virtualenv/FFmpeg
1
0
Fork 0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2024-12-18 03:19:31 +02:00
Code Issues Releases Activity
b78243c504
FFmpeg/libavcodec/aaccoder_twoloop.h

764 lines
35 KiB
C
Raw Normal View History

AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
/*
* AAC encoder twoloop coder
* 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 encoder twoloop coder
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.
2015-10-11 22:29:50 +02:00
* @author Konstantin Shishkov, Claudio Freire
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
*/
/**
* This file contains a template for the twoloop coder function.
* It needs to be provided, externally, as an already included declaration,
* the following functions from aacenc_quantization/util.h. They're not included
* explicitly here to make it possible to provide alternative implementations:
* - quantize_band_cost
* - abs_pow34_v
* - find_max_val
* - find_min_book
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.
2015-10-11 22:29:50 +02:00
* - find_form_factor
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
*/
#ifndef AVCODEC_AACCODER_TWOLOOP_H
#define AVCODEC_AACCODER_TWOLOOP_H
#include <float.h>
#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.
2015-10-11 22:29:50 +02:00
#include "mathops.h"
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
#include "avcodec.h"
#include "put_bits.h"
#include "aac.h"
#include "aacenc.h"
#include "aactab.h"
#include "aacenctab.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.
2015-10-11 22:29:50 +02:00
/** Frequency in Hz for lower limit of noise substitution **/
#define NOISE_LOW_LIMIT 4000
#define sclip(x) av_clip(x,60,218)
aaccoder_twoloop.h: simplify and comment ff_pns_bits()
2015-10-12 02:42:43 +02:00
/* Reflects the cost to change codebooks */
static inline int ff_pns_bits(SingleChannelElement *sce, int w, int 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.
2015-10-11 22:29:50 +02:00
{
aaccoder_twoloop.h: simplify and comment ff_pns_bits()
2015-10-12 02:42:43 +02:00
return (!g || !sce->zeroes[w*16+g-1] || !sce->can_pns[w*16+g-1]) ? 9 : 5;
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.
2015-10-11 22:29:50 +02:00
}
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
/**
* two-loop quantizers search taken from ISO 13818-7 Appendix C
*/
static void search_for_quantizers_twoloop(AVCodecContext *avctx,
AACEncContext *s,
SingleChannelElement *sce,
const float 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.
2015-10-11 22:29:50 +02:00
int start = 0, i, w, w2, g, recomprd;
int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
/ ((avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : avctx->channels)
* (lambda / 120.f);
int refbits = destbits;
int toomanybits, toofewbits;
char nzs[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
2015-12-01 08:28:36 +02:00
uint8_t nextband[128];
AAC encoder: fix signed integer overflow Clamp scalefactors by coef2minsf to avoid undefined behavior caused by signed integer overflow. It also avoids clipping of coefficients so it should avoid artifacts as well, on very rare corner cases.
2016-03-30 15:05:03 +02:00
int maxsf[128], minsf[128];
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.
2015-10-11 22:29:50 +02:00
float dists[128] = { 0 }, qenergies[128] = { 0 }, uplims[128], euplims[128], energies[128];
float maxvals[128], spread_thr_r[128];
float min_spread_thr_r, max_spread_thr_r;
/**
* rdlambda controls the maximum tolerated distortion. Twoloop
* will keep iterating until it fails to lower it or it reaches
* ulimit * rdlambda. Keeping it low increases quality on difficult
* signals, but lower it too much, and bits will be taken from weak
doc: fix spelling errors Thanks to Mathieu Malaterre <malat@debian.org> for reporting the Que/Queue typo. (https://bugs.debian.org/839542) Reviewed-by: Lou Logan <lou@lrcd.com> Signed-off-by: Andreas Cadhalpun <Andreas.Cadhalpun@googlemail.com>
2016-10-13 23:08:01 +02:00
* signals, creating "holes". A balance is necessary.
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.
2015-10-11 22:29:50 +02:00
* rdmax and rdmin specify the relative deviation from rdlambda
* allowed for tonality compensation
*/
float rdlambda = av_clipf(2.0f * 120.f / lambda, 0.0625f, 16.0f);
const float nzslope = 1.5f;
float rdmin = 0.03125f;
float rdmax = 1.0f;
/**
* sfoffs controls an offset of optmium allocation that will be
* applied based on lambda. Keep it real and modest, the loop
* will take care of the rest, this just accelerates convergence
*/
float sfoffs = av_clipf(log2f(120.0f / lambda) * 4.0f, -5, 10);
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
2015-12-01 08:28:36 +02:00
int fflag, minscaler, maxscaler, nminscaler;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
int its = 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.
2015-10-11 22:29:50 +02:00
int maxits = 30;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
int allz = 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.
2015-10-11 22:29:50 +02:00
int tbits;
int cutoff = 1024;
int pns_start_pos;
AAC encoder: fix assertion error re SF differences Intermediate results can indeed violate SF delta. Instead of asserting there, just make the code safe, and assert on the final result. Also re-clamp SFs more often in short windows (which tend to violate the restriction when encoding the switch from one window to the other)
2015-10-12 04:00:46 +02:00
int prev;
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.
