Go to the documentation of this file.
44 #define PSY_3GPP_THR_SPREAD_HI 1.5f // spreading factor for low-to-hi threshold spreading (15 dB/Bark)
45 #define PSY_3GPP_THR_SPREAD_LOW 3.0f // spreading factor for hi-to-low threshold spreading (30 dB/Bark)
47 #define PSY_3GPP_EN_SPREAD_HI_L1 2.0f
49 #define PSY_3GPP_EN_SPREAD_HI_L2 1.5f
51 #define PSY_3GPP_EN_SPREAD_HI_S 1.5f
53 #define PSY_3GPP_EN_SPREAD_LOW_L 3.0f
55 #define PSY_3GPP_EN_SPREAD_LOW_S 2.0f
57 #define PSY_3GPP_RPEMIN 0.01f
58 #define PSY_3GPP_RPELEV 2.0f
60 #define PSY_3GPP_C1 3.0f
61 #define PSY_3GPP_C2 1.3219281f
62 #define PSY_3GPP_C3 0.55935729f
64 #define PSY_SNR_1DB 7.9432821e-1f
65 #define PSY_SNR_25DB 3.1622776e-3f
67 #define PSY_3GPP_SAVE_SLOPE_L -0.46666667f
68 #define PSY_3GPP_SAVE_SLOPE_S -0.36363637f
69 #define PSY_3GPP_SAVE_ADD_L -0.84285712f
70 #define PSY_3GPP_SAVE_ADD_S -0.75f
71 #define PSY_3GPP_SPEND_SLOPE_L 0.66666669f
72 #define PSY_3GPP_SPEND_SLOPE_S 0.81818181f
73 #define PSY_3GPP_SPEND_ADD_L -0.35f
74 #define PSY_3GPP_SPEND_ADD_S -0.26111111f
75 #define PSY_3GPP_CLIP_LO_L 0.2f
76 #define PSY_3GPP_CLIP_LO_S 0.2f
77 #define PSY_3GPP_CLIP_HI_L 0.95f
78 #define PSY_3GPP_CLIP_HI_S 0.75f
80 #define PSY_3GPP_AH_THR_LONG 0.5f
81 #define PSY_3GPP_AH_THR_SHORT 0.63f
83 #define PSY_PE_FORGET_SLOPE 511
91 #define PSY_3GPP_BITS_TO_PE(bits) ((bits) * 1.18f)
92 #define PSY_3GPP_PE_TO_BITS(bits) ((bits) / 1.18f)
95 #define PSY_LAME_FIR_LEN 21
96 #define AAC_BLOCK_SIZE_LONG 1024
97 #define AAC_BLOCK_SIZE_SHORT 128
98 #define AAC_NUM_BLOCKS_SHORT 8
99 #define PSY_LAME_NUM_SUBBLOCKS 3
220 -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2,
221 -3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2,
222 -5.52212e-17 * 2, -0.313819 * 2
235 int lower_range = 12, upper_range = 12;
243 for (
i = 1;
i < 13;
i++) {
284 return 13.3f *
atanf(0.00076
f *
f) + 3.5f *
atanf((
f / 7500.0
f) * (
f / 7500.0
f));
295 return 3.64 * pow(
f, -0.8)
296 - 6.8 *
exp(-0.6 * (
f - 3.4) * (
f - 3.4))
297 + 6.0 *
exp(-0.15 * (
f - 8.7) * (
f - 8.7))
298 + (0.6 + 0.04 * add) * 0.001 *
f *
f *
f *
f;
305 float prev, minscale, minath, minsnr, pe_min;
309 const float num_bark =
calc_bark((
float)bandwidth);
315 if (!
ctx->model_priv_data)
317 pctx =
ctx->model_priv_data;
322 chan_bitrate = (
int)(chan_bitrate / 120.0 * (
ctx->avctx->global_quality ?
ctx->avctx->global_quality : 120));
330 ctx->bitres.size -=
ctx->bitres.size % 8;
333 for (j = 0; j < 2; j++) {
335 const uint8_t *band_sizes =
ctx->bands[j];
336 float line_to_frequency =
ctx->avctx->sample_rate / (j ? 256.f : 2048.0f);
337 float avg_chan_bits = chan_bitrate * (j ? 128.0f : 1024.0f) /
ctx->avctx->sample_rate;
346 for (
g = 0;
g <
ctx->num_bands[j];
g++) {
349 coeffs[
g].
barks = (bark + prev) / 2.0;
352 for (
g = 0;
g <
ctx->num_bands[j] - 1;
g++) {
354 float bark_width = coeffs[
g+1].
