|
|
|
|
/*
|
|
|
|
|
* AAC coefficients encoder
|
|
|
|
|
* Copyright (C) 2008-2009 Konstantin Shishkov
|
|
|
|
|
*
|
|
|
|
|
* This file is part of FFmpeg.
|
|
|
|
|
*
|
|
|
|
|
* FFmpeg is free software; you can redistribute it and/or
|
|
|
|
|
* modify it under the terms of the GNU Lesser General Public
|
|
|
|
|
* License as published by the Free Software Foundation; either
|
|
|
|
|
* version 2.1 of the License, or (at your option) any later version.
|
|
|
|
|
*
|
|
|
|
|
* FFmpeg is distributed in the hope that it will be useful,
|
|
|
|
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
|
|
* Lesser General Public License for more details.
|
|
|
|
|
*
|
|
|
|
|
* You should have received a copy of the GNU Lesser General Public
|
|
|
|
|
* License along with FFmpeg; if not, write to the Free Software
|
|
|
|
|
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* @file
|
|
|
|
|
* AAC coefficients encoder
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
/***********************************
|
|
|
|
|
* TODOs:
|
|
|
|
|
* speedup quantizer selection
|
|
|
|
|
* add sane pulse detection
|
|
|
|
|
***********************************/
|
|
|
|
|
|
|
|
|
|
#include "libavutil/libm.h" // brought forward to work around cygwin header breakage
|
|
|
|
|
|
|
|
|
|
#include <float.h>
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
|
|
|
|
|
#include "libavutil/mathematics.h"
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
#include "mathops.h"
|
|
|
|
|
#include "avcodec.h"
|
|
|
|
|
#include "put_bits.h"
|
|
|
|
|
#include "aac.h"
|
|
|
|
|
#include "aacenc.h"
|
|
|
|
|
#include "aactab.h"
|
|
|
|
|
#include "aacenctab.h"
|
|
|
|
|
#include "aacenc_utils.h"
|
|
|
|
|
#include "aacenc_quantization.h"
|
|
|
|
|
|
|
|
|
|
#include "aacenc_is.h"
|
|
|
|
|
#include "aacenc_tns.h"
|
|
|
|
|
#include "aacenc_ltp.h"
|
|
|
|
|
#include "aacenc_pred.h"
|
|
|
|
|
|
|
|
|
|
#include "libavcodec/aaccoder_twoloop.h"
|
|
|
|
|
|
|
|
|
|
/* Parameter of f(x) = a*(lambda/100), defines the maximum fourier spread
|
|
|
|
|
* beyond which no PNS is used (since the SFBs contain tone rather than noise) */
|
AAC encoder: improve SF range utilization
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
9 years ago
|
|
|
#define NOISE_SPREAD_THRESHOLD 0.9f
|
|
|
|
|
|
|
|
|
|
/* Parameter of f(x) = a*(100/lambda), defines how much PNS is allowed to
|
|
|
|
|
* replace low energy non zero bands */
|
|
|
|
|
#define NOISE_LAMBDA_REPLACE 1.948f
|
aaccoder: Implement Perceptual Noise Substitution for AAC
This commit implements the perceptual noise substitution AAC extension. This is a proof of concept
implementation, and as such, is not enabled by default. This is the fourth revision of this patch,
made after some problems were noted out. Any changes made since the previous revisions have been indicated.
In order to extend the encoder to use an additional codebook, the array holding each codebook has been
modified with two additional entries - 13 for the NOISE_BT codebook and 12 which has a placeholder function.
The cost system was modified to skip the 12th entry using an array to map the input and outputs it has. It
also does not accept using the 13th codebook for any band which is not marked as containing noise, thereby
restricting its ability to arbitrarily choose it for bands. The use of arrays allows the system to be easily
extended to allow for intensity stereo encoding, which uses additional codebooks.
The 12th entry in the codebook function array points to a function which stops the execution of the program
by calling an assert with an always 'false' argument. It was pointed out in an email discussion with
Claudio Freire that having a 'NULL' entry can result in unexpected behaviour and could be used as
a security hole. There is no danger of this function being called during encoding due to the codebook maps introduced.
Another change from version 1 of the patch is the addition of an argument to the encoder, '-aac_pns' to
enable and disable the PNS. This currently defaults to disable the PNS, as it is experimental.
The switch will be removed in the future, when the algorithm to select noise bands has been improved.
The current algorithm simply compares the energy to the threshold (multiplied by a constant) to determine
noise, however the FFPsyBand structure contains other useful figures to determine which bands carry noise more accurately.
Some of the sample files provided triggered an assertion when the parameter to tune the threshold was set to
a value of '2.2'. Claudio Freire reported the problem's source could be in the range of the scalefactor
indices for noise and advised to measure the minimal index and clip anything above the maximum allowed
value. This has been implemented and all the files which used to trigger the asserion now encode without error.
The third revision of the problem also removes unneded variabes and comparisons. All of them were
redundant and were of little use for when the PNS implementation would be extended.
The fourth revision moved the clipping of the noise scalefactors outside the second loop of the two-loop
algorithm in order to prevent their redundant calculations. Also, freq_mult has been changed to a float
variable due to the fact that rounding errors can prove to be a problem at low frequencies.
Considerations were taken whether the entire expression could be evaluated inside the expression
, but in the end it was decided that it would be for the best if just the type of the variable were
to change. Claudio Freire reported the two problems. There is no change of functionality
(except for low sampling frequencies) so the spectral demonstrations at the end of this commit's message were not updated.
Finally, the way energy values are converted to scalefactor indices has changed since the first commit,
as per the suggestion of Claudio Freire. This may still have some drawbacks, but unlike the first commit
it works without having redundant offsets and outputs what the decoder expects to have, in terms of the
ranges of the scalefactor indices.
Some spectral comparisons: https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/Original.png (original),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS_NO.png (encoded without PNS),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS1.2.png (encoded with PNS, const = 1.2),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/Difference1.png (spectral difference).
The constant is the value which multiplies the threshold when it gets compared to the energy, larger
values means more noise will be substituded by PNS values. Example when const = 2.2:
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS_2.2.png
Reviewed-by: Claudio Freire <klaussfreire@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
10 years ago
|
|
|
|
|
|
|
|
#include "libavcodec/aaccoder_trellis.h"
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* structure used in optimal codebook search
|
|
|
|
|
*/
|
|
|
|
|
typedef struct BandCodingPath {
|
|
|
|
|
int prev_idx; ///< pointer to the previous path point
|
|
|
|
|
float cost; ///< path cost
|
|
|
|
|
int run;
|
|
|
|
|
} BandCodingPath;
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Encode band info for single window group bands.
|
|
|
|
|
*/
|
|
|
|
|
static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
|
|
|
|
|
int win, int group_len, const float lambda)
|
|
|
|
|
{
|
|
|
|
|
BandCodingPath path[120][CB_TOT_ALL];
|
|
|
|
|
int w, swb, cb, start, size;
|
|
|
|
|
int i, j;
|
|
|
|
|
const int max_sfb = sce->ics.max_sfb;
|
|
|
|
|
const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
|
|
|
|
|
const int run_esc = (1 << run_bits) - 1;
|
|
|
|
|
int idx, ppos, count;
|
|
|
|
|
int stackrun[120], stackcb[120], stack_len;
|
|
|
|
|
float next_minrd = INFINITY;
|
|
|
|
|
int next_mincb = 0;
|
|
|
|
|
|
|
|
|
|
s->abs_pow34(s->scoefs, sce->coeffs, 1024);
|
|
|
|
|
start = win*128;
|
|
|
|
|
for (cb = 0; cb < CB_TOT_ALL; cb++) {
|
|
|
|
|
path[0][cb].cost = 0.0f;
|
|
|
|
|
path[0][cb].prev_idx = -1;
|
|
|
|
|
path[0][cb].run = 0;
|
|
|
|
|
}
|
|
|
|
|
for (swb = 0; swb < max_sfb; swb++) {
|
|
|
|
|
size = sce->ics.swb_sizes[swb];
|
|
|
|
|
if (sce->zeroes[win*16 + swb]) {
|
|
|
|
|
for (cb = 0; cb < CB_TOT_ALL; cb++) {
|
|
|
|
|
path[swb+1][cb].prev_idx = cb;
|
|
|
|
|
path[swb+1][cb].cost = path[swb][cb].cost;
|
|
|
|
|
path[swb+1][cb].run = path[swb][cb].run + 1;
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
float minrd = next_minrd;
|
|
|
|
|
int mincb = next_mincb;
|
|
|
|
|
next_minrd = INFINITY;
|
|
|
|
|
next_mincb = 0;
|
|
|
|
|
for (cb = 0; cb < CB_TOT_ALL; cb++) {
|
|
|
|
|
float cost_stay_here, cost_get_here;
|
|
|
|
|
float rd = 0.0f;
|
|
|
|
|
if (cb >= 12 && sce->band_type[win*16+swb] < aac_cb_out_map[cb] ||
|
|
|
|
|
cb < aac_cb_in_map[sce->band_type[win*16+swb]] && sce->band_type[win*16+swb] > aac_cb_out_map[cb]) {
|
|
|
|
|
path[swb+1][cb].prev_idx = -1;
|
|
|
|
|
path[swb+1][cb].cost = INFINITY;
|
|
|
|
|
path[swb+1][cb].run = path[swb][cb].run + 1;
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
for (w = 0; w < group_len; w++) {
|
|
|
|
|
FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(win+w)*16+swb];
|
|
|
|
|
rd += quantize_band_cost(s, &sce->coeffs[start + w*128],
|
|
|
|
|
&s->scoefs[start + w*128], size,
|
aaccoder: Implement Perceptual Noise Substitution for AAC
This commit implements the perceptual noise substitution AAC extension. This is a proof of concept
implementation, and as such, is not enabled by default. This is the fourth revision of this patch,
made after some problems were noted out. Any changes made since the previous revisions have been indicated.
