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>
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>
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>
Make sure to properly enforce maximum scalefactor delta in PNS
and I/S, not as an absolute range but relative to the previous
and first scalefactor, the latter important since PNS uses an
implicit inicial scalefactor derived from the global gain, and
encoded in a different way, which makes a big difference in the
allowed range.
1. Fix sf_idx and band_type addressing to address only the first
subwindow in the group (others could hold garbage values)
2. Don't step on ms_mask when is_mask is set. I/S selection
already sets the ms_mask properly and shouldn't be overridden.
3. Use mid/sid cb/sf when computing coding error, as should be
since those are the cb/sfs that will eventually be set.
4. Fix distortion computation on multi-subwindow groups (was
subtracting the bits terms multiple times)
5. Clear ms_mask when one side uses PNS and the other doesn't.
When using PNS, ms_mask signals correlated noise, which can be
detected just like regular M/S detection, so we don't skip
noise bands, but when only one side uses PNS setting the flag
can confuse some encoders, so avoid that.
Fix possible SF delta violation that would cause an
eventual assertion failure in some corner cases (esp
on very low bitrates) when marking bands for PNS due
to misuse of the sf_delta utilities
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>
Fix OOB access in search_for_pns which was using
w2 outside the window group loop, and fix a typo
in which it was checking sf_idx instead of band_type
Reviewed-by: Andreas Cadhalpun <andreas.cadhalpun@googlemail.com>
Signed-off-by: Andreas Cadhalpun <Andreas.Cadhalpun@googlemail.com>
If minq is negative, the range of sf_idx can be larger than
SCALE_MAX_DIFF allows, causing assertion failures later in
encode_scale_factors.
Reviewed-by: Claudio Freire <klaussfreire@gmail.com>
Signed-off-by: Andreas Cadhalpun <Andreas.Cadhalpun@googlemail.com>
This patch does 4 things, all of which interact and thus it
woudln't be possible to commit them separately without causing
either quality regressions or assertion failures.
Fate comparison targets don't all reflect improvements in
quality, yet listening tests show substantially improved quality
and stability.
1. Increase SF range utilization.
The spec requires SF delta values to be constrained within the
range -60..60. The previous code was applying that range to
the whole SF array and not only the deltas of consecutive values,
because doing so requires smarter code: zeroing or otherwise
skipping a band may invalidate lots of SF choices.
This patch implements that logic to allow the coders to utilize
the full dynamic range of scalefactors, increasing quality quite
considerably, and fixing delta-SF-related assertion failures,
since now the limitation is enforced rather than asserted.
2. PNS tweaks
The previous modification makes big improvements in twoloop's
efficiency, and every time that happens PNS logic needs to be
tweaked accordingly to avoid it from stepping all over twoloop's
decisions. This patch includes modifications of the sort.
3. Account for lowpass cutoff during PSY analysis
The closer PSY's allocation is to final allocation the better
the quality is, and given these modifications, twoloop is now
very efficient at avoiding holes. Thus, to compute accurate
thresholds, PSY needs to account for the lowpass applied
implicitly during twoloop (by zeroing high bands).
This patch makes twoloop set the cutoff in psymodel's context
the first time it runs, and makes PSY account for it during
threshold computation, making PE and threshold computations
closer to the final allocation and thus achieving better
subjective quality.
4. Tweaks to RC lambda tracking loop in relation to PNS
Without this tweak some corner cases cause quality regressions.
Basically, lambda needs to react faster to overall bitrate
efficiency changes since now PNS can be quite successful in
enforcing maximum bitrates, when PSY allocates too many bits
to the lower bands, suppressing the signals RC logic uses to
lower lambda in those cases and causing aggressive PNS.
This tweak makes PNS much less aggressive, though it can still
use some further tweaks.
Also update MIPS specializations and adjust fuzz
Also in lavc/mips/aacpsy_mips.h: remove trailing whitespace
When both M/S coding and PNS are enabled, scalefactors
and coding books would be mistakenly clobbered when setting
the M/S flag on PNS'd bands. The flag needs to be set to
signal the generation of correlated noise, but the scalefactors,
coefficients and the coding books need to be kept intact.
Long Term Prediction allows for prediction of spectral coefficients
via the previously decoded time-dependent samples. This feature
works well with harmonic content 2 or more frames long, like speech,
human or non-human, piano music or any constant tones at very low
bitrates.
