac: use AVCodec.encode2()
g726enc: use AVCodec.encode2()
...
Conflicts:
configure
libavcodec/Makefile
libavcodec/ac3enc.c
libavcodec/adxenc.c
libavcodec/libgsm.c
libavcodec/libvorbis.c
libavcodec/vorbisenc.c
libavcodec/wmaenc.c
tests/ref/acodec/g722
tests/ref/lavf/asf
tests/ref/lavf/ffm
tests/ref/lavf/mkv
tests/ref/lavf/mpg
tests/ref/lavf/rm
tests/ref/lavf/ts
tests/ref/seek/lavf_asf
tests/ref/seek/lavf_ffm
tests/ref/seek/lavf_mkv
tests/ref/seek/lavf_mpg
tests/ref/seek/lavf_rm
tests/ref/seek/lavf_ts
Merged-by: Michael Niedermayer <michaelni@gmx.at>
Current code compares the desired recording time with InputStream.pts,
which has a very unclear meaning. Change the code to use actual
timestamps of the frames passed to the encoder.
In several tests, one less frame is encoded, which is more correct.
In the idroq test one more frame is encoded, which is again more
correct.
Behavior with stream copy should be unchanged.
This makes the AC3 encoder use the shared fixed-point MDCT rather
than its own implementation. The checksum changes are due to
different rounding in the MDCT.
Signed-off-by: Mans Rullgard <mans@mansr.com>
This increases the accuracy of coefficients, leading to improved quality.
Rescaling of the coefficients to full 25-bit accuracy is done rather than
offsetting the exponent values. This requires coefficient scaling to be done
before determining the rematrixing strategy. Also, the rematrixing strategy
calculation must use 64-bit math to prevent overflow due to the higher
precision coefficients.
This is to match the value in every (E-)AC-3 file from commercial sources.
It has a negligible effect on audio quality.
Signed-off-by: Mans Rullgard <mans@mansr.com>
This patch changes the exponent difference threshold in the exponent
strategy decision function of the AC-3 encoder. I tested lowering in
increments of 100. From 1000 down to 500 generally increased in quality
with each step, but 400 was generally much worse.
Signed-off-by: Mans Rullgard <mans@mansr.com>
This gives slightly better quality in PEAQ tests.
Code 3 gives a dBpb value of 2816 = -132dB (128 psd units = -6dB), which
corresponds to 22 bits. Since the exponents have an offset applied, the
16-bit source looks like 24-bit source to the bit allocation routine.
So using dBpb code=3 is a closer match to the exponent range.
Regression test refs updated for acodec-ac3, lavf-rm, and seek-ac3_rm.
Originally committed as revision 26144 to svn://svn.ffmpeg.org/ffmpeg/trunk
This avoids a 16-bit overflow in mdct512() due to a -32768 value in costab.
References updated for acodec-ac3, lavf-rm, and seek-ac3_rm tests.
Thanks to Måns Rullgård for finding the bug.
Originally committed as revision 26071 to svn://svn.ffmpeg.org/ffmpeg/trunk
With this change, the output is checked immediately after each test
has run. This means commands like "make regtest-mpeg2" can now be
used to run a single test and get meaningful results.
By default, make will abort if any test fails. To run all tests
regardless, use make -k.
Originally committed as revision 21254 to svn://svn.ffmpeg.org/ffmpeg/trunk
