The branch, master has been updated
via a06aee359b4626ad4084e398baef4650ed499f71 (commit)
via e11dfb3f09fe63463cc931b0f52f60ebf4830cc8 (commit)
via d907f7d555c4147908493aaefe92af08a47a3df3 (commit)
via 9bded2145fac2e467ad9315b16e6783f36804ef7 (commit)
from d479a4926ac470f2825905c808409eb9dec8441f (commit)
- Log -----------------------------------------------------------------
commit a06aee359b4626ad4084e398baef4650ed499f71
Author: Agent45 <[email protected]>
AuthorDate: Sun Nov 2 17:34:52 2025 +0000
Commit: Lynne <[email protected]>
CommitDate: Tue Nov 4 09:55:51 2025 +0000
revert 458a942481151ede27478e6ac6d9d2866d438b84
revert avcodec/aacenc: add bitrate threshold for PNS
diff --git a/libavcodec/aaccoder.c b/libavcodec/aaccoder.c
index ddebdfd53d..7f1c4cdcc1 100644
--- a/libavcodec/aaccoder.c
+++ b/libavcodec/aaccoder.c
@@ -58,9 +58,6 @@
* replace low energy non zero bands */
#define NOISE_LAMBDA_REPLACE 1.948f
-/* Bitrate threshold (in bits/sec/channel) above which PNS is disabled. */
-#define PNS_BITRATE_LIMIT 64000.0f
-
#include "libavcodec/aaccoder_trellis.h"
typedef float (*quantize_and_encode_band_func)(struct AACEncContext *s,
PutBitContext *pb,
@@ -516,7 +513,6 @@ static void search_for_pns(AACEncContext *s, AVCodecContext
*avctx, SingleChanne
? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
: (avctx->bit_rate / avctx->ch_layout.nb_channels);
- int pns_at_low_bitrate = frame_bit_rate < PNS_BITRATE_LIMIT;
frame_bit_rate *= 1.15f;
if (avctx->cutoff > 0) {
@@ -540,7 +536,7 @@ static void search_for_pns(AACEncContext *s, AVCodecContext
*avctx, SingleChanne
const int start = wstart+sce->ics.swb_offset[g];
const float freq = (start-wstart)*freq_mult;
const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
- if (!pns_at_low_bitrate || freq < NOISE_LOW_LIMIT ||
(start-wstart) >= cutoff) {
+ if (freq < NOISE_LOW_LIMIT || (start-wstart) >= cutoff) {
if (!sce->zeroes[w*16+g])
prev_sf = sce->sf_idx[w*16+g];
continue;
@@ -653,7 +649,6 @@ static void mark_pns(AACEncContext *s, AVCodecContext
*avctx, SingleChannelEleme
? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
: (avctx->bit_rate / avctx->ch_layout.nb_channels);
- int pns_at_low_bitrate = frame_bit_rate < PNS_BITRATE_LIMIT;
frame_bit_rate *= 1.15f;
if (avctx->cutoff > 0) {
@@ -672,7 +667,7 @@ static void mark_pns(AACEncContext *s, AVCodecContext
*avctx, SingleChannelEleme
const int start = sce->ics.swb_offset[g];
const float freq = start*freq_mult;
const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
- if (!pns_at_low_bitrate || freq < NOISE_LOW_LIMIT || start >=
cutoff) {
+ if (freq < NOISE_LOW_LIMIT || start >= cutoff) {
sce->can_pns[w*16+g] = 0;
continue;
}
commit e11dfb3f09fe63463cc931b0f52f60ebf4830cc8
Author: Agent45 <[email protected]>
AuthorDate: Sat Nov 1 22:34:33 2025 +0000
Commit: Lynne <[email protected]>
CommitDate: Tue Nov 4 09:55:51 2025 +0000
trim trailing whitespace in aacpsy.c
diff --git a/libavcodec/aacpsy.c b/libavcodec/aacpsy.c
index b5e5be9f78..fb6ff75795 100644
--- a/libavcodec/aacpsy.c
+++ b/libavcodec/aacpsy.c
@@ -913,7 +913,6 @@ static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx,
const float *audio,
p = FFMAX(p, fabsf(*pf));
pch->prev_energy_subshort[i] = energy_subshort[i +
PSY_LAME_NUM_SUBBLOCKS] = p;
energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p;
-
/* NOTE: The indexes below are [i + 3 - 2] in the LAME source.
Compare each sub-block to sub-block - 2 */
if (p > energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS - 2])
p = p / energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS - 2];
commit d907f7d555c4147908493aaefe92af08a47a3df3
Author: Agent45 <[email protected]>
AuthorDate: Sat Nov 1 20:41:24 2025 +0000
Commit: Lynne <[email protected]>
CommitDate: Tue Nov 4 09:55:51 2025 +0000
avcodec/aacpsy: fix attack detection logic and subblock indexing
Fix several indexing errors in attack detection logic and refine transient
handling in the AAC psychoacoustic model.