In order to extend the encoder to use an additional codebook, the array holding each codebook has been
modified with two additional entries - 13 for the NOISE_BT codebook and 12 which has a placeholder function.
The cost system was modified to skip the 12th entry using an array to map the input and outputs it has. It
also does not accept using the 13th codebook for any band which is not marked as containing noise, thereby
restricting its ability to arbitrarily choose it for bands. The use of arrays allows the system to be easily
extended to allow for intensity stereo encoding, which uses additional codebooks.
The 12th entry in the codebook function array points to a function which stops the execution of the program
by calling an assert with an always 'false' argument. It was pointed out in an email discussion with
Claudio Freire that having a 'NULL' entry can result in unexpected behaviour and could be used as
a security hole. There is no danger of this function being called during encoding due to the codebook maps introduced.
Another change from version 1 of the patch is the addition of an argument to the encoder, '-aac_pns' to
enable and disable the PNS. This currently defaults to disable the PNS, as it is experimental.
The switch will be removed in the future, when the algorithm to select noise bands has been improved.
The current algorithm simply compares the energy to the threshold (multiplied by a constant) to determine
noise, however the FFPsyBand structure contains other useful figures to determine which bands carry noise more accurately.
Some of the sample files provided triggered an assertion when the parameter to tune the threshold was set to
a value of '2.2'. Claudio Freire reported the problem's source could be in the range of the scalefactor
indices for noise and advised to measure the minimal index and clip anything above the maximum allowed
value. This has been implemented and all the files which used to trigger the asserion now encode without error.
The third revision of the problem also removes unneded variabes and comparisons. All of them were
redundant and were of little use for when the PNS implementation would be extended.
The fourth revision moved the clipping of the noise scalefactors outside the second loop of the two-loop
algorithm in order to prevent their redundant calculations. Also, freq_mult has been changed to a float
variable due to the fact that rounding errors can prove to be a problem at low frequencies.
Considerations were taken whether the entire expression could be evaluated inside the expression
, but in the end it was decided that it would be for the best if just the type of the variable were
to change. Claudio Freire reported the two problems. There is no change of functionality
(except for low sampling frequencies) so the spectral demonstrations at the end of this commit's message were not updated.
Finally, the way energy values are converted to scalefactor indices has changed since the first commit,
as per the suggestion of Claudio Freire. This may still have some drawbacks, but unlike the first commit
it works without having redundant offsets and outputs what the decoder expects to have, in terms of the
ranges of the scalefactor indices.
Some spectral comparisons: https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/Original.png (original),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS_NO.png (encoded without PNS),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS1.2.png (encoded with PNS, const = 1.2),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/Difference1.png (spectral difference).
The constant is the value which multiplies the threshold when it gets compared to the energy, larger
values means more noise will be substituded by PNS values. Example when const = 2.2:
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS_2.2.png
Reviewed-by: Claudio Freire <klaussfreire@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
10 years ago
|
|
|
sce->sf_idx[(win+w)*16+swb], aac_cb_out_map[cb],
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
lambda / band->threshold, INFINITY, NULL, NULL, 0);
|
|
|
|
|
}
|
|
|
|
|
cost_stay_here = path[swb][cb].cost + rd;
|
|
|
|
|
cost_get_here = minrd + rd + run_bits + 4;
|
|
|
|
|
if ( run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
|
|
|
|
|
!= run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
|
|
|
|
|
cost_stay_here += run_bits;
|
|
|
|
|
if (cost_get_here < cost_stay_here) {
|
|
|
|
|
path[swb+1][cb].prev_idx = mincb;
|
|
|
|
|
path[swb+1][cb].cost = cost_get_here;
|
|
|
|
|
path[swb+1][cb].run = 1;
|
|
|
|
|
} else {
|
|
|
|
|
path[swb+1][cb].prev_idx = cb;
|
|
|
|
|
path[swb+1][cb].cost = cost_stay_here;
|
|
|
|
|
path[swb+1][cb].run = path[swb][cb].run + 1;
|
|
|
|
|
}
|
|
|
|
|
if (path[swb+1][cb].cost < next_minrd) {
|
|
|
|
|
next_minrd = path[swb+1][cb].cost;
|
|
|
|
|
next_mincb = cb;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
start += sce->ics.swb_sizes[swb];
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
//convert resulting path from backward-linked list
|
|
|
|
|
stack_len = 0;
|
|
|
|
|
idx = 0;
|
|
|
|
|
for (cb = 1; cb < CB_TOT_ALL; cb++)
|
|
|
|
|
if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
|
|
|
|
|
idx = cb;
|
|
|
|
|
ppos = max_sfb;
|
|
|
|
|
while (ppos > 0) {
|
|
|
|
|
av_assert1(idx >= 0);
|
|
|
|
|
cb = idx;
|
|
|
|
|
stackrun[stack_len] = path[ppos][cb].run;
|
|
|
|
|
stackcb [stack_len] = cb;
|
|
|
|
|
idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
|
|
|
|
|
ppos -= path[ppos][cb].run;
|
|
|
|
|
stack_len++;
|
|
|
|
|
}
|
|
|
|
|
//perform actual band info encoding
|
|
|
|
|
start = 0;
|
|
|
|
|
for (i = stack_len - 1; i >= 0; i--) {
|
aaccoder: Implement Perceptual Noise Substitution for AAC
This commit implements the perceptual noise substitution AAC extension. This is a proof of concept
implementation, and as such, is not enabled by default. This is the fourth revision of this patch,
made after some problems were noted out. Any changes made since the previous revisions have been indicated.
In order to extend the encoder to use an additional codebook, the array holding each codebook has been
modified with two additional entries - 13 for the NOISE_BT codebook and 12 which has a placeholder function.
The cost system was modified to skip the 12th entry using an array to map the input and outputs it has. It
also does not accept using the 13th codebook for any band which is not marked as containing noise, thereby
restricting its ability to arbitrarily choose it for bands. The use of arrays allows the system to be easily
extended to allow for intensity stereo encoding, which uses additional codebooks.
The 12th entry in the codebook function array points to a function which stops the execution of the program
by calling an assert with an always 'false' argument. It was pointed out in an email discussion with
Claudio Freire that having a 'NULL' entry can result in unexpected behaviour and could be used as
a security hole. There is no danger of this function being called during encoding due to the codebook maps introduced.
Another change from version 1 of the patch is the addition of an argument to the encoder, '-aac_pns' to
enable and disable the PNS. This currently defaults to disable the PNS, as it is experimental.
The switch will be removed in the future, when the algorithm to select noise bands has been improved.
The current algorithm simply compares the energy to the threshold (multiplied by a constant) to determine
noise, however the FFPsyBand structure contains other useful figures to determine which bands carry noise more accurately.
Some of the sample files provided triggered an assertion when the parameter to tune the threshold was set to
a value of '2.2'. Claudio Freire reported the problem's source could be in the range of the scalefactor
indices for noise and advised to measure the minimal index and clip anything above the maximum allowed
value. This has been implemented and all the files which used to trigger the asserion now encode without error.
The third revision of the problem also removes unneded variabes and comparisons. All of them were
redundant and were of little use for when the PNS implementation would be extended.
The fourth revision moved the clipping of the noise scalefactors outside the second loop of the two-loop
algorithm in order to prevent their redundant calculations. Also, freq_mult has been changed to a float
variable due to the fact that rounding errors can prove to be a problem at low frequencies.