It should be noted that the current coder is highly efficient and
the rate control system is unable to encode files at extremely
low bitrates (less than 14kbps seems to be impossible) so this
extension isn't capable of optimum operation. Dramatic difference
is observable with some types of audio and speech but for the most
part the audiable differences are subtle. The spectrum looks better
however so the encoder is able to harvest the additional bits that
this feature provies, should the user choose to enable it. So
it's best to enable this feature only if encoding at the absolutely
lowest bitrate that the encoder is capable of.
This finalizes merging of the work in the patches in ticket #2686.
Improvements to twoloop and RC logic are extensive.
The non-exhaustive list of twoloop improvments includes:
- Tweaks to distortion limits on the RD optimization phase of twoloop
- Deeper search in twoloop
- PNS information marking to let twoloop decide when to use it
(turned out having the decision made separately wasn't working)
- Tonal band detection and priorization
- Better band energy conservation rules
- Strict hole avoidance
For rate control:
- Use psymodel's bit allocation to allow proper use of the bit
reservoir. Don't work against the bit reservoir by moving lambda
in the opposite direction when psymodel decides to allocate more/less
bits to a frame.
- Retry the encode if the effective rate lies outside a reasonable
margin of psymodel's allocation or the selected ABR.
- Log average lambda at the end. Useful info for everyone, but especially
for tuning of the various encoder constants that relate to lambda
feedback.
Psy:
- Do not apply lowpass with a FIR filter, instead just let the coder
zero bands above the cutoff. The FIR filter induces group delay,
and while zeroing bands causes ripple, it's lost in the quantization
noise.
- Experimental VBR bit allocation code
- Tweak automatic lowpass filter threshold to maximize audio bandwidth
at all bitrates while still providing acceptable, stable quality.
I/S:
- Phase decision fixes. Unrelated to #2686, but the bugs only surfaced
when the merge was finalized. Measure I/S band energy accounting for
phase, and prevent I/S and M/S from being applied both.
PNS:
- Avoid marking short bands with PNS when they're part of a window
group in which there's a large variation of energy from one window
to the next. PNS can't preserve those and the effect is extremely
noticeable.
M/S:
- Implement BMLD protection similar to the specified in
ISO-IEC/13818:7-2003, Appendix C Section 6.1. Since M/S decision
doesn't conform to section 6.1, a different method had to be
implemented, but should provide equivalent protection.
- Move the decision logic closer to the method specified in
ISO-IEC/13818:7-2003, Appendix C Section 6.1. Specifically,
make sure M/S needs less bits than dual stereo.
- Don't apply M/S in bands that are using I/S
Now, this of course needed adjustments in the compare targets and
fuzz factors of the AAC encoder's fate tests, but if wondering why
the targets go up (more distortion), consider the previous coder
was using too many bits on LF content (far more than required by
psy), and thus those signals will now be more distorted, not less.
The extra distortion isn't audible though, I carried extensive
ABX testing to make sure.
A very similar patch was also extensively tested by Kamendo2 in
the context of #2686.
This patch tweaks search_for_pns to be both more
aggressive and more careful when applying PNS. On
the one side, it will again try to use PNS on zero
(or effectively zero) bands. For this, both zeroes
and band_type have to be checked (some ZERO bands
aren't marked in zeroes). On the other side, a more
accurate rate-distortion measure avoids using PNS
where it would cause audible distortion.
Also fixed a small bug in the computation of freq
that caused PNS usage on low-frequency bands during
8-short windows. This allows re-enabling PNS during
8-short.
This patch refactors the AAC coders to reuse code
between the MIPS port and the regular, portable C code.
There were two main functions that had to use
hand-optimized versions of quantization code:
- search_for_quantizers_twoloop
- codebook_trellis_rate
Those two were split into their own template header
files so they can be inlined inside both the MIPS port
and the generic code. In each context, they'll link
to their specialized implementations, and thus be
optimized by the compiler.
This approach I believe is better than maintaining
several copies of each function. As past experience has
proven, having to keep those in sync was error prone.
In this way, they will remain in sync by default.