2015-10-11 22:29:50 +02:00
/**
* zeroscale controls a multiplier of the threshold, if band energy
* is below this, a zero is forced. Keep it lower than 1, unless
* low lambda is used, because energy < threshold doesn't mean there's
* no audible signal outright, it's just energy. Also make it rise
* slower than rdlambda, as rdscale has due compensation with
* noisy band depriorization below, whereas zeroing logic is rather dumb
*/
float zeroscale;
if (lambda > 120.f) {
zeroscale = av_clipf(powf(120.f / lambda, 0.25f), 0.0625f, 1.0f);
} else {
zeroscale = 1.f;
}
if (s->psy.bitres.alloc >= 0) {
/**
* Psy granted us extra bits to use, from the reservoire
* adjust for lambda except what psy already did
*/
destbits = s->psy.bitres.alloc
* (lambda / (avctx->global_quality ? avctx->global_quality : 120));
}
if (avctx->flags & CODEC_FLAG_QSCALE) {
/**
* Constant Q-scale doesn't compensate MS coding on its own
* No need to be overly precise, this only controls RD
* adjustment CB limits when going overboard
*/
aacenc: add support for changing options based on a profile This commit adds the ability for a profile to set the default options, as well as for the user to override such options by simply stating them in the command line while still keeping the same profile, as long as those options are still permitted by the profile. Example: setting the profile to aac_low (the default) will turn PNS and IS on. They can be disabled by -aac_pns 0 and -aac_is 0, respectively. Turning on -aac_pred 1 will cause the profile to be elevated to aac_main, as long as no options forbidding aac_main have been entered (like AAC-LTP, which will be pushed soon). A useful feature is that by setting the profile to mpeg2_aac_low, all MPEG4 features will be disabled and if the user tries to enable them then the program will exit with an error. This profile is signalled with the same bitstream as aac_low (MPEG4) but some devices and decoders will fail if any MPEG4 features have been enabled.
2015-10-12 17:50:10 +02:00
if (s->options.mid_side && s->cur_type == TYPE_CPE)
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.
2015-10-11 22:29:50 +02:00
destbits *= 2;
/**
* When using a constant Q-scale, don't adjust bits, just use RD
* Don't let it go overboard, though... 8x psy target is enough
*/
toomanybits = 5800;
toofewbits = destbits / 16;
/** Don't offset scalers, just RD */
sfoffs = sce->ics.num_windows - 1;
rdlambda = sqrtf(rdlambda);
/** search further */
maxits *= 2;
} else {
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
2015-12-01 08:28:36 +02:00
/* When using ABR, be strict, but a reasonable leeway is
* critical to allow RC to smoothly track desired bitrate
* without sudden quality drops that cause audible artifacts.
* Symmetry is also desirable, to avoid systematic bias.
*/
toomanybits = destbits + destbits/8;
toofewbits = destbits - destbits/8;
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.
2015-10-11 22:29:50 +02:00
sfoffs = 0;
rdlambda = sqrtf(rdlambda);
}
/** and zero out above cutoff frequency */
{
int wlen = 1024 / sce->ics.num_windows;
int bandwidth;
/**
* Scale, psy gives us constant quality, this LP only scales
* bitrate by lambda, so we save bits on subjectively unimportant HF
* rather than increase quantization noise. Adjust nominal bitrate
* to effective bitrate according to encoding parameters,
* AAC_CUTOFF_FROM_BITRATE is calibrated for effective bitrate.
*/
float rate_bandwidth_multiplier = 1.5f;
int frame_bit_rate = (avctx->flags & CODEC_FLAG_QSCALE)
? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
: (avctx->bit_rate / avctx->channels);
/** Compensate for extensions that increase efficiency */
if (s->options.pns || s->options.intensity_stereo)
frame_bit_rate *= 1.15f;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
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.
2015-10-11 22:29:50 +02:00
if (avctx->cutoff > 0) {
bandwidth = avctx->cutoff;
} else {
bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
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
2015-12-01 08:28:36 +02:00
s->psy.cutoff = bandwidth;
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.
2015-10-11 22:29:50 +02:00
}
cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
pns_start_pos = NOISE_LOW_LIMIT * 2 * wlen / avctx->sample_rate;
}
/**
* for values above this the decoder might end up in an endless loop
* due to always having more bits than what can be encoded.
*/
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
destbits = FFMIN(destbits, 5800);
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.
2015-10-11 22:29:50 +02:00
toomanybits = FFMIN(toomanybits, 5800);
toofewbits = FFMIN(toofewbits, 5800);
/**
* XXX: some heuristic to determine initial quantizers will reduce search time
* determine zero bands and upper distortion limits
*/
min_spread_thr_r = -1;
max_spread_thr_r = -1;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
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.
2015-10-11 22:29:50 +02:00
for (g = start = 0; g < sce->ics.num_swb; start += sce->ics.swb_sizes[g++]) {
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
int nz = 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.