Considerations were taken whether the entire expression could be evaluated inside the expression
, but in the end it was decided that it would be for the best if just the type of the variable were
to change. Claudio Freire reported the two problems. There is no change of functionality
(except for low sampling frequencies) so the spectral demonstrations at the end of this commit's message were not updated.
Finally, the way energy values are converted to scalefactor indices has changed since the first commit,
as per the suggestion of Claudio Freire. This may still have some drawbacks, but unlike the first commit
it works without having redundant offsets and outputs what the decoder expects to have, in terms of the
ranges of the scalefactor indices.
Some spectral comparisons: https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/Original.png (original),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS_NO.png (encoded without PNS),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS1.2.png (encoded with PNS, const = 1.2),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/Difference1.png (spectral difference).
The constant is the value which multiplies the threshold when it gets compared to the energy, larger
values means more noise will be substituded by PNS values. Example when const = 2.2:
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS_2.2.png
Reviewed-by: Claudio Freire <klaussfreire@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
10 years ago
|
|
|
cb = aac_cb_out_map[stackcb[i]];
|
|
|
|
|
put_bits(&s->pb, 4, cb);
|
|
|
|
|
count = stackrun[i];
|
aaccoder: Implement Perceptual Noise Substitution for AAC
This commit implements the perceptual noise substitution AAC extension. This is a proof of concept
implementation, and as such, is not enabled by default. This is the fourth revision of this patch,
made after some problems were noted out. Any changes made since the previous revisions have been indicated.
In order to extend the encoder to use an additional codebook, the array holding each codebook has been
modified with two additional entries - 13 for the NOISE_BT codebook and 12 which has a placeholder function.
The cost system was modified to skip the 12th entry using an array to map the input and outputs it has. It
also does not accept using the 13th codebook for any band which is not marked as containing noise, thereby
restricting its ability to arbitrarily choose it for bands. The use of arrays allows the system to be easily
extended to allow for intensity stereo encoding, which uses additional codebooks.
The 12th entry in the codebook function array points to a function which stops the execution of the program
by calling an assert with an always 'false' argument. It was pointed out in an email discussion with
Claudio Freire that having a 'NULL' entry can result in unexpected behaviour and could be used as
a security hole. There is no danger of this function being called during encoding due to the codebook maps introduced.
Another change from version 1 of the patch is the addition of an argument to the encoder, '-aac_pns' to
enable and disable the PNS. This currently defaults to disable the PNS, as it is experimental.
The switch will be removed in the future, when the algorithm to select noise bands has been improved.
The current algorithm simply compares the energy to the threshold (multiplied by a constant) to determine
noise, however the FFPsyBand structure contains other useful figures to determine which bands carry noise more accurately.
Some of the sample files provided triggered an assertion when the parameter to tune the threshold was set to
a value of '2.2'. Claudio Freire reported the problem's source could be in the range of the scalefactor
indices for noise and advised to measure the minimal index and clip anything above the maximum allowed
value. This has been implemented and all the files which used to trigger the asserion now encode without error.
The third revision of the problem also removes unneded variabes and comparisons. All of them were
redundant and were of little use for when the PNS implementation would be extended.
The fourth revision moved the clipping of the noise scalefactors outside the second loop of the two-loop
algorithm in order to prevent their redundant calculations. Also, freq_mult has been changed to a float
variable due to the fact that rounding errors can prove to be a problem at low frequencies.
Considerations were taken whether the entire expression could be evaluated inside the expression
, but in the end it was decided that it would be for the best if just the type of the variable were
to change. Claudio Freire reported the two problems. There is no change of functionality
(except for low sampling frequencies) so the spectral demonstrations at the end of this commit's message were not updated.
Finally, the way energy values are converted to scalefactor indices has changed since the first commit,
as per the suggestion of Claudio Freire. This may still have some drawbacks, but unlike the first commit
it works without having redundant offsets and outputs what the decoder expects to have, in terms of the
ranges of the scalefactor indices.
Some spectral comparisons: https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/Original.png (original),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS_NO.png (encoded without PNS),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS1.2.png (encoded with PNS, const = 1.2),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/Difference1.png (spectral difference).
The constant is the value which multiplies the threshold when it gets compared to the energy, larger
values means more noise will be substituded by PNS values. Example when const = 2.2:
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS_2.2.png
Reviewed-by: Claudio Freire <klaussfreire@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
10 years ago
|
|
|
memset(sce->zeroes + win*16 + start, !cb, count);
|
|
|
|
|
//XXX: memset when band_type is also uint8_t
|
|
|
|
|
for (j = 0; j < count; j++) {
|
aaccoder: Implement Perceptual Noise Substitution for AAC
This commit implements the perceptual noise substitution AAC extension. This is a proof of concept
implementation, and as such, is not enabled by default. This is the fourth revision of this patch,
made after some problems were noted out. Any changes made since the previous revisions have been indicated.
In order to extend the encoder to use an additional codebook, the array holding each codebook has been
modified with two additional entries - 13 for the NOISE_BT codebook and 12 which has a placeholder function.
The cost system was modified to skip the 12th entry using an array to map the input and outputs it has. It
also does not accept using the 13th codebook for any band which is not marked as containing noise, thereby
restricting its ability to arbitrarily choose it for bands. The use of arrays allows the system to be easily
extended to allow for intensity stereo encoding, which uses additional codebooks.
The 12th entry in the codebook function array points to a function which stops the execution of the program
by calling an assert with an always 'false' argument. It was pointed out in an email discussion with
Claudio Freire that having a 'NULL' entry can result in unexpected behaviour and could be used as
a security hole. There is no danger of this function being called during encoding due to the codebook maps introduced.
Another change from version 1 of the patch is the addition of an argument to the encoder, '-aac_pns' to
enable and disable the PNS. This currently defaults to disable the PNS, as it is experimental.
The switch will be removed in the future, when the algorithm to select noise bands has been improved.
The current algorithm simply compares the energy to the threshold (multiplied by a constant) to determine
noise, however the FFPsyBand structure contains other useful figures to determine which bands carry noise more accurately.
Some of the sample files provided triggered an assertion when the parameter to tune the threshold was set to
a value of '2.2'. Claudio Freire reported the problem's source could be in the range of the scalefactor
indices for noise and advised to measure the minimal index and clip anything above the maximum allowed
value. This has been implemented and all the files which used to trigger the asserion now encode without error.
The third revision of the problem also removes unneded variabes and comparisons. All of them were
redundant and were of little use for when the PNS implementation would be extended.
The fourth revision moved the clipping of the noise scalefactors outside the second loop of the two-loop
algorithm in order to prevent their redundant calculations. Also, freq_mult has been changed to a float
variable due to the fact that rounding errors can prove to be a problem at low frequencies.
Considerations were taken whether the entire expression could be evaluated inside the expression
, but in the end it was decided that it would be for the best if just the type of the variable were
to change. Claudio Freire reported the two problems. There is no change of functionality
(except for low sampling frequencies) so the spectral demonstrations at the end of this commit's message were not updated.
Finally, the way energy values are converted to scalefactor indices has changed since the first commit,
as per the suggestion of Claudio Freire. This may still have some drawbacks, but unlike the first commit
it works without having redundant offsets and outputs what the decoder expects to have, in terms of the
ranges of the scalefactor indices.
Some spectral comparisons: https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/Original.png (original),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS_NO.png (encoded without PNS),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS1.2.png (encoded with PNS, const = 1.2),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/Difference1.png (spectral difference).