Also, an implementation of the dequantized output
argument for the optimized quantize_and_encode
functions is included in the patch. While the current
implementation of search_for_pred still isn't using
it, future iterations of main prediction probably will.
It should not imply any measurable performance hit while
not being used.
Makes more sense as users usually set the -cutoff option
to low pass filter the signal. The encoder will still over
shoot slightly when encoding normal coefficients however
that's normal.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit changes a few things about the noise substitution
logic:
- Brings back the quantization factor (reduced to 3) during
scalefactor index calculations.
- Rejects any zeroed bands. They should be inaudiable and it's
a waste transmitting the scalefactor indices for these.
- Uses swb_offsets instead of incrementing a 'start' with every
window group size.
- Rejects all PNS during short windows.
Overall improves quality. There was a plan to use the lfg system
to create the random numbers instead of using whatever the decoder
uses but for now this works fine. Entropy is far from important here.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit once again improves the PNS implementation by scaling the
thresholds with frequency. The thresholds get looser as the frequency
increases since higher frequencies are basically noise to human ears.
Also, this introduces quantization error correction for PNS. Should
the error be too much, no PNS will be used. The energy_ratio is used
to regulate the actual encoded PNS energy: if the generated PNS
energy is higher than the energy from the psy system, energy_ratio
is used to correct it so that hopefully once requantized and
transmitted the value in the decoder will be closer to what the
encoder has.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit rewrites the PNS implementation and significantly
improves sonic quality.
The previous implementation marked an incredibly big amount
of SFBs to predict when there was no need for this and this
resulted in quite a large amount of artifacts. Also the
quantization was incorrect (av_clip(4+log2f(...))) which
led to 3x the intensity for PNS values leading to even more
artifacts.
This commit rewrites the PNS search function and introduces
a major change: the PNS values are synthesized and are compared
to the current coefficients in addition to passing through
the revised checks to see whether PNS can be used.
This decreases distortions and makes the current PNS implementation
mainly focused on replacing any low-power non-zero bands as well
as adding any zeroed bands back.
The current encoder's performance is enough (especially with
IS) so PNS isn't really required except to fill in the occasional
few bands as well as extend any zeroed high frequency, so this
combination which is already enabled by default works
to get as much quality as it can within the bits allowed.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
The specifications explicitly state to use roundf() which
also rounds half-integer values away from zero.
This does fix a few IS artifacts.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit abandons the way the specifications state to
quantize the coefficients, makes use of the new LPC float
functions and is much better.
The original way of converting non-normalized float samples
to int32_t which out LPC system expects was wrong and it was
wrong to assume the coefficients that are generated are also
valid. It was essentially a full garbage-in, garbage-out
system and it definitely shows when looking at spectrals
and listening. The high frequencies were very overattenuated.
The new LPC function performs the analysis directly.
The specifications state to quantize the coefficients into
four bit index values using an asin() function which of course
had to have ugly ternary operators because the function turns
negative if the coefficients are negative which when encoding
causes invalid bitstream to get generated.
This deviates from this by using the direct TNS tables, which
are fairly small since you only have 4 bits at most for index
values. The LPC values are directly quantized against the tables
and are then used to perform filtering after the requantization,
which simply fetches the array values.
The end result is that TNS works much better now and doesn't
attenuate anything but the actual signal, e.g. TNS removes
quantization errors and does it's job correctly now.
It might be enabled by default soon since it doesn't hurt and
helps reduce nastyness at low bitrates.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit completely alters the algorithm of prediction.
The original commit which introduced prediction was completely
incorrect to even remotely care about what the actual coefficients
contain or whether any options were enabled. Not my actual fault.
This commit treats prediction the way the decoder does and expects
to do: like lossy encryption. Everything related to prediction now
happens at the very end but just before quantization and encoding
of coefficients. On the decoder side, prediction happens before
anything has had a chance to even access the coefficients.
Also the original implementation had problems because it actually
touched the band_type of special bands which already had their
scalefactor indices marked and it's a wonder the asserion wasn't
triggered when transmitting those.
Overall, this now drastically increases audio quality and you should
think about enabling it if you don't plan on playing anything encoded
on really old low power ultra-embedded devices since they might not
support decoding of prediction or AAC-Main. Though the specifications
were written ages ago and as times change so do the FLOPS.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit finalizes AAC-Main profile encoding support
by implementing all mandatory and optional tools available
in the specifications and current decoders.