2015-10-11 22:29:50 +02:00
float uplim = 0.0f, energy = 0.0f, spread = 0.0f;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(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.
2015-10-11 22:29:50 +02:00
if (start >= cutoff || band->energy <= (band->threshold * zeroscale) || band->threshold == 0.0f) {
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
sce->zeroes[(w+w2)*16+g] = 1;
continue;
}
nz = 1;
}
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.
2015-10-11 22:29:50 +02:00
if (!nz) {
uplim = 0.0f;
} else {
nz = 0;
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 * zeroscale) || band->threshold == 0.0f)
continue;
uplim += band->threshold;
energy += band->energy;
spread += band->spread;
nz++;
}
}
uplims[w*16+g] = uplim;
energies[w*16+g] = energy;
nzs[w*16+g] = nz;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
sce->zeroes[w*16+g] = !nz;
allz |= nz;
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
2015-12-01 08:28:36 +02:00
if (nz && sce->can_pns[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.
2015-10-11 22:29:50 +02:00
spread_thr_r[w*16+g] = energy * nz / (uplim * spread);
if (min_spread_thr_r < 0) {
min_spread_thr_r = max_spread_thr_r = spread_thr_r[w*16+g];
} else {
min_spread_thr_r = FFMIN(min_spread_thr_r, spread_thr_r[w*16+g]);
max_spread_thr_r = FFMAX(max_spread_thr_r, spread_thr_r[w*16+g]);
}
}
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
}
}
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.
2015-10-11 22:29:50 +02:00
/** Compute initial scalers */
minscaler = 65535;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
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;
}
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.
2015-10-11 22:29:50 +02:00
/**
* log2f-to-distortion ratio is, technically, 2 (1.5db = 4, but it's power vs level so it's 2).
* But, as offsets are applied, low-frequency signals are too sensitive to the induced distortion,
* so we make scaling more conservative by choosing a lower log2f-to-distortion ratio, and thus
* more robust.
*/
sce->sf_idx[w*16+g] = av_clip(
SCALE_ONE_POS
+ 1.75*log2f(FFMAX(0.00125f,uplims[w*16+g]) / sce->ics.swb_sizes[g])
+ sfoffs,
60, SCALE_MAX_POS);
minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
}
}
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.
2015-10-11 22:29:50 +02:00
/** Clip */
minscaler = av_clip(minscaler, SCALE_ONE_POS - SCALE_DIV_512, SCALE_MAX_POS - SCALE_DIV_512);
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] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF - 1);
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
if (!allz)
return;
aacenc: add SIMD optimizations for abs_pow34 and quantization Performance improvements: quant_bands: with: 681 decicycles in quant_bands, 8388453 runs, 155 skips without: 1190 decicycles in quant_bands, 8388386 runs, 222 skips Around 42% for the function Twoloop coder: abs_pow34: with/without: 7.82s/8.17s Around 4% for the entire encoder Both: with/without: 7.15s/8.17s Around 12% for the entire encoder Fast coder: abs_pow34: with/without: 3.40s/3.77s Around 10% for the entire encoder Both: with/without: 3.02s/3.77s Around 20% faster for the entire encoder Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com> Tested-by: Michael Niedermayer <michael@niedermayer.cc> Reviewed-by: James Almer <jamrial@gmail.com>
2016-10-08 16:59:14 +02:00
s->abs_pow34(s->scoefs, sce->coeffs, 1024);
AAC encoder: memoize quantize_band_cost The bulk of calls to quantize_band_cost are replaced by a call to a version that memoizes, greatly improving performance, since during coefficient search there is a great deal of repeat work. Memoization cannot always be applied, so do this in a different function, and leave the original as-is.
2015-10-12 08:56:22 +02:00
ff_quantize_band_cost_cache_init(s);
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
AAC encoder: fix signed integer overflow Clamp scalefactors by coef2minsf to avoid undefined behavior caused by signed integer overflow. It also avoids clipping of coefficients so it should avoid artifacts as well, on very rare corner cases.
2016-03-30 15:05:03 +02:00
for (i = 0; i < sizeof(minsf) / sizeof(minsf[0]); ++i)
minsf[i] = 0;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
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;
AAC encoder: fix undefined behavior Fix uninitialized access of minsf in short windows Fix potential invocation of coef2minsf(0)
2016-03-30 23:34:08 +02:00
int minsfidx;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
maxvals[w*16+g] = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled);
AAC encoder: fix initialization of minsf In some situations (exactly zeroed DC coeffs) minsf would be initialized with garbage
2016-04-03 20:03:53 +02:00
if (maxvals[w*16+g] > 0) {
AAC encoder: fix undefined behavior Fix uninitialized access of minsf in short windows Fix potential invocation of coef2minsf(0)
2016-03-30 23:34:08 +02:00
minsfidx = coef2minsf(maxvals[w*16+g]);
AAC encoder: fix initialization of minsf In some situations (exactly zeroed DC coeffs) minsf would be initialized with garbage
2016-04-03 20:03:53 +02:00
for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
minsf[(w+w2)*16+g] = minsfidx;
}
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
start += 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.