- Change PSY_LAME_NUM_SUBBLOCKS from 3 to 2 to ensure full coverage of all
1024 MDCT samples, with each subblock containing exactly 1024 / (8 * 2) = 64
samplesâmatching LAMEâs empirical design.
- Introduce next_attack0_zero state flag to stabilize attack[0] prediction
across frames.
- Adjust attack threshold presets.
These changes improve the handling of periodic signals such as trumpet,
especially under low bitrate conditions.
diff --git a/libavcodec/aacpsy.c b/libavcodec/aacpsy.c
index 7a9c2ce6ad..b5e5be9f78 100644
--- a/libavcodec/aacpsy.c
+++ b/libavcodec/aacpsy.c
@@ -97,7 +97,7 @@ enum {
#define AAC_BLOCK_SIZE_LONG 1024 ///< long block size
#define AAC_BLOCK_SIZE_SHORT 128 ///< short block size
#define AAC_NUM_BLOCKS_SHORT 8 ///< number of blocks in a short sequence
-#define PSY_LAME_NUM_SUBBLOCKS 3 ///< Number of sub-blocks in each short
block
+#define PSY_LAME_NUM_SUBBLOCKS 2 ///< Number of sub-blocks in each short
block
/**
* @}
@@ -133,6 +133,7 @@ typedef struct AacPsyChannel{
float attack_threshold; ///< attack threshold for this channel
float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS];
int prev_attack; ///< attack value for the last short
block in the previous sequence
+ int next_attack0_zero; ///< whether attack[0] of the next frame
is zero
}AacPsyChannel;
/**
@@ -181,19 +182,19 @@ typedef struct PsyLamePreset {
static const PsyLamePreset psy_abr_map[] = {
/* TODO: Tuning. These were taken from LAME. */
/* kbps/ch st_lrm */
- { 8, 6.60},
- { 16, 6.60},
- { 24, 6.60},
- { 32, 6.60},
- { 40, 6.60},
- { 48, 6.60},
- { 56, 6.60},
- { 64, 6.40},
- { 80, 6.00},
- { 96, 5.60},
- {112, 5.20},
- {128, 5.20},
- {160, 5.20}
+ { 8, 7.60},
+ { 16, 7.60},
+ { 24, 7.60},
+ { 32, 7.60},
+ { 40, 7.60},
+ { 48, 7.60},
+ { 56, 7.60},
+ { 64, 7.40},
+ { 80, 7.00},
+ { 96, 6.60},
+ {112, 6.20},
+ {128, 6.20},
+ {160, 6.20}
};
/**
@@ -900,8 +901,8 @@ static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx,
const float *audio,
/* Calculate the energies of each sub-shortblock */
for (i = 0; i < PSY_LAME_NUM_SUBBLOCKS; i++) {
energy_subshort[i] = pch->prev_energy_subshort[i +
((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS)];
- assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) *
PSY_LAME_NUM_SUBBLOCKS + 1)] > 0);
- attack_intensity[i] = energy_subshort[i] /
pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) *
PSY_LAME_NUM_SUBBLOCKS + 1)];
+ assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) *
PSY_LAME_NUM_SUBBLOCKS - 2)] > 0);
+ attack_intensity[i] = energy_subshort[i] /
pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) *
PSY_LAME_NUM_SUBBLOCKS - 2)];
energy_short[0] += energy_subshort[i];
}
@@ -912,17 +913,12 @@ static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx,
const float *audio,
p = FFMAX(p, fabsf(*pf));
pch->prev_energy_subshort[i] = energy_subshort[i +
PSY_LAME_NUM_SUBBLOCKS] = p;
energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p;
- /* NOTE: The indexes below are [i + 3 - 2] in the LAME source.
- * Obviously the 3 and 2 have some significance, or this
would be just [i + 1]
- * (which is what we use here). What the 3 stands for is
ambiguous, as it is both
- * number of short blocks, and the number of sub-short
blocks.
- * It seems that LAME is comparing each sub-block to
sub-block + 1 in the
- * previous block.
- */
- if (p > energy_subshort[i + 1])
- p = p / energy_subshort[i + 1];
- else if (energy_subshort[i + 1] > p * 10.0f)
- p = energy_subshort[i + 1] / (p * 10.0f);
+
+ /* NOTE: The indexes below are [i + 3 - 2] in the LAME source.