The AAC-Main profile reqires that prediction support be
present (although decoders don't require it to be enabled)
for an encoder to be deemed capable of AAC-Main encoding,
as well as TNS, PNS and IS, all of which were implemented
with previous commits or earlier of this year.
Users are encouraged to test the new functionality using either
-profile:a aac_main or -aac_pred 1, the former of which will enable
the prediction option by default and the latter will change the
profile to AAC-Main. No other options shall be changed by enabling
either, it's currently up to the users to decide what's best.
The current implementation works best using M/S and/or IS,
so users are also welcome to enable both options and any
other options (TNS, PNS) for maximum quality.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit implements temporal noise shaping support in the
encoder, along with an -aac_tns option to toggle it on or off
(off by default for now). TNS will increase audio quality
and reduce quantization noise by applying a multitap FIR filter
across allowed coefficients and transmit side information to the
decoder so it could create an inverse filter.
Users are encouraged to test the new functionality by enabling
-aac_tns 1 during encoding.
No major bugs are observable at this time so after a while if no
new problems appear and if the current implementation is deemed
of high enough quality and stability it will be enabled by default,
possibly at the same time the encoder has its experimental flag
removed and becomes the standard aac encoder in ffmpeg.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit moves the intensity stereo implementation
out from aaccoder and into a separate file. This was
possible using the previous commits.
This commit also drastically improves the IS implementation
by making it phase invariant e.g. it will always choose the
best possible phase regardless of whether M/S coding is on
or most of the coefficients have identical phases.
This also increases the quality and reduces any distortions
introduced by enablind intensity stereo.
Users are encouraged to test it out using the -aac_is 1
parameter as it has always been.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit moves the quantizer to a separate header file.
This allows the quantizer to be used from a separate files outside
of aaccoder without having to put another function pointer and will
result in a slight speedup as the compiler can do more optimizations.
This is required for commits following.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit moves the resetting of special bands (above RESERVED_BT)
to the main frame encoding function rather than the way it was done
previously in their corresponding search_for_... functions.
The reason why special bands need to be reset is that while normal
bands get chosen for every frame by the coder (twoloop by default)
the coders do not touch any special sfbs and will therefore
make them persist throughout the file.
If we zero them out any bands left unmarked will be chosen by
the second part of the coder (the trellis function in aaccoder.c).
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit only changes the coding style to a saner way
of accessing coefficients (makes more sense to get the
memory address of a coefficients and start from there
rather than adding arbitrary numbers to offset a pointer).
Some compilers might detect an out of bounds access easier.
Also the way M/S and IS coefficients are calculated has been
changed, but should still have the same result (with the exception
that IS now applies from the normal coefficients rather than the
pristine ones, this is needed for upcoming commits).
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
As well as tables littered everywhere, functions were spread
out all across the encoder's files. This moves them to a single
place where they can be used by either the encoder's main files
or additional encoder files. Additionally, it changes the type
of some to 'inline' to enable us to simply put them in a header
file and possibly gain some speed due to compiler optimizations.
Signed-off-by: Claudio Freire <klaussfreire@gmail.com>
This commit moves any tables specific to the encoder from aacenc
and aaccoder to a separate file called 'aacenctab.c/.h'.
This was done as a clean up attempt as the encoder was filled with
tables pasted in between functions which made it confusing to follow
and track where each table and definition had been used.
This commit solves this by simply exporting the smaller tables out to
the aacenctab.h while the larger ones are compiled using aacenctab.c
and are referenced from the header file.
Signed-off-by: Claudio Freire <klaussfreire@gmail.com>
This commit removes a redundant argument from the functions in aaccoder.
The argument lambda was redundant as it was just a copy of s->lambda,
to which all functions have access to anyway. This cleans up the function
pointers a bit which is helpful as there are a lot of other search_for_*
functions under development and with them populated it gets messy.
Reviewed-by: Claudio Freire <klaussfreire@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Avoid clipping due to quantization noise to produce audible
artifacts, by detecting near-clipping signals and both attenuating
them a little and encoding escape-encoded bands (usually the
loudest) rounding towards zero instead of nearest, which tends to
decrease overall energy and thus clipping.