2015-10-11 22:29:50 +02:00
/**
* Scale uplims to match rate distortion to quality
* bu applying noisy band depriorization and tonal band priorization.
* Maxval-energy ratio gives us an idea of how noisy/tonal the band is.
* If maxval^2 ~ energy, then that band is mostly noise, and we can relax
* rate distortion requirements.
*/
memcpy(euplims, uplims, sizeof(euplims));
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
/** psy already priorizes transients to some extent */
float de_psy_factor = (sce->ics.num_windows > 1) ? 8.0f / sce->ics.group_len[w] : 1.0f;
start = w*128;
for (g = 0; g < sce->ics.num_swb; g++) {
if (nzs[g] > 0) {
float cleanup_factor = ff_sqrf(av_clipf(start / (cutoff * 0.75f), 1.0f, 2.0f));
float energy2uplim = find_form_factor(
sce->ics.group_len[w], sce->ics.swb_sizes[g],
uplims[w*16+g] / (nzs[g] * sce->ics.swb_sizes[w]),
sce->coeffs + start,
nzslope * cleanup_factor);
energy2uplim *= de_psy_factor;
if (!(avctx->flags & CODEC_FLAG_QSCALE)) {
/** In ABR, we need to priorize less and let rate control do its thing */
energy2uplim = sqrtf(energy2uplim);
}
energy2uplim = FFMAX(0.015625f, FFMIN(1.0f, energy2uplim));
uplims[w*16+g] *= av_clipf(rdlambda * energy2uplim, rdmin, rdmax)
* sce->ics.group_len[w];
energy2uplim = find_form_factor(
sce->ics.group_len[w], sce->ics.swb_sizes[g],
uplims[w*16+g] / (nzs[g] * sce->ics.swb_sizes[w]),
sce->coeffs + start,
2.0f);
energy2uplim *= de_psy_factor;
if (!(avctx->flags & CODEC_FLAG_QSCALE)) {
/** In ABR, we need to priorize less and let rate control do its thing */
energy2uplim = sqrtf(energy2uplim);
}
energy2uplim = FFMAX(0.015625f, FFMIN(1.0f, energy2uplim));
euplims[w*16+g] *= av_clipf(rdlambda * energy2uplim * sce->ics.group_len[w],
0.5f, 1.0f);
}
start += sce->ics.swb_sizes[g];
}
}
for (i = 0; i < sizeof(maxsf) / sizeof(maxsf[0]); ++i)
maxsf[i] = SCALE_MAX_POS;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
//perform two-loop search
//outer loop - improve quality
do {
//inner loop - quantize spectrum to fit into given number of bits
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.
2015-10-11 22:29:50 +02:00
int overdist;
int qstep = its ? 1 : 32;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
do {
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.
2015-10-11 22:29:50 +02:00
int changed = 0;
AAC encoder: fix assertion error re SF differences Intermediate results can indeed violate SF delta. Instead of asserting there, just make the code safe, and assert on the final result. Also re-clamp SFs more often in short windows (which tend to violate the restriction when encoding the switch from one window to the other)
2015-10-12 04:00:46 +02:00
prev = -1;
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.
2015-10-11 22:29:50 +02:00
recomprd = 0;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
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;
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.
2015-10-11 22:29:50 +02:00
float qenergy = 0.0f;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
start += 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.
2015-10-11 22:29:50 +02:00
if (sce->can_pns[w*16+g]) {
/** PNS isn't free */
tbits += ff_pns_bits(sce, w, g);
}
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
continue;
}
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;
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.
2015-10-11 22:29:50 +02:00
float sqenergy;
AAC encoder: memoize quantize_band_cost The bulk of calls to quantize_band_cost are replaced by a call to a version that memoizes, greatly improving performance, since during coefficient search there is a great deal of repeat work. Memoization cannot always be applied, so do this in a different function, and leave the original as-is.
2015-10-12 08:56:22 +02:00
dist += quantize_band_cost_cached(s, w + w2, g, coefs + w2*128,
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
scaled + w2*128,
sce->ics.swb_sizes[g],
sce->sf_idx[w*16+g],
cb,
1.0f,
INFINITY,
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.
2015-10-11 22:29:50 +02:00
&b, &sqenergy,
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
0);
bits += b;
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.
2015-10-11 22:29:50 +02:00
qenergy += sqenergy;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
}
dists[w*16+g] = dist - bits;
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.
2015-10-11 22:29:50 +02:00
qenergies[w*16+g] = qenergy;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
if (prev != -1) {
AAC encoder: fix assertion error re SF differences Intermediate results can indeed violate SF delta. Instead of asserting there, just make the code safe, and assert on the final result. Also re-clamp SFs more often in short windows (which tend to violate the restriction when encoding the switch from one window to the other)
2015-10-12 04:00:46 +02:00
int sfdiff = av_clip(sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO, 0, 2*SCALE_MAX_DIFF);
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.
2015-10-11 22:29:50 +02:00
bits += ff_aac_scalefactor_bits[sfdiff];
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
}
tbits += bits;
start += sce->ics.swb_sizes[g];
prev = 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.