barks - coeffs->
barks;
357 coeff->spread_low[1] =
ff_exp10(-bark_width * en_spread_low);
359 pe_min = bark_pe * bark_width;
360 minsnr =
exp2(pe_min / band_sizes[
g]) - 1.5f;
364 for (
g = 0;
g <
ctx->num_bands[j];
g++) {
365 minscale =
ath(start * line_to_frequency,
ATH_ADD);
366 for (
i = 1;
i < band_sizes[
g];
i++)
368 coeffs[
g].
ath = minscale - minath;
369 start += band_sizes[
g];
401 0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36
409 const int16_t *audio,
415 int attack_ratio = br <= 16000 ? 18 : 10;
418 uint8_t grouping = 0;
424 int switch_to_eight = 0;
425 float sum = 0.0, sum2 = 0.0;
428 for (
i = 0;
i < 8;
i++) {
429 for (j = 0; j < 128; j++) {
436 for (
i = 0;
i < 8;
i++) {
437 if (
s[
i] > pch->win_energy * attack_ratio) {
443 pch->win_energy = pch->win_energy*7/8 + sum2/64;
445 wi.window_type[1] = prev_type;
453 grouping = pch->next_grouping;
469 pch->next_window_seq = next_type;
471 for (
i = 0;
i < 3;
i++)
472 wi.window_type[
i] = prev_type;
483 for (
i = 0;
i < 8;
i++) {
484 if (!((grouping >>
i) & 1))
486 wi.grouping[lastgrp]++;
503 float clipped_pe, bit_save, bit_spend, bit_factor, fill_level, forgetful_min_pe;
507 fill_level =
av_clipf((
float)
ctx->fill_level /
size, clip_low, clip_high);
509 bit_save = (fill_level + bitsave_add) * bitsave_slope;
510 assert(bit_save <= 0.3f && bit_save >= -0.05000001
f);
511 bit_spend = (fill_level + bitspend_add) * bitspend_slope;
512 assert(bit_spend <= 0.5f && bit_spend >= -0.1
f);
519 bit_factor = 1.0f - bit_save + ((bit_spend - bit_save) / (
ctx->pe.max -
ctx->pe.min)) * (clipped_pe -
ctx->pe.min);
527 ctx->pe.min =
FFMIN(pe, forgetful_min_pe);
533 ctx->frame_bits * bit_factor,
563 float thr_avg, reduction;
565 if(active_lines == 0.0)
568 thr_avg =
exp2f((
a - pe) / (4.0
f * active_lines));
569 reduction =
exp2f((
a - desired_pe) / (4.0
f * active_lines)) - thr_avg;
571 return FFMAX(reduction, 0.0
f);
577 float thr = band->
thr;
581 thr =
sqrtf(thr) + reduction;
599 #ifndef calc_thr_3gpp
601 const uint8_t *band_sizes,
const float *coefs,
const int cutoff)
604 int start = 0, wstart = 0;
607 for (
g = 0;
g < num_bands;
g++) {
610 float form_factor = 0.0f;
613 if (wstart < cutoff) {
614 for (
i = 0;
i < band_sizes[
g];
i++) {
615 band->
energy += coefs[start+
i] * coefs[start+
i];
623 start += band_sizes[
g];
624 wstart += band_sizes[
g];
630 #ifndef psy_hp_filter
644 hpfsmpl[
i] = (sum1 + sum2) * 32768.0
f;
658 float desired_bits, desired_pe, delta_pe, reduction=
NAN, spread_en[128] = {0};
659 float a = 0.0f, active_lines = 0.0f, norm_fac = 0.0f;
660 float pe = pctx->chan_bitrate > 32000 ? 0.0f :
FFMAX(50.0
f, 100.0
f - pctx->chan_bitrate * 100.0f / 32000.0f);
661 const int num_bands =
ctx->num_bands[wi->num_windows == 8];
662 const uint8_t *band_sizes =
ctx->bands[wi->num_windows == 8];
663 AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8];
666 const int cutoff = bandwidth * 2048 / wi->num_windows /
ctx->avctx->sample_rate;
669 calc_thr_3gpp(wi, num_bands, pch, band_sizes, coefs, cutoff);
672 for (
w = 0;
w < wi->num_windows*16;
w += 16) {
676 spread_en[0] =
bands[0].energy;
677 for (
g = 1;
g < num_bands;
g++) {
679 spread_en[
w+
g] =
FFMAX(
bands[
g].energy, spread_en[
w+
g-1] * coeffs[
g].spread_hi[1]);
681 for (
g = num_bands - 2;
g >= 0;
g--) {
683 spread_en[
w+
g] =
FFMAX(spread_en[
w+
g], spread_en[
w+
g+1] * coeffs[
g].spread_low[1]);
686 for (
g = 0;
g < num_bands;
g++) {
701 if (spread_en[
w+
g] * avoid_hole_thr > band->
energy || coeffs[
g].min_snr > 1.0f)
714 desired_pe = pe * (
ctx->avctx->global_quality ?