Compare each sub-block to sub-block - 2 */
+ if (p > energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS - 2])
+ p = p / energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS - 2];
+ else if (energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS - 2] > p *
10.0f)
+ p = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS - 2] / (p *
10.0f);
else
p = 0.0;
attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p;
@@ -943,7 +939,7 @@ static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx,
const float *audio,
const float v = energy_short[i];
const float m = FFMAX(u, v);
if (m < 40000) { /* (2) */
- if (u < 1.7f * v && v < 1.7f * u) { /* (1) */
+ if (u < 2.3f * v && v < 2.3f * u) { /* (1) */
if (i == 1 && attacks[0] < attacks[i])
attacks[0] = 0;
attacks[i] = 0;
@@ -951,13 +947,19 @@ static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx,
const float *audio,
}
att_sum += attacks[i];
}
+ if (pch->next_attack0_zero)
+ attacks[0] = 0;
+ if (attacks[AAC_NUM_BLOCKS_SHORT] == 0)
+ pch->next_attack0_zero = 1;
+ else
+ pch->next_attack0_zero = 0;
if (attacks[0] <= pch->prev_attack)
attacks[0] = 0;
att_sum += attacks[0];
- /* 3 below indicates the previous attack happened in the last
sub-block of the previous sequence */
- if (pch->prev_attack == 3 || att_sum) {
+ /* If the previous attack happened in the last sub-block of the
previous sequence, or if there's a new attack, use short window */
+ if (pch->prev_attack == PSY_LAME_NUM_SUBBLOCKS || att_sum) {
uselongblock = 0;
for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++)
@@ -1007,7 +1009,7 @@ static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx,
const float *audio,
}
pch->next_grouping = window_grouping[grouping];
- pch->prev_attack = attacks[8];
+ pch->prev_attack = attacks[AAC_NUM_BLOCKS_SHORT - 1];
return wi;
}
commit 9bded2145fac2e467ad9315b16e6783f36804ef7
Author: Agent45 <[email protected]>
AuthorDate: Sat Nov 1 19:49:05 2025 +0000
Commit: Lynne <[email protected]>
CommitDate: Tue Nov 4 09:55:51 2025 +0000
avcodec/aacenc: add bitrate threshold for PNS
diff --git a/libavcodec/aaccoder.c b/libavcodec/aaccoder.c
index 7f1c4cdcc1..ddebdfd53d 100644
--- a/libavcodec/aaccoder.c
+++ b/libavcodec/aaccoder.c
@@ -58,6 +58,9 @@
* replace low energy non zero bands */
#define NOISE_LAMBDA_REPLACE 1.948f
+/* Bitrate threshold (in bits/sec/channel) above which PNS is disabled. */
+#define PNS_BITRATE_LIMIT 64000.0f
+
#include "libavcodec/aaccoder_trellis.h"
typedef float (*quantize_and_encode_band_func)(struct AACEncContext *s,
PutBitContext *pb,
@@ -513,6 +516,7 @@ static void search_for_pns(AACEncContext *s, AVCodecContext
*avctx, SingleChanne
? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
: (avctx->bit_rate / avctx->ch_layout.nb_channels);
+ int pns_at_low_bitrate = frame_bit_rate < PNS_BITRATE_LIMIT;
frame_bit_rate *= 1.15f;
if (avctx->cutoff > 0) {
@@ -536,7 +540,7 @@ static void search_for_pns(AACEncContext *s, AVCodecContext
*avctx, SingleChanne
const int start = wstart+sce->ics.swb_offset[g];
const float freq = (start-wstart)*freq_mult;
const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
- if (freq < NOISE_LOW_LIMIT || (start-wstart) >= cutoff) {
+ if (!pns_at_low_bitrate || freq < NOISE_LOW_LIMIT ||
(start-wstart) >= cutoff) {
if (!sce->zeroes[w*16+g])
prev_sf = sce->sf_idx[w*16+g];
continue;
@@ -649,6 +653,7 @@ static void mark_pns(AACEncContext *s, AVCodecContext
*avctx, SingleChannelEleme
? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
: (avctx->bit_rate / avctx->ch_layout.nb_channels);
+ int pns_at_low_bitrate = frame_bit_rate < PNS_BITRATE_LIMIT;
frame_bit_rate *= 1.15f;
if (avctx->cutoff > 0) {
@@ -667,7 +672,7 @@ static void mark_pns(AACEncContext *s, AVCodecContext
*avctx, SingleChannelEleme
const int start = sce->ics.swb_offset[g];
const float freq = start*freq_mult;
const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
- if (freq < NOISE_LOW_LIMIT || start >= cutoff) {
+ if (!pns_at_low_bitrate || freq < NOISE_LOW_LIMIT || start >=
cutoff) {
sce->can_pns[w*16+g] = 0;
continue;
}
-----------------------------------------------------------------------
Summary of changes:
libavcodec/aacpsy.c | 63 +++++++++++++++++++++++++++--------------------------
1 file changed, 32 insertions(+), 31 deletions(-)
hooks/post-receive
--
_______________________________________________
ffmpeg-cvslog mailing list -- [email protected]
To unsubscribe send an email to [email protected]