2015-10-11 22:29:50 +02:00
if (tbits > toomanybits) {
recomprd = 1;
for (i = 0; i < 128; i++) {
if (sce->sf_idx[i] < (SCALE_MAX_POS - SCALE_DIV_512)) {
int maxsf_i = (tbits > 5800) ? SCALE_MAX_POS : maxsf[i];
int new_sf = FFMIN(maxsf_i, sce->sf_idx[i] + qstep);
if (new_sf != sce->sf_idx[i]) {
sce->sf_idx[i] = new_sf;
changed = 1;
}
}
}
} else if (tbits < toofewbits) {
recomprd = 1;
for (i = 0; i < 128; i++) {
if (sce->sf_idx[i] > SCALE_ONE_POS) {
AAC encoder: fix signed integer overflow Clamp scalefactors by coef2minsf to avoid undefined behavior caused by signed integer overflow. It also avoids clipping of coefficients so it should avoid artifacts as well, on very rare corner cases.
2016-03-30 15:05:03 +02:00
int new_sf = FFMAX3(minsf[i], SCALE_ONE_POS, sce->sf_idx[i] - qstep);
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.
2015-10-11 22:29:50 +02:00
if (new_sf != sce->sf_idx[i]) {
sce->sf_idx[i] = new_sf;
changed = 1;
}
}
}
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
}
qstep >>= 1;
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.
2015-10-11 22:29:50 +02:00
if (!qstep && tbits > toomanybits && sce->sf_idx[0] < 217 && changed)
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
qstep = 1;
} while (qstep);
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.
2015-10-11 22:29:50 +02:00
overdist = 1;
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
2015-12-01 08:28:36 +02:00
fflag = tbits < toofewbits;
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.
2015-10-11 22:29:50 +02:00
for (i = 0; i < 2 && (overdist || recomprd); ++i) {
if (recomprd) {
/** Must recompute distortion */
AAC encoder: fix assertion error re SF differences Intermediate results can indeed violate SF delta. Instead of asserting there, just make the code safe, and assert on the final result. Also re-clamp SFs more often in short windows (which tend to violate the restriction when encoding the switch from one window to the other)
2015-10-12 04:00:46 +02:00
prev = -1;
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.
2015-10-11 22:29:50 +02:00
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;
float qenergy = 0.0f;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
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.
2015-10-11 22:29:50 +02:00
if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
start += sce->ics.swb_sizes[g];
if (sce->can_pns[w*16+g]) {
/** PNS isn't free */
tbits += ff_pns_bits(sce, w, g);
}
continue;
}
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;
float sqenergy;
AAC encoder: memoize quantize_band_cost The bulk of calls to quantize_band_cost are replaced by a call to a version that memoizes, greatly improving performance, since during coefficient search there is a great deal of repeat work. Memoization cannot always be applied, so do this in a different function, and leave the original as-is.
2015-10-12 08:56:22 +02:00
dist += quantize_band_cost_cached(s, w + w2, g, coefs + w2*128,
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.
2015-10-11 22:29:50 +02:00
scaled + w2*128,
sce->ics.swb_sizes[g],
sce->sf_idx[w*16+g],
cb,
1.0f,
INFINITY,
&b, &sqenergy,
0);
bits += b;
qenergy += sqenergy;
}
dists[w*16+g] = dist - bits;
qenergies[w*16+g] = qenergy;
if (prev != -1) {
AAC encoder: fix assertion error re SF differences Intermediate results can indeed violate SF delta. Instead of asserting there, just make the code safe, and assert on the final result. Also re-clamp SFs more often in short windows (which tend to violate the restriction when encoding the switch from one window to the other)
2015-10-12 04:00:46 +02:00
int sfdiff = av_clip(sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO, 0, 2*SCALE_MAX_DIFF);
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.
2015-10-11 22:29:50 +02:00
bits += ff_aac_scalefactor_bits[sfdiff];
}
tbits += bits;
start += sce->ics.swb_sizes[g];
prev = sce->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
2015-12-01 08:28:36 +02:00
if (!i && s->options.pns && its > maxits/2 && tbits > toofewbits) {
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.
2015-10-11 22:29:50 +02:00
float maxoverdist = 0.0f;
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
2015-12-01 08:28:36 +02:00
float ovrfactor = 1.f+(maxits-its)*16.f/maxits;
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.
2015-10-11 22:29:50 +02:00
overdist = recomprd = 0;
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
for (g = start = 0; g < sce->ics.num_swb; start += sce->ics.swb_sizes[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
2015-12-01 08:28:36 +02:00
if (!sce->zeroes[w*16+g] && sce->sf_idx[w*16+g] > SCALE_ONE_POS && dists[w*16+g] > uplims[w*16+g]*ovrfactor) {
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.