Currently fate tests measure numerical error so this change makes
tests using asynth (which are near clipping) report higher error
not less, because of window attenuation. Yet, they sound better,
not worse (albeit subtle, other samples aren't subtle at all).
Only measuring psychoacoustically weighted error would make for
a representative test, so that will be left for a future patch.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This commit moves the generation of ff_aac_pow34sf_tab[] out of the
encoder and into the table generator. The original commit log for
this table in 2011 actually mentions that it should be moved outside
but this never happened.
This is the first commit which cleans up the encoder a little.
Reviewed-by: Claudio Freire <klaussfreire@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This commit implements intensity stereo coding support
to the native aac encoder. This is a way to increase the efficiency
of the encoder by zeroing the right channel's spectral coefficients
(in a channel pair) and rederiving them in the decoder using information
from the scalefactor indices of special band types. This commit
confomrs to the official ISO 13818-7 specifications, although due to
their ambiguity certain deviations have been taken to ensure maximum
sound quality. This commit has been extensively tested and has shown
to not result in audiable audio artifacts unless in extreme cases.
This commit also adds an option, aac_is, which has the value of
0 by default. Intensity Stereo is part of the scalable aac profile
and is thus non-default.
The way IS coding works is that it rederives the right channel's
spectral coefficients from the left channel via the scalefactor
index values left in the right channel. Since an entire band's
spectral coefficients do not need to be coded, the encoder's
efficiency jumps up and it unzeroes some high frequency values
which it previously did not have enough bits to encode. That way
less information is lost than the information lost by rederiving
the spectral coefficients with some error. This is why the
filesize of files encoded with IS do not decrease significantly.
Users wishing that IS coding should reduce filesize are expected
to reduce their encoding bitrates appropriately.
This is V2 of the commit. The old version did not mark ms_mask as
0 since M/S and IS coding are incompactible, which resulted in
distortions with M/S coding enabled. This version also improves
phase detection by measuring it for every spectral coefficient in
the band and using a simple majority rule to determine whether the
coefficients are in or out of phase. Also, the energy values per
spectral coefficient were changed as to reflect the
official specifications.
Reviewed-by: Claudio Freire <klaussfreire@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
This commit finalizes the PNS implementation previously added to the encoder
by moving it to a seperate function search_for_pns() and thus making it
coder-generic. This new implementation makes use of the spread field of
the psy bands and the lambda quality feedback paremeter. The spread of the
spectrum in a band prevents PNS from being used excessively and thus preserve
more phase information in high frequencies. The lambda parameter allows
the number of PNS-marked bands to vary based on the lambda parameter and the
amount of bits available, making better choices on which bands are to be marked
as noise. Comparisons with the previous PNS implementation can be found
here: https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/
This is V2 of the patch, the changes from the previous version being that this
version uses the new band->spread metric from aacpsy and normalizes the
energy using the group size. These changes were suggested by Claudio Freire
on the mailing list. Another change is the use of lambda to alter the
frequency threshold. This change makes the actual threshold frequencies
vary between +-2Khz of what's specified, depending on frame encoding performance.
Reviewed-by: Claudio Freire <klaussfreire@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
This commit undoes commit c5d4f87e81
and removes PNS band marking from the twoloop coder, which has
been reimplemented in a better way in this series of patches.
Reviewed-by: Claudio Freire <klaussfreire@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
This commit enables the function added with commit 7c10b87 and uses that
new function for setting any special scalefactor indices. This commit does
not change the behaviour of the encoder since no bands are being marked as
either NOISE_BT(due to the previous PNS implementation removed in the
previous commit) or INTENSITY_BT2/INTENSITY_BT.
Reviewed-by: Claudio Freire <klaussfreire@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
There were some mistakes in the code for M/S stereo, this commit fixes them.
The start variable was not being reset for every window and every access to
the coefficients was incorrect as well. This fixes that by properly
addressing the coefficients using both windows and setting the start on every window to zero.
Reviewed-by: Claudio Freire <klaussfreire@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
This commit adds support for the coding of intensity stereo scalefactor indices.
It does not do any marking of such bands and as such does no functional changes
to the encoder. It removes any old twoloop specific code for PNS and moves it
into a seperate function which handles setting of scalefactor indices for
PNS and IS bands.
Reviewed-by: Claudio Freire <klaussfreire@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