The constant is the value which multiplies the threshold when it gets compared to the energy, larger
values means more noise will be substituded by PNS values. Example when const = 2.2:
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS_2.2.png
Reviewed-by: Claudio Freire <klaussfreire@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
10 years ago
|
|
|
sce->band_type[win*16 + start] = cb;
|
|
|
|
|
start++;
|
|
|
|
|
}
|
|
|
|
|
while (count >= run_esc) {
|
|
|
|
|
put_bits(&s->pb, run_bits, run_esc);
|
|
|
|
|
count -= run_esc;
|
|
|
|
|
}
|
|
|
|
|
put_bits(&s->pb, run_bits, count);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
typedef struct TrellisPath {
|
|
|
|
|
float cost;
|
|
|
|
|
int prev;
|
|
|
|
|
} TrellisPath;
|
|
|
|
|
|
|
|
|
|
#define TRELLIS_STAGES 121
|
|
|
|
|
#define TRELLIS_STATES (SCALE_MAX_DIFF+1)
|
|
|
|
|
|
|
|
|
|
static void set_special_band_scalefactors(AACEncContext *s, SingleChannelElement *sce)
|
|
|
|
|
{
|
|
|
|
|
int w, g;
|
|
|
|
|
int prevscaler_n = -255, prevscaler_i = 0;
|
|
|
|
|
int bands = 0;
|
|
|
|
|
|
|
|
|
|
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
|
|
|
|
|
for (g = 0; g < sce->ics.num_swb; g++) {
|
|
|
|
|
if (sce->zeroes[w*16+g])
|
|
|
|
|
continue;
|
|
|
|
|
if (sce->band_type[w*16+g] == INTENSITY_BT || sce->band_type[w*16+g] == INTENSITY_BT2) {
|
|
|
|
|
sce->sf_idx[w*16+g] = av_clip(roundf(log2f(sce->is_ener[w*16+g])*2), -155, 100);
|
|
|
|
|
bands++;
|
|
|
|
|
} else if (sce->band_type[w*16+g] == NOISE_BT) {
|
|
|
|
|
sce->sf_idx[w*16+g] = av_clip(3+ceilf(log2f(sce->pns_ener[w*16+g])*2), -100, 155);
|
|
|
|
|
if (prevscaler_n == -255)
|
|
|
|
|
prevscaler_n = sce->sf_idx[w*16+g];
|
|
|
|
|
bands++;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (!bands)
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
/* Clip the scalefactor indices */
|
|
|
|
|
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
|
|
|
|
|
for (g = 0; g < sce->ics.num_swb; g++) {
|
|
|
|
|
if (sce->zeroes[w*16+g])
|
|
|
|
|
continue;
|
|
|
|
|
if (sce->band_type[w*16+g] == INTENSITY_BT || sce->band_type[w*16+g] == INTENSITY_BT2) {
|
|
|
|
|
sce->sf_idx[w*16+g] = prevscaler_i = av_clip(sce->sf_idx[w*16+g], prevscaler_i - SCALE_MAX_DIFF, prevscaler_i + SCALE_MAX_DIFF);
|
|
|
|
|
} else if (sce->band_type[w*16+g] == NOISE_BT) {
|
|
|
|
|
sce->sf_idx[w*16+g] = prevscaler_n = av_clip(sce->sf_idx[w*16+g], prevscaler_n - SCALE_MAX_DIFF, prevscaler_n + SCALE_MAX_DIFF);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
|
|
|
|
|
SingleChannelElement *sce,
|
|
|
|
|
const float lambda)
|
|
|
|
|
{
|
|
|
|
|
int q, w, w2, g, start = 0;
|
|
|
|
|
int i, j;
|
|
|
|
|
int idx;
|
|
|
|
|
TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
|
|
|
|
|
int bandaddr[TRELLIS_STAGES];
|
|
|
|
|
int minq;
|
|
|
|
|
float mincost;
|
|
|
|
|
float q0f = FLT_MAX, q1f = 0.0f, qnrgf = 0.0f;
|
|
|
|
|
int q0, q1, qcnt = 0;
|
|
|
|
|
|
|
|
|
|
for (i = 0; i < 1024; i++) {
|
|
|
|
|
float t = fabsf(sce->coeffs[i]);
|
|
|
|
|
if (t > 0.0f) {
|
|
|
|
|
q0f = FFMIN(q0f, t);
|
|
|
|
|
q1f = FFMAX(q1f, t);
|
|
|
|
|
qnrgf += t*t;
|
|
|
|
|
qcnt++;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (!qcnt) {
|
|
|
|
|
memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
|
|
|
|
|
memset(sce->zeroes, 1, sizeof(sce->zeroes));
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
//minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
|
|
|
|
|
q0 = av_clip(coef2minsf(q0f), 0, SCALE_MAX_POS-1);
|
|
|
|
|
//maximum scalefactor index is when maximum coefficient after quantizing is still not zero
|
|
|
|
|
q1 = av_clip(coef2maxsf(q1f), 1, SCALE_MAX_POS);
|
|
|
|
|
if (q1 - q0 > 60) {
|
|
|
|
|
int q0low = q0;
|
|
|
|
|
int q1high = q1;
|
|
|
|
|
//minimum scalefactor index is when maximum nonzero coefficient after quantizing is not clipped
|
|
|
|
|
int qnrg = av_clip_uint8(log2f(sqrtf(qnrgf/qcnt))*4 - 31 + SCALE_ONE_POS - SCALE_DIV_512);
|
|
|
|
|
q1 = qnrg + 30;
|
|
|
|
|
q0 = qnrg - 30;
|
|
|
|
|
if (q0 < q0low) {
|
|
|
|
|
q1 += q0low - q0;
|
|
|
|
|
q0 = q0low;
|
|
|
|
|
} else if (q1 > q1high) {
|
|
|
|
|
q0 -= q1 - q1high;
|
|
|
|
|
q1 = q1high;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
// q0 == q1 isn't really a legal situation
|
|
|
|
|
if (q0 == q1) {
|
|
|
|
|
// the following is indirect but guarantees q1 != q0 && q1 near q0
|
|
|
|
|
q1 = av_clip(q0+1, 1, SCALE_MAX_POS);
|
|
|
|
|
q0 = av_clip(q1-1, 0, SCALE_MAX_POS - 1);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
for (i = 0; i < TRELLIS_STATES; i++) {
|
|
|
|
|
paths[0][i].cost = 0.0f;
|
|
|
|
|
paths[0][i].prev = -1;
|
|
|
|
|
}
|
|
|
|
|
for (j = 1; j < TRELLIS_STAGES; j++) {
|
|
|
|
|
for (i = 0; i < TRELLIS_STATES; i++) {
|
|
|
|
|
paths[j][i].cost = INFINITY;
|
|
|
|
|
paths[j][i].prev = -2;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
idx = 1;
|
|
|
|
|
s->abs_pow34(s->scoefs, sce->coeffs, 1024);
|
|
|
|
|
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
|
|
|
|
|
start = w*128;
|
|
|
|
|
for (g = 0; g < sce->ics.num_swb; g++) {
|
|
|
|
|
const float *coefs = &sce->coeffs[start];
|
|
|
|
|
float qmin, qmax;
|
|
|
|
|
int nz = 0;
|
|
|
|
|
|
|
|
|
|
bandaddr[idx] = w * 16 + g;
|
|
|
|
|
qmin = INT_MAX;
|
|
|
|
|
qmax = 0.0f;
|
|
|
|
|
for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
|
|
|
|
|
FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
|
|
|
|
|
if (band->energy <= band->threshold || band->threshold == 0.0f) {
|
|
|
|
|
sce->zeroes[(w+w2)*16+g] = 1;
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
sce->zeroes[(w+w2)*16+g] = 0;
|
|
|
|
|
nz = 1;
|
|
|
|
|
for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
|
|
|
|
|
float t = fabsf(coefs[w2*128+i]);
|
|
|
|
|
if (t > 0.0f)
|
|
|
|
|
qmin = FFMIN(qmin, t);
|
|
|
|
|
qmax = FFMAX(qmax, t);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
if (nz) {
|
|
|
|
|
int minscale, maxscale;
|
|
|
|
|
float minrd = INFINITY;
|
|
|
|
|
float maxval;
|
|
|
|
|
//minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
|
|
|
|
|
minscale = coef2minsf(qmin);
|
|
|
|
|
//maximum scalefactor index is when maximum coefficient after quantizing is still not zero
|
|
|
|
|
maxscale = coef2maxsf(qmax);
|
|
|
|
|
minscale = av_clip(minscale - q0, 0, TRELLIS_STATES - 1);
|
|
|
|
|
maxscale = av_clip(maxscale - q0, 0, TRELLIS_STATES);
|
|
|
|
|
if (minscale == maxscale) {
|
|
|
|
|
maxscale = av_clip(minscale+1, 1, TRELLIS_STATES);
|
|
|
|
|
minscale = av_clip(maxscale-1, 0, TRELLIS_STATES - 1);
|
|
|
|
|
}
|
|
|
|
|
maxval = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], s->scoefs+start);
|
|
|
|
|
for (q = minscale; q < maxscale; q++) {
|
|
|
|
|
float dist = 0;
|
|
|
|
|
int cb = find_min_book(maxval, sce->sf_idx[w*16+g]);
|
|
|
|
|
for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
|
|
|
|
|
FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
|
|
|
|
|
dist += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
q + q0, cb, lambda / band->threshold, INFINITY, NULL, NULL, 0);
|
|
|
|
|
}
|
|
|
|
|
minrd = FFMIN(minrd, dist);
|
|
|
|
|
|
|
|
|
|
for (i = 0; i < q1 - q0; i++) {
|
|
|
|
|
float cost;
|
|
|
|
|
cost = paths[idx - 1][i].cost + dist
|
|
|
|
|
+ ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
|
|
|
|
|
if (cost < paths[idx][q].cost) {
|
|
|
|
|
paths[idx][q].cost = cost;
|
|
|
|
|
paths[idx][q].prev = i;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
for (q = 0; q < q1 - q0; q++) {
|
|
|
|
|
paths[idx][q].cost = paths[idx - 1][q].cost + 1;
|
|
|
|
|
paths[idx][q].prev = q;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
sce->zeroes[w*16+g] = !nz;
|
|
|
|
|
start += sce->ics.swb_sizes[g];
|
|
|
|
|
idx++;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
idx--;
|
|
|
|
|
mincost = paths[idx][0].cost;
|
|
|
|
|
minq = 0;
|
|
|
|
|
for (i = 1; i < TRELLIS_STATES; i++) {
|
|
|
|
|
if (paths[idx][i].cost < mincost) {
|
|
|
|
|
mincost = paths[idx][i].cost;
|
|
|
|
|
minq = i;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
while (idx) {
|
|
|
|
|
sce->sf_idx[bandaddr[idx]] = minq + q0;
|
|
|
|
|
minq = FFMAX(paths[idx][minq].prev, 0);
|
|
|
|
|
idx--;
|
|
|
|
|
}
|
|
|
|
|
//set the same quantizers inside window groups
|
|
|
|
|
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
|
|
|
|
|
for (g = 0; g < sce->ics.num_swb; g++)
|
|
|
|
|
for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
|
|
|
|
|
sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
|
|
|
|
|
SingleChannelElement *sce,
|
|
|
|
|
const float lambda)
|
|
|
|
|
{
|
|
|
|
|
int start = 0, i, w, w2, g;
|
|
|
|
|
int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels * (lambda / 120.f);
|
|
|
|
|
float dists[128] = { 0 }, uplims[128] = { 0 };
|
|
|
|
|
float maxvals[128];
|
|
|
|
|
int fflag, minscaler;
|
|
|
|
|
int its = 0;
|
|
|
|
|
int allz = 0;
|
|
|
|
|
float minthr = INFINITY;
|
|
|
|
|
|
|
|
|
|
// for values above this the decoder might end up in an endless loop
|
|
|
|
|
// due to always having more bits than what can be encoded.