2015-10-11 22:29:50 +02:00
float ovrdist = dists[w*16+g] / FFMAX(uplims[w*16+g],euplims[w*16+g]);
maxoverdist = FFMAX(maxoverdist, ovrdist);
overdist++;
}
}
}
if (overdist) {
/* We have overdistorted bands, trade for zeroes (that can be noise)
* Zero the bands in the lowest 1.25% spread-energy-threshold ranking
*/
float minspread = max_spread_thr_r;
float maxspread = min_spread_thr_r;
float zspread;
int zeroable = 0;
int zeroed = 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
2015-12-01 08:28:36 +02:00
int maxzeroed, zloop;
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.
2015-10-11 22:29:50 +02:00
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
for (g = start = 0; g < sce->ics.num_swb; start += sce->ics.swb_sizes[g++]) {
if (start >= pns_start_pos && !sce->zeroes[w*16+g] && sce->can_pns[w*16+g]) {
minspread = FFMIN(minspread, spread_thr_r[w*16+g]);
maxspread = FFMAX(maxspread, spread_thr_r[w*16+g]);
zeroable++;
}
}
}
zspread = (maxspread-minspread) * 0.0125f + minspread;
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
2015-12-01 08:28:36 +02:00
/* Don't PNS everything even if allowed. It suppresses bit starvation signals from RC,
* and forced the hand of the later search_for_pns step.
* Instead, PNS a fraction of the spread_thr_r range depending on how starved for bits we are,
* and leave further PNSing to search_for_pns if worthwhile.
*/
zspread = FFMIN3(min_spread_thr_r * 8.f, zspread,
((toomanybits - tbits) * min_spread_thr_r + (tbits - toofewbits) * max_spread_thr_r) / (toomanybits - toofewbits + 1));
maxzeroed = FFMIN(zeroable, FFMAX(1, (zeroable * its + maxits - 1) / (2 * maxits)));
for (zloop = 0; zloop < 2; zloop++) {
/* Two passes: first distorted stuff - two birds in one shot and all that,
* then anything viable. Viable means not zero, but either CB=zero-able
* (too high SF), not SF <= 1 (that means we'd be operating at very high
* quality, we don't want PNS when doing VHQ), PNS allowed, and within
* the lowest ranking percentile.
*/
float loopovrfactor = (zloop) ? 1.0f : ovrfactor;
int loopminsf = (zloop) ? (SCALE_ONE_POS - SCALE_DIV_512) : SCALE_ONE_POS;
int mcb;
for (g = sce->ics.num_swb-1; g > 0 && zeroed < maxzeroed; g--) {
if (sce->ics.swb_offset[g] < pns_start_pos)
continue;
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
if (!sce->zeroes[w*16+g] && sce->can_pns[w*16+g] && spread_thr_r[w*16+g] <= zspread
&& sce->sf_idx[w*16+g] > loopminsf
&& (dists[w*16+g] > loopovrfactor*uplims[w*16+g] || !(mcb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]))
|| (mcb <= 1 && dists[w*16+g] > FFMIN(uplims[w*16+g], euplims[w*16+g]))) ) {
sce->zeroes[w*16+g] = 1;
sce->band_type[w*16+g] = 0;
zeroed++;
}
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.
2015-10-11 22:29:50 +02:00
}
}
}
if (zeroed)
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
2015-12-01 08:28:36 +02:00
recomprd = fflag = 1;
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.
2015-10-11 22:29:50 +02:00
} else {
overdist = 0;
}
}
}
minscaler = SCALE_MAX_POS;
maxscaler = 0;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
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.
2015-10-11 22:29:50 +02:00
for (g = 0; g < sce->ics.num_swb; g++) {
if (!sce->zeroes[w*16+g]) {
minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
maxscaler = FFMAX(maxscaler, sce->sf_idx[w*16+g]);
}
}
}
minscaler = nminscaler = av_clip(minscaler, SCALE_ONE_POS - SCALE_DIV_512, SCALE_MAX_POS - SCALE_DIV_512);
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
2015-12-01 08:28:36 +02:00
prev = -1;
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.
2015-10-11 22:29:50 +02:00
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
/** Start with big steps, end up fine-tunning */
int depth = (its > maxits/2) ? ((its > maxits*2/3) ? 1 : 3) : 10;
int edepth = depth+2;
float uplmax = its / (maxits*0.25f) + 1.0f;
uplmax *= (tbits > destbits) ? FFMIN(2.0f, tbits / (float)FFMAX(1,destbits)) : 1.0f;
start = w * 128;
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
for (g = 0; g < sce->ics.num_swb; g++) {
int prevsc = sce->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
2015-12-01 08:28:36 +02:00
if (prev < 0 && !sce->zeroes[w*16+g])
prev = sce->sf_idx[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.
2015-10-11 22:29:50 +02:00
if (!sce->zeroes[w*16+g]) {
const float *coefs = sce->coeffs + start;
const float *scaled = s->scoefs + start;
int cmb = find_min_book(maxvals[w*16+g], sce->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
2015-12-01 08:28:36 +02:00
int mindeltasf = FFMAX(0, prev - SCALE_MAX_DIFF);
int maxdeltasf = FFMIN(SCALE_MAX_POS - SCALE_DIV_512, prev + SCALE_MAX_DIFF);
AAC encoder: fix signed integer overflow Clamp scalefactors by coef2minsf to avoid undefined behavior caused by signed integer overflow. It also avoids clipping of coefficients so it should avoid artifacts as well, on very rare corner cases.