|
|
|
|
|
destbits = FFMIN(destbits, 5800);
|
|
|
|
|
//some heuristic to determine initial quantizers will reduce search time
|
|
|
|
|
//determine zero bands and upper limits
|
|
|
|
|
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
|
|
|
|
|
start = 0;
|
|
|
|
|
for (g = 0; g < sce->ics.num_swb; g++) {
|
|
|
|
|
int nz = 0;
|
|
|
|
|
float uplim = 0.0f, energy = 0.0f;
|
|
|
|
|
for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
|
|
|
|
|
FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
|
|
|
|
|
uplim += band->threshold;
|
|
|
|
|
energy += band->energy;
|
|
|
|
|
if (band->energy <= band->threshold || band->threshold == 0.0f) {
|
|
|
|
|
sce->zeroes[(w+w2)*16+g] = 1;
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
nz = 1;
|
|
|
|
|
}
|
|
|
|
|
uplims[w*16+g] = uplim *512;
|
|
|
|
|
sce->band_type[w*16+g] = 0;
|
|
|
|
|
sce->zeroes[w*16+g] = !nz;
|
|
|
|
|
if (nz)
|
|
|
|
|
minthr = FFMIN(minthr, uplim);
|
|
|
|
|
allz |= nz;
|
|
|
|
|
start += sce->ics.swb_sizes[g];
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
|
|
|
|
|
for (g = 0; g < sce->ics.num_swb; g++) {
|
|
|
|
|
if (sce->zeroes[w*16+g]) {
|
|
|
|
|
sce->sf_idx[w*16+g] = SCALE_ONE_POS;
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2f(uplims[w*16+g]/minthr)*4,59);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (!allz)
|
|
|
|
|
return;
|
|
|
|
|
s->abs_pow34(s->scoefs, sce->coeffs, 1024);
|
|
|
|
|
ff_quantize_band_cost_cache_init(s);
|
|
|
|
|
|
|
|
|
|
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
|
|
|
|
|
start = w*128;
|
|
|
|
|
for (g = 0; g < sce->ics.num_swb; g++) {
|
|
|
|
|
const float *scaled = s->scoefs + start;
|
|
|
|
|
maxvals[w*16+g] = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled);
|
|
|
|
|
start += sce->ics.swb_sizes[g];
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
//perform two-loop search
|
|
|
|
|
//outer loop - improve quality
|
|
|
|
|
do {
|
|
|
|
|
int tbits, qstep;
|
|
|
|
|
minscaler = sce->sf_idx[0];
|
|
|
|
|
//inner loop - quantize spectrum to fit into given number of bits
|
|
|
|
|
qstep = its ? 1 : 32;
|
|
|
|
|
do {
|
|
|
|
|
int prev = -1;
|
|
|
|
|
tbits = 0;
|
|
|
|
|
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
|
|
|
|
|
start = w*128;
|
|
|
|
|
for (g = 0; g < sce->ics.num_swb; g++) {
|
|
|
|
|
const float *coefs = sce->coeffs + start;
|
|
|
|
|
const float *scaled = s->scoefs + start;
|
|
|
|
|
int bits = 0;
|
|
|
|
|
int cb;
|
|
|
|
|
float dist = 0.0f;
|
|
|
|
|
|
|
|
|
|
if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
|
|
|
|
|
start += sce->ics.swb_sizes[g];
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
|
|
|
|
|
cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
|
|
|
|
|
for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
|
|
|
|
|
int b;
|
|
|
|
|
dist += quantize_band_cost_cached(s, w + w2, g,
|
|
|
|
|
coefs + w2*128,
|
|
|
|
|
scaled + w2*128,
|
|
|
|
|
sce->ics.swb_sizes[g],
|
|
|
|
|
sce->sf_idx[w*16+g],
|
|
|
|
|
cb, 1.0f, INFINITY,
|
|
|
|
|
&b, NULL, 0);
|
|
|
|
|
bits += b;
|
|
|
|
|
}
|
|
|
|
|
dists[w*16+g] = dist - bits;
|
|
|
|
|
if (prev != -1) {
|
|
|
|
|
bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
|
|
|
|
|
}
|
|
|
|
|
tbits += bits;
|
|
|
|
|
start += sce->ics.swb_sizes[g];
|
|
|
|
|
prev = sce->sf_idx[w*16+g];
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
if (tbits > destbits) {
|
|
|
|
|
for (i = 0; i < 128; i++)
|
|
|
|
|
if (sce->sf_idx[i] < 218 - qstep)
|
|
|
|
|
sce->sf_idx[i] += qstep;
|
|
|
|
|
} else {
|
|
|
|
|
for (i = 0; i < 128; i++)
|
|
|
|
|
if (sce->sf_idx[i] > 60 - qstep)
|
|
|
|
|
sce->sf_idx[i] -= qstep;
|
|
|
|
|
}
|
|
|
|
|
qstep >>= 1;
|
|
|
|
|
if (!qstep && tbits > destbits*1.02 && sce->sf_idx[0] < 217)
|
|
|
|
|
qstep = 1;
|
|
|
|
|
} while (qstep);
|
|
|
|
|
|
|
|
|
|
fflag = 0;
|
|
|
|
|
minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
|
|
|
|
|
|
|
|
|
|
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
|
|
|
|
|
for (g = 0; g < sce->ics.num_swb; g++) {
|
|
|
|
|
int prevsc = sce->sf_idx[w*16+g];
|
|
|
|
|
if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60) {
|
|
|
|
|
if (find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]-1))
|
|
|
|
|
sce->sf_idx[w*16+g]--;
|
|
|
|
|
else //Try to make sure there is some energy in every band
|
|
|
|
|
sce->sf_idx[w*16+g]-=2;
|
|
|
|
|
}
|
|
|
|
|
sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
|
|
|
|
|
sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
|
|
|
|
|
if (sce->sf_idx[w*16+g] != prevsc)
|
|
|
|
|
fflag = 1;
|
|
|
|
|
sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
its++;
|
|
|
|
|
} while (fflag && its < 10);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChannelElement *sce)
|
|
|
|
|
{
|
|
|
|
|
FFPsyBand *band;
|
|
|
|
|
int w, g, w2, i;
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
int wlen = 1024 / sce->ics.num_windows;
|
|
|
|
|
int bandwidth, cutoff;
|
|
|
|
|
float *PNS = &s->scoefs[0*128], *PNS34 = &s->scoefs[1*128];
|
|
|
|
|
float *NOR34 = &s->scoefs[3*128];
|
AAC encoder: improve SF range utilization
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
9 years ago
|
|
|
uint8_t nextband[128];
|
|
|
|
|
const float lambda = s->lambda;
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
const float freq_mult = avctx->sample_rate*0.5f/wlen;
|
|
|
|
|
const float thr_mult = NOISE_LAMBDA_REPLACE*(100.0f/lambda);
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
const float spread_threshold = FFMIN(0.75f, NOISE_SPREAD_THRESHOLD*FFMAX(0.5f, lambda/100.f));
|
|
|
|
|
const float dist_bias = av_clipf(4.f * 120 / lambda, 0.25f, 4.0f);
|
|
|
|
|
const float pns_transient_energy_r = FFMIN(0.7f, lambda / 140.f);
|
|
|
|
|
|
|
|
|
|
int refbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
|
|
|
|
|
/ ((avctx->flags & AV_CODEC_FLAG_QSCALE) ? 2.0f : avctx->channels)
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
* (lambda / 120.f);
|
|
|
|
|
|
|
|
|
|
/** Keep this in sync with twoloop's cutoff selection */
|
|
|
|
|
float rate_bandwidth_multiplier = 1.5f;
|
AAC encoder: improve SF range utilization
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
9 years ago
|
|
|
int prev = -1000, prev_sf = -1;
|
|
|
|
|
int frame_bit_rate = (avctx->flags & AV_CODEC_FLAG_QSCALE)
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
|
|
|
|
|
: (avctx->bit_rate / avctx->channels);
|
|
|
|
|
|
|
|
|
|
frame_bit_rate *= 1.15f;
|
|
|
|
|
|
|
|
|
|
if (avctx->cutoff > 0) {
|
|
|
|
|
bandwidth = avctx->cutoff;
|
|
|
|
|
} else {
|
|
|
|
|
bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
|
|
|
|
|
|
|
|
|
|
memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type));
|
AAC encoder: improve SF range utilization
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
9 years ago
|
|
|
ff_init_nextband_map(sce, nextband);
|
|
|
|
|
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
|
|
|
|
|
int wstart = w*128;
|
|
|
|
|
for (g = 0; g < sce->ics.num_swb; g++) {
|
|
|
|
|
int noise_sfi;
|
|
|
|
|
float dist1 = 0.0f, dist2 = 0.0f, noise_amp;
|
|
|
|
|
float pns_energy = 0.0f, pns_tgt_energy, energy_ratio, dist_thresh;
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
float sfb_energy = 0.0f, threshold = 0.0f, spread = 2.0f;
|
|
|
|
|
float min_energy = -1.0f, max_energy = 0.0f;
|
|
|
|
|
const int start = wstart+sce->ics.swb_offset[g];
|
|
|
|
|
const float freq = (start-wstart)*freq_mult;
|
|
|
|
|
const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
|
|
|
|
|
if (freq < NOISE_LOW_LIMIT || (start-wstart) >= cutoff) {
|
|
|
|
|
if (!sce->zeroes[w*16+g])
|
|
|
|
|
prev_sf = sce->sf_idx[w*16+g];
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
|
|
|
|
|
band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
|
|
|
|
|
sfb_energy += band->energy;
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
spread = FFMIN(spread, band->spread);
|
|
|
|
|
threshold += band->threshold;
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
if (!w2) {
|
|
|
|
|
min_energy = max_energy = band->energy;
|
|
|
|
|
} else {
|
|
|
|
|
min_energy = FFMIN(min_energy, band->energy);
|
|
|
|
|
max_energy = FFMAX(max_energy, band->energy);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Ramps down at ~8000Hz and loosens the dist threshold */
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
dist_thresh = av_clipf(2.5f*NOISE_LOW_LIMIT/freq, 0.5f, 2.5f) * dist_bias;
|
|
|
|
|
|
|
|
|
|
/* PNS is acceptable when all of these are true:
|
|
|
|
|
* 1. high spread energy (noise-like band)
|
|
|
|
|
* 2. near-threshold energy (high PE means the random nature of PNS content will be noticed)
|
|
|
|
|
* 3. on short window groups, all windows have similar energy (variations in energy would be destroyed by PNS)
|
|
|
|
|
*
|
|
|
|
|
* At this stage, point 2 is relaxed for zeroed bands near the noise threshold (hole avoidance is more important)
|
|
|
|
|
*/
|
AAC encoder: improve SF range utilization
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
9 years ago
|
|
|
if ((!sce->zeroes[w*16+g] && !ff_sfdelta_can_remove_band(sce, nextband, prev_sf, w*16+g)) ||
|
|
|
|
|
((sce->zeroes[w*16+g] || !sce->band_alt[w*16+g]) && sfb_energy < threshold*sqrtf(1.0f/freq_boost)) || spread < spread_threshold ||
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
(!sce->zeroes[w*16+g] && sce->band_alt[w*16+g] && sfb_energy > threshold*thr_mult*freq_boost) ||
|
|
|
|
|
min_energy < pns_transient_energy_r * max_energy ) {
|
|
|
|
|
sce->pns_ener[w*16+g] = sfb_energy;
|
AAC encoder: improve SF range utilization
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
9 years ago
|
|
|
if (!sce->zeroes[w*16+g])
|
|
|
|
|
prev_sf = sce->sf_idx[w*16+g];
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
pns_tgt_energy = sfb_energy*FFMIN(1.0f, spread*spread);
|
|
|
|
|
noise_sfi = av_clip(roundf(log2f(pns_tgt_energy)*2), -100, 155); /* Quantize */
|
|
|
|
|
noise_amp = -ff_aac_pow2sf_tab[noise_sfi + POW_SF2_ZERO]; /* Dequantize */
|
AAC encoder: improve SF range utilization
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
9 years ago
|
|
|
if (prev != -1000) {
|
|
|
|
|
int noise_sfdiff = noise_sfi - prev + SCALE_DIFF_ZERO;
|
|
|
|
|
if (noise_sfdiff < 0 || noise_sfdiff > 2*SCALE_MAX_DIFF) {
|
|
|
|
|
if (!sce->zeroes[w*16+g])
|
|
|
|
|
prev_sf = sce->sf_idx[w*16+g];
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
|
|
|
|
|
float band_energy, scale, pns_senergy;
|
|
|
|
|
const int start_c = (w+w2)*128+sce->ics.swb_offset[g];
|
|
|
|
|
band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
|
|
|
|
|
for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
|
|
|
|
|
s->random_state = lcg_random(s->random_state);
|
|
|
|
|
PNS[i] = s->random_state;
|
|
|
|
|
}
|
|
|
|
|
band_energy = s->fdsp->scalarproduct_float(PNS, PNS, sce->ics.swb_sizes[g]);
|
|
|
|
|
scale = noise_amp/sqrtf(band_energy);
|
|
|
|
|
s->fdsp->vector_fmul_scalar(PNS, PNS, scale, sce->ics.swb_sizes[g]);
|
|
|
|
|
pns_senergy = s->fdsp->scalarproduct_float(PNS, PNS, sce->ics.swb_sizes[g]);
|
|
|
|
|
pns_energy += pns_senergy;
|
|
|
|
|
s->abs_pow34(NOR34, &sce->coeffs[start_c], sce->ics.swb_sizes[g]);
|
|
|
|
|
s->abs_pow34(PNS34, PNS, sce->ics.swb_sizes[g]);
|
|
|
|
|
dist1 += quantize_band_cost(s, &sce->coeffs[start_c],
|
|
|
|
|
NOR34,
|
|
|
|
|
sce->ics.swb_sizes[g],
|
|
|
|
|
sce->sf_idx[(w+w2)*16+g],
|
|
|
|
|
sce->band_alt[(w+w2)*16+g],
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
lambda/band->threshold, INFINITY, NULL, NULL, 0);
|
|
|
|
|
/* Estimate rd on average as 5 bits for SF, 4 for the CB, plus spread energy * lambda/thr */
|
|
|
|
|
dist2 += band->energy/(band->spread*band->spread)*lambda*dist_thresh/band->threshold;
|
|
|
|
|
}
|
|
|
|
|
if (g && sce->band_type[w*16+g-1] == NOISE_BT) {
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
dist2 += 5;
|
|
|
|
|
} else {
|
|
|
|
|
dist2 += 9;
|
|
|
|
|
}
|
|
|
|
|
energy_ratio = pns_tgt_energy/pns_energy; /* Compensates for quantization error */
|
|
|
|
|
sce->pns_ener[w*16+g] = energy_ratio*pns_tgt_energy;
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
if (sce->zeroes[w*16+g] || !sce->band_alt[w*16+g] || (energy_ratio > 0.85f && energy_ratio < 1.25f && dist2 < dist1)) {
|
|
|
|
|
sce->band_type[w*16+g] = NOISE_BT;
|
|
|
|
|
sce->zeroes[w*16+g] = 0;
|
AAC encoder: improve SF range utilization
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
9 years ago
|
|
|
prev = noise_sfi;
|
|
|
|
|
} else {
|
|
|
|
|
if (!sce->zeroes[w*16+g])
|
|
|
|
|
prev_sf = sce->sf_idx[w*16+g];
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
static void mark_pns(AACEncContext *s, AVCodecContext *avctx, SingleChannelElement *sce)
|
|
|
|
|
{
|
|
|
|
|
FFPsyBand *band;
|