2016-03-30 15:05:03 +02:00
if ((!cmb || dists[w*16+g] > uplims[w*16+g]) && sce->sf_idx[w*16+g] > FFMAX(mindeltasf, minsf[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.
2015-10-11 22:29:50 +02:00
/* Try to make sure there is some energy in every nonzero band
* NOTE: This algorithm must be forcibly imbalanced, pushing harder
* on holes or more distorted bands at first, otherwise there's
* no net gain (since the next iteration will offset all bands
* on the opposite direction to compensate for extra bits)
*/
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
2015-12-01 08:28:36 +02:00
for (i = 0; i < edepth && sce->sf_idx[w*16+g] > mindeltasf; ++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.
2015-10-11 22:29:50 +02:00
int cb, bits;
float dist, qenergy;
int mb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]-1);
cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
dist = qenergy = 0.f;
bits = 0;
if (!cb) {
maxsf[w*16+g] = FFMIN(sce->sf_idx[w*16+g]-1, maxsf[w*16+g]);
} else if (i >= depth && dists[w*16+g] < euplims[w*16+g]) {
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
2015-12-01 08:28:36 +02:00
/* !g is the DC band, it's important, since quantization error here
* applies to less than a cycle, it creates horrible intermodulation
* distortion if it doesn't stick to what psy requests
*/
if (!g && sce->ics.num_windows > 1 && dists[w*16+g] >= euplims[w*16+g])
maxsf[w*16+g] = FFMIN(sce->sf_idx[w*16+g], maxsf[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.
2015-10-11 22:29:50 +02:00
for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
int b;
float sqenergy;
AAC encoder: memoize quantize_band_cost The bulk of calls to quantize_band_cost are replaced by a call to a version that memoizes, greatly improving performance, since during coefficient search there is a great deal of repeat work. Memoization cannot always be applied, so do this in a different function, and leave the original as-is.
2015-10-12 08:56:22 +02:00
dist += quantize_band_cost_cached(s, w + w2, g, coefs + w2*128,
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.
2015-10-11 22:29:50 +02:00
scaled + w2*128,
sce->ics.swb_sizes[g],
sce->sf_idx[w*16+g]-1,
cb,
1.0f,
INFINITY,
&b, &sqenergy,
0);
bits += b;
qenergy += sqenergy;
}
sce->sf_idx[w*16+g]--;
dists[w*16+g] = dist - bits;
qenergies[w*16+g] = qenergy;
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
2015-12-01 08:28:36 +02:00
if (mb && (sce->sf_idx[w*16+g] < mindeltasf || (
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.
2015-10-11 22:29:50 +02:00
(dists[w*16+g] < FFMIN(uplmax*uplims[w*16+g], euplims[w*16+g]))
&& (fabsf(qenergies[w*16+g]-energies[w*16+g]) < euplims[w*16+g])
) )) {
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
2015-12-01 08:28:36 +02:00
} else if (tbits > toofewbits && sce->sf_idx[w*16+g] < FFMIN(maxdeltasf, maxsf[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.
2015-10-11 22:29:50 +02:00
&& (dists[w*16+g] < FFMIN(euplims[w*16+g], uplims[w*16+g]))
&& (fabsf(qenergies[w*16+g]-energies[w*16+g]) < euplims[w*16+g])
) {
/** Um... over target. Save bits for more important stuff. */
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
2015-12-01 08:28:36 +02:00
for (i = 0; i < depth && sce->sf_idx[w*16+g] < maxdeltasf; ++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.
2015-10-11 22:29:50 +02:00
int cb, bits;
float dist, qenergy;
cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]+1);
if (cb > 0) {
dist = qenergy = 0.f;
bits = 0;
for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
int b;
float sqenergy;
AAC encoder: memoize quantize_band_cost The bulk of calls to quantize_band_cost are replaced by a call to a version that memoizes, greatly improving performance, since during coefficient search there is a great deal of repeat work. Memoization cannot always be applied, so do this in a different function, and leave the original as-is.
2015-10-12 08:56:22 +02:00
dist += quantize_band_cost_cached(s, w + w2, g, coefs + w2*128,
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.
2015-10-11 22:29:50 +02:00
scaled + w2*128,
sce->ics.swb_sizes[g],
sce->sf_idx[w*16+g]+1,
cb,
1.0f,
INFINITY,
&b, &sqenergy,
0);
bits += b;
qenergy += sqenergy;
}
dist -= bits;
if (dist < FFMIN(euplims[w*16+g], uplims[w*16+g])) {
sce->sf_idx[w*16+g]++;
dists[w*16+g] = dist;
qenergies[w*16+g] = qenergy;
} else {
break;
}
} else {
maxsf[w*16+g] = FFMIN(sce->sf_idx[w*16+g], maxsf[w*16+g]);
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
2015-12-01 08:28:36 +02:00
prev = sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], mindeltasf, maxdeltasf);
if (sce->sf_idx[w*16+g] != prevsc)
fflag = 1;
nminscaler = FFMIN(nminscaler, sce->sf_idx[w*16+g]);
sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
}
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.
2015-10-11 22:29:50 +02:00
start += sce->ics.swb_sizes[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
2015-12-01 08:28:36 +02:00
/** SF difference limit violation risk. Must re-clamp. */
prev = -1;
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]) {
int prevsf = sce->sf_idx[w*16+g];
if (prev < 0)
prev = prevsf;
sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], prev - SCALE_MAX_DIFF, prev + SCALE_MAX_DIFF);
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.
2015-10-11 22:29:50 +02:00
sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->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
2015-12-01 08:28:36 +02:00
prev = sce->sf_idx[w*16+g];
if (!fflag && prevsf != sce->sf_idx[w*16+g])
fflag = 1;
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.
2015-10-11 22:29:50 +02:00
}
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
}
}
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
2015-12-01 08:28:36 +02:00
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
its++;
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.
2015-10-11 22:29:50 +02:00
} while (fflag && its < maxits);
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
2015-12-01 08:28:36 +02:00
/** Scout out next nonzero bands */
ff_init_nextband_map(sce, nextband);
AAC encoder: fix assertion error re SF differences Intermediate results can indeed violate SF delta. Instead of asserting there, just make the code safe, and assert on the final result. Also re-clamp SFs more often in short windows (which tend to violate the restriction when encoding the switch from one window to the other)
2015-10-12 04:00:46 +02:00
prev = -1;
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.
2015-10-11 22:29:50 +02:00
for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
/** Make sure proper codebooks are set */
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
2015-12-01 08:28:36 +02:00
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.
2015-10-11 22:29:50 +02:00
if (!sce->zeroes[w*16+g]) {
sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
if (sce->band_type[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
2015-12-01 08:28:36 +02:00
if (!ff_sfdelta_can_remove_band(sce, nextband, prev, w*16+g)) {
/** Cannot zero out, make sure it's not attempted */
sce->band_type[w*16+g] = 1;
} else {
sce->zeroes[w*16+g] = 1;
sce->band_type[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.
2015-10-11 22:29:50 +02:00
}
} else {
sce->band_type[w*16+g] = 0;
}
AAC encoder: fix assertion error re SF differences Intermediate results can indeed violate SF delta. Instead of asserting there, just make the code safe, and assert on the final result. Also re-clamp SFs more often in short windows (which tend to violate the restriction when encoding the switch from one window to the other)
2015-10-12 04:00:46 +02:00
/** Check that there's no SF delta range violations */
if (!sce->zeroes[w*16+g]) {
if (prev != -1) {
aaccoder_twoloop: Mark sfdiff as av_unused Silences warning when building without assertions Signed-off-by: Claudio Freire <klaussfreire@gmail.com>
2015-11-26 08:46:09 +02:00
av_unused int sfdiff = sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO;
AAC encoder: fix assertion error re SF differences Intermediate results can indeed violate SF delta. Instead of asserting there, just make the code safe, and assert on the final result. Also re-clamp SFs more often in short windows (which tend to violate the restriction when encoding the switch from one window to the other)
2015-10-12 04:00:46 +02:00
av_assert1(sfdiff >= 0 && sfdiff <= 2*SCALE_MAX_DIFF);
AAC encoder: fix wrong gain sacalefactor being set In some conditions, where the first band was being zeroed mainly, the wrong global gain scalefactor would be written to the stream since it's always taken from the first band regardless of whether it's been marked as zero or not. So, always make sure it contians something useful.
2015-11-26 08:37:29 +02:00
} else if (sce->zeroes[0]) {
/** Set global gain to something useful */
sce->sf_idx[0] = sce->sf_idx[w*16+g];
AAC encoder: fix assertion error re SF differences Intermediate results can indeed violate SF delta. Instead of asserting there, just make the code safe, and assert on the final result. Also re-clamp SFs more often in short windows (which tend to violate the restriction when encoding the switch from one window to the other)
2015-10-12 04:00:46 +02:00
}
prev = 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.
2015-10-11 22:29:50 +02:00
}
}
AAC encoder: refactor to resynchronize MIPS port This patch refactors the AAC coders to reuse code between the MIPS port and the regular, portable C code. There were two main functions that had to use hand-optimized versions of quantization code: - search_for_quantizers_twoloop - codebook_trellis_rate Those two were split into their own template header files so they can be inlined inside both the MIPS port and the generic code. In each context, they'll link to their specialized implementations, and thus be optimized by the compiler. This approach I believe is better than maintaining several copies of each function. As past experience has proven, having to keep those in sync was error prone. In this way, they will remain in sync by default. Also, an implementation of the dequantized output argument for the optimized quantize_and_encode functions is included in the patch. While the current implementation of search_for_pred still isn't using it, future iterations of main prediction probably will. It should not imply any measurable performance hit while not being used.
2015-09-15 08:59:45 +02:00
}
#endif /* AVCODEC_AACCODER_TWOLOOP_H */
Reference in New Issue Copy Permalink