|
|
|
|
int w, g, w2;
|
|
|
|
|
int wlen = 1024 / sce->ics.num_windows;
|
|
|
|
|
int bandwidth, cutoff;
|
|
|
|
|
const float lambda = s->lambda;
|
|
|
|
|
const float freq_mult = avctx->sample_rate*0.5f/wlen;
|
|
|
|
|
const float spread_threshold = FFMIN(0.75f, NOISE_SPREAD_THRESHOLD*FFMAX(0.5f, lambda/100.f));
|
|
|
|
|
const float pns_transient_energy_r = FFMIN(0.7f, lambda / 140.f);
|
|
|
|
|
|
|
|
|
|
int refbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
|
|
|
|
|
/ ((avctx->flags & AV_CODEC_FLAG_QSCALE) ? 2.0f : avctx->channels)
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
* (lambda / 120.f);
|
|
|
|
|
|
|
|
|
|
/** Keep this in sync with twoloop's cutoff selection */
|
|
|
|
|
float rate_bandwidth_multiplier = 1.5f;
|
|
|
|
|
int frame_bit_rate = (avctx->flags & AV_CODEC_FLAG_QSCALE)
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
|
|
|
|
|
: (avctx->bit_rate / avctx->channels);
|
|
|
|
|
|
|
|
|
|
frame_bit_rate *= 1.15f;
|
|
|
|
|
|
|
|
|
|
if (avctx->cutoff > 0) {
|
|
|
|
|
bandwidth = avctx->cutoff;
|
|
|
|
|
} else {
|
|
|
|
|
bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
|
|
|
|
|
|
|
|
|
|
memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type));
|
|
|
|
|
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
|
|
|
|
|
for (g = 0; g < sce->ics.num_swb; g++) {
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
float sfb_energy = 0.0f, threshold = 0.0f, spread = 2.0f;
|
|
|
|
|
float min_energy = -1.0f, max_energy = 0.0f;
|
|
|
|
|
const int start = sce->ics.swb_offset[g];
|
|
|
|
|
const float freq = start*freq_mult;
|
|
|
|
|
const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
|
|
|
|
|
if (freq < NOISE_LOW_LIMIT || start >= cutoff) {
|
|
|
|
|
sce->can_pns[w*16+g] = 0;
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
|
|
|
|
|
band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
|
|
|
|
|
sfb_energy += band->energy;
|
|
|
|
|
spread = FFMIN(spread, band->spread);
|
|
|
|
|
threshold += band->threshold;
|
|
|
|
|
if (!w2) {
|
|
|
|
|
min_energy = max_energy = band->energy;
|
|
|
|
|
} else {
|
|
|
|
|
min_energy = FFMIN(min_energy, band->energy);
|
|
|
|
|
max_energy = FFMAX(max_energy, band->energy);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* PNS is acceptable when all of these are true:
|
|
|
|
|
* 1. high spread energy (noise-like band)
|
|
|
|
|
* 2. near-threshold energy (high PE means the random nature of PNS content will be noticed)
|
|
|
|
|
* 3. on short window groups, all windows have similar energy (variations in energy would be destroyed by PNS)
|
|
|
|
|
*/
|
|
|
|
|
sce->pns_ener[w*16+g] = sfb_energy;
|
|
|
|
|
if (sfb_energy < threshold*sqrtf(1.5f/freq_boost) || spread < spread_threshold || min_energy < pns_transient_energy_r * max_energy) {
|
|
|
|
|
sce->can_pns[w*16+g] = 0;
|
|
|
|
|
} else {
|
|
|
|
|
sce->can_pns[w*16+g] = 1;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void search_for_ms(AACEncContext *s, ChannelElement *cpe)
|
|
|
|
|
{
|
AAC encoder: improve SF range utilization
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
9 years ago
|
|
|
int start = 0, i, w, w2, g, sid_sf_boost, prev_mid, prev_side;
|
|
|
|
|
uint8_t nextband0[128], nextband1[128];
|
|
|
|
|
float *M = s->scoefs + 128*0, *S = s->scoefs + 128*1;
|
|
|
|
|
float *L34 = s->scoefs + 128*2, *R34 = s->scoefs + 128*3;
|
|
|
|
|
float *M34 = s->scoefs + 128*4, *S34 = s->scoefs + 128*5;
|
|
|
|
|
const float lambda = s->lambda;
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
const float mslambda = FFMIN(1.0f, lambda / 120.f);
|
|
|
|
|
SingleChannelElement *sce0 = &cpe->ch[0];
|
|
|
|
|
SingleChannelElement *sce1 = &cpe->ch[1];
|
|
|
|
|
if (!cpe->common_window)
|
|
|
|
|
return;
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
|
AAC encoder: improve SF range utilization
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
9 years ago
|
|
|
/** Scout out next nonzero bands */
|
|
|
|
|
ff_init_nextband_map(sce0, nextband0);
|
|
|
|
|
ff_init_nextband_map(sce1, nextband1);
|
|
|
|
|
|
|
|
|
|
prev_mid = sce0->sf_idx[0];
|
|
|
|
|
prev_side = sce1->sf_idx[0];
|
|
|
|
|
for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
|
|
|
|
|
start = 0;
|
|
|
|
|
for (g = 0; g < sce0->ics.num_swb; g++) {
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
float bmax = bval2bmax(g * 17.0f / sce0->ics.num_swb) / 0.0045f;
|
|
|
|
|
if (!cpe->is_mask[w*16+g])
|
|
|
|
|
cpe->ms_mask[w*16+g] = 0;
|
|
|
|
|
if (!sce0->zeroes[w*16+g] && !sce1->zeroes[w*16+g] && !cpe->is_mask[w*16+g]) {
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
float Mmax = 0.0f, Smax = 0.0f;
|
|
|
|
|
|
|
|
|
|
/* Must compute mid/side SF and book for the whole window group */
|
|
|
|
|
for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
|
|
|
|
|
for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
|
|
|
|
|
M[i] = (sce0->coeffs[start+(w+w2)*128+i]
|
|
|
|
|
+ sce1->coeffs[start+(w+w2)*128+i]) * 0.5;
|
|
|
|
|
S[i] = M[i]
|
|
|
|
|
- sce1->coeffs[start+(w+w2)*128+i];
|
|
|
|
|
}
|
|
|
|
|
s->abs_pow34(M34, M, sce0->ics.swb_sizes[g]);
|
|
|
|
|
s->abs_pow34(S34, S, sce0->ics.swb_sizes[g]);
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
for (i = 0; i < sce0->ics.swb_sizes[g]; i++ ) {
|
|
|
|
|
Mmax = FFMAX(Mmax, M34[i]);
|
|
|
|
|
Smax = FFMAX(Smax, S34[i]);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
for (sid_sf_boost = 0; sid_sf_boost < 4; sid_sf_boost++) {
|
|
|
|
|
float dist1 = 0.0f, dist2 = 0.0f;
|
|
|
|
|
int B0 = 0, B1 = 0;
|
|
|
|
|
int minidx;
|
|
|
|
|
int mididx, sididx;
|
|
|
|
|
int midcb, sidcb;
|
|
|
|
|
|
|
|
|
|
minidx = FFMIN(sce0->sf_idx[w*16+g], sce1->sf_idx[w*16+g]);
|
AAC encoder: improve SF range utilization
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
9 years ago
|
|
|
mididx = av_clip(minidx, 0, SCALE_MAX_POS - SCALE_DIV_512);
|
|
|
|
|
sididx = av_clip(minidx - sid_sf_boost * 3, 0, SCALE_MAX_POS - SCALE_DIV_512);
|
|
|
|
|
if (sce0->band_type[w*16+g] != NOISE_BT && sce1->band_type[w*16+g] != NOISE_BT
|
AAC encoder: improve SF range utilization
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
9 years ago
|
|
|
&& ( !ff_sfdelta_can_replace(sce0, nextband0, prev_mid, mididx, w*16+g)
|
|
|
|
|
|| !ff_sfdelta_can_replace(sce1, nextband1, prev_side, sididx, w*16+g))) {
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
/* scalefactor range violation, bad stuff, will decrease quality unacceptably */
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
|
AAC encoder: improve SF range utilization
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
9 years ago
|
|
|
midcb = find_min_book(Mmax, mididx);
|
|
|
|
|
sidcb = find_min_book(Smax, sididx);
|
|
|
|
|
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
/* No CB can be zero */
|
|
|
|
|
midcb = FFMAX(1,midcb);
|
|
|
|
|
sidcb = FFMAX(1,sidcb);
|
|
|
|
|
|
|
|
|
|
for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
|
|
|
|
|
FFPsyBand *band0 = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g];
|
|
|
|
|
FFPsyBand *band1 = &s->psy.ch[s->cur_channel+1].psy_bands[(w+w2)*16+g];
|
|
|
|
|
float minthr = FFMIN(band0->threshold, band1->threshold);
|
|
|
|
|
int b1,b2,b3,b4;
|
|
|
|
|
for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
|
|
|
|
|
M[i] = (sce0->coeffs[start+(w+w2)*128+i]
|
|
|
|
|
+ sce1->coeffs[start+(w+w2)*128+i]) * 0.5;
|
|
|
|
|
S[i] = M[i]
|
|
|
|
|
- sce1->coeffs[start+(w+w2)*128+i];
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
s->abs_pow34(L34, sce0->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
|
|
|
|
|
s->abs_pow34(R34, sce1->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
|
|
|
|
|
s->abs_pow34(M34, M, sce0->ics.swb_sizes[g]);
|
|
|
|
|
s->abs_pow34(S34, S, sce0->ics.swb_sizes[g]);
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
dist1 += quantize_band_cost(s, &sce0->coeffs[start + (w+w2)*128],
|
|
|
|
|
L34,
|
|
|
|
|
sce0->ics.swb_sizes[g],
|
|
|
|
|
sce0->sf_idx[w*16+g],
|
|
|
|
|
sce0->band_type[w*16+g],
|
|
|
|
|
lambda / (band0->threshold + FLT_MIN), INFINITY, &b1, NULL, 0);
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
dist1 += quantize_band_cost(s, &sce1->coeffs[start + (w+w2)*128],
|
|
|
|
|
R34,
|
|
|
|
|
sce1->ics.swb_sizes[g],
|
|
|
|
|
sce1->sf_idx[w*16+g],
|
|
|
|
|
sce1->band_type[w*16+g],
|
|
|
|
|
lambda / (band1->threshold + FLT_MIN), INFINITY, &b2, NULL, 0);
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
dist2 += quantize_band_cost(s, M,
|
|
|
|
|
M34,
|
|
|
|
|
sce0->ics.swb_sizes[g],
|
|
|
|
|
mididx,
|
|
|
|
|
midcb,
|
|
|
|
|
lambda / (minthr + FLT_MIN), INFINITY, &b3, NULL, 0);
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
dist2 += quantize_band_cost(s, S,
|
|
|
|
|
S34,
|
|
|
|
|
sce1->ics.swb_sizes[g],
|
|
|
|
|
sididx,
|
|
|
|
|
sidcb,
|
|
|
|
|
mslambda / (minthr * bmax + FLT_MIN), INFINITY, &b4, NULL, 0);
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
B0 += b1+b2;
|
|
|
|
|
B1 += b3+b4;
|
|
|
|
|
dist1 -= b1+b2;
|
|
|
|
|
dist2 -= b3+b4;
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
}
|
|
|
|
|
cpe->ms_mask[w*16+g] = dist2 <= dist1 && B1 < B0;
|
|
|
|
|
if (cpe->ms_mask[w*16+g]) {
|
|
|
|
|
if (sce0->band_type[w*16+g] != NOISE_BT && sce1->band_type[w*16+g] != NOISE_BT) {
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
sce0->sf_idx[w*16+g] = mididx;
|
|
|
|
|
sce1->sf_idx[w*16+g] = sididx;
|
|
|
|
|
sce0->band_type[w*16+g] = midcb;
|
|
|
|
|
sce1->band_type[w*16+g] = sidcb;
|
|
|
|
|
} else if ((sce0->band_type[w*16+g] != NOISE_BT) ^ (sce1->band_type[w*16+g] != NOISE_BT)) {
|
|
|
|
|
/* ms_mask unneeded, and it confuses some decoders */
|
|
|
|
|
cpe->ms_mask[w*16+g] = 0;
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
}
|
|
|
|
|
break;
|
|
|
|
|
} else if (B1 > B0) {
|
|
|
|
|
/* More boost won't fix this */
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
AAC encoder: improve SF range utilization
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
9 years ago
|
|
|
if (!sce0->zeroes[w*16+g] && sce0->band_type[w*16+g] < RESERVED_BT)
|
|
|
|
|
prev_mid = sce0->sf_idx[w*16+g];
|
|
|
|
|
if (!sce1->zeroes[w*16+g] && !cpe->is_mask[w*16+g] && sce1->band_type[w*16+g] < RESERVED_BT)
|
|
|
|
|
prev_side = sce1->sf_idx[w*16+g];
|
|
|
|
|
start += sce0->ics.swb_sizes[g];
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
const AACCoefficientsEncoder ff_aac_coders[AAC_CODER_NB] = {
|
|
|
|
|
[AAC_CODER_ANMR] = {
|
|
|
|
|
search_for_quantizers_anmr,
|
|
|
|
|
encode_window_bands_info,
|
|
|
|
|
quantize_and_encode_band,
|
|
|
|
|
ff_aac_encode_tns_info,
|
|
|
|
|
ff_aac_encode_ltp_info,
|
|
|
|
|
ff_aac_encode_main_pred,
|
|
|
|
|
ff_aac_adjust_common_pred,
|
|
|
|
|
ff_aac_adjust_common_ltp,
|
|
|
|
|
ff_aac_apply_main_pred,
|
|
|
|
|
ff_aac_apply_tns,
|
|
|
|
|
ff_aac_update_ltp,
|
|
|
|
|
ff_aac_ltp_insert_new_frame,
|
|
|
|
|
set_special_band_scalefactors,
|
|
|
|
|
search_for_pns,
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
mark_pns,
|
|
|
|
|
ff_aac_search_for_tns,
|
|
|
|
|
ff_aac_search_for_ltp,
|
|
|
|
|
search_for_ms,
|
|
|
|
|
ff_aac_search_for_is,
|
|
|
|
|
ff_aac_search_for_pred,
|
|
|
|
|
},
|
|
|
|
|
[AAC_CODER_TWOLOOP] = {
|
|
|
|
|
search_for_quantizers_twoloop,
|
|
|
|
|
codebook_trellis_rate,
|
|
|
|
|
quantize_and_encode_band,
|
|
|
|
|
ff_aac_encode_tns_info,
|
|
|
|
|
ff_aac_encode_ltp_info,
|
|
|
|
|
ff_aac_encode_main_pred,
|
|
|
|
|
ff_aac_adjust_common_pred,
|
|
|
|
|
ff_aac_adjust_common_ltp,
|
|
|
|
|
ff_aac_apply_main_pred,
|
|
|
|
|
ff_aac_apply_tns,
|
|
|
|
|
ff_aac_update_ltp,
|
|
|
|
|
ff_aac_ltp_insert_new_frame,
|
|
|
|
|
set_special_band_scalefactors,
|
|
|
|
|
search_for_pns,
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
mark_pns,
|
|
|
|
|
ff_aac_search_for_tns,
|
|
|
|
|
ff_aac_search_for_ltp,
|
|
|
|
|
search_for_ms,
|
|
|
|
|
ff_aac_search_for_is,
|
|
|
|
|
ff_aac_search_for_pred,
|
|
|
|
|
},
|
|
|
|
|
[AAC_CODER_FAST] = {
|
|
|
|
|
search_for_quantizers_fast,
|
|
|
|
|
codebook_trellis_rate,
|
|
|
|
|
quantize_and_encode_band,
|
|
|
|
|
ff_aac_encode_tns_info,
|
|
|
|
|
ff_aac_encode_ltp_info,
|
|
|
|
|
ff_aac_encode_main_pred,
|
|
|
|
|
ff_aac_adjust_common_pred,
|
|
|
|
|
ff_aac_adjust_common_ltp,
|
|
|
|
|
ff_aac_apply_main_pred,
|
|
|
|
|
ff_aac_apply_tns,
|
|
|
|
|
ff_aac_update_ltp,
|
|
|
|
|
ff_aac_ltp_insert_new_frame,
|
|
|
|
|
set_special_band_scalefactors,
|
|
|
|
|
search_for_pns,
|
AAC encoder: Extensive improvements
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.
9 years ago
|
|
|
mark_pns,
|
|
|
|
|
ff_aac_search_for_tns,
|
|
|
|
|
ff_aac_search_for_ltp,
|
|
|
|
|
search_for_ms,
|
|
|
|
|
ff_aac_search_for_is,
|
|
|
|
|
ff_aac_search_for_pred,
|
|
|
|
|
},
|
|
|
|
|
};
|