ctx->avctx->global_quality : 120) / (2 * 2.5
f * 120.0
f);
719 if (
ctx->bitres.bits > 0) {
724 pctx->pe.max =
FFMAX(pe, pctx->pe.max);
725 pctx->pe.min =
FFMIN(pe, pctx->pe.min);
734 if (
ctx->bitres.bits > 0)
739 ctx->bitres.alloc = desired_bits;
741 if (desired_pe < pe) {
743 for (
w = 0;
w < wi->num_windows*16;
w += 16) {
748 for (
g = 0;
g < num_bands;
g++) {
760 for (
i = 0;
i < 2;
i++) {
761 float pe_no_ah = 0.0f, desired_pe_no_ah;
762 active_lines =
a = 0.0f;
763 for (
w = 0;
w < wi->num_windows*16;
w += 16) {
764 for (
g = 0;
g < num_bands;
g++) {
768 pe_no_ah += band->
pe;
774 desired_pe_no_ah =
FFMAX(desired_pe - (pe - pe_no_ah), 0.0
f);
775 if (active_lines > 0.0
f)
779 for (
w = 0;
w < wi->num_windows*16;
w += 16) {
780 for (
g = 0;
g < num_bands;
g++) {
783 if (active_lines > 0.0
f)
786 if (band->
thr > 0.0f)
793 delta_pe = desired_pe - pe;
794 if (
fabs(delta_pe) > 0.05
f * desired_pe)
798 if (pe < 1.15
f * desired_pe) {
800 norm_fac = norm_fac ? 1.0f / norm_fac : 0;
801 for (
w = 0;
w < wi->num_windows*16;
w += 16) {
802 for (
g = 0;
g < num_bands;
g++) {
806 float delta_sfb_pe = band->
norm_fac * norm_fac * delta_pe;
807 float thr = band->
thr;
819 while (pe > desired_pe &&
g--) {
820 for (
w = 0;
w < wi->num_windows*16;
w+= 16) {
833 for (
w = 0;
w < wi->num_windows*16;
w += 16) {
834 for (
g = 0;
g < num_bands;
g++) {
845 memcpy(pch->prev_band, pch->band,
sizeof(pch->band));
854 for (ch = 0; ch < group->
num_ch; ch++)
881 ctx->next_window_seq = blocktype;
885 const float *la,
int channel,
int prev_type)
890 int uselongblock = 1;
897 const float *pf = hpfsmpl;
912 energy_short[0] += energy_subshort[
i];
918 for (; pf < pfe; pf++)
929 if (p > energy_subshort[
i + 1])
930 p = p / energy_subshort[
i + 1];
931 else if (energy_subshort[
i + 1] > p * 10.0
f)
932 p = energy_subshort[
i + 1] / (p * 10.0f);
941 if (attack_intensity[
i] > pch->attack_threshold)
949 const float u = energy_short[
i - 1];
950 const float v = energy_short[
i];
951 const float m =
FFMAX(
u, v);
953 if (
u < 1.7
f * v && v < 1.7
f *
u) {
954 if (
i == 1 && attacks[0] < attacks[
i])
959 att_sum += attacks[
i];
962 if (attacks[0] <= pch->prev_attack)
965 att_sum += attacks[0];
967 if (pch->prev_attack == 3 || att_sum) {
971 if (attacks[
i] && attacks[
i-1])
996 for (
i = 0;
i < 8;
i++) {
997 if (!((pch->next_grouping >>
i) & 1))
1009 for (
i = 0;
i < 9;
i++) {
1017 pch->prev_attack = attacks[8];
1024 .
name =
"3GPP TS 26.403-inspired model",
float spread_low[2]
spreading factor for low-to-high threshold spreading in long frame
static av_always_inline double ff_exp10(double x)
Compute 10^x for floating point values.
static av_cold int psy_3gpp_init(FFPsyContext *ctx)
Filter the word “frame” indicates either a video frame or a group of audio as stored in an AVFrame structure Format for each input and each output the list of supported formats For video that means pixel format For audio that means channel sample they are references to shared objects When the negotiation mechanism computes the intersection of the formats supported at each end of a all references to both lists are replaced with a reference to the intersection And when a single format is eventually chosen for a link amongst the remaining all references to the list are updated That means that if a filter requires that its input and output have the same format amongst a supported all it has to do is use a reference to the same list of formats query_formats can leave some formats unset and return AVERROR(EAGAIN) to cause the negotiation mechanism toagain later. That can be used by filters with complex requirements to use the format negotiated on one link to set the formats supported on another. Frame references ownership and permissions
static av_unused FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx, const int16_t *audio, const int16_t *la, int channel, int prev_type)
Tell encoder which window types to use.
static float lame_calc_attack_threshold(int bitrate)
Calculate the ABR attack threshold from the above LAME psymodel table.
#define u(width, name, range_min, range_max)
#define PSY_PE_FORGET_SLOPE
static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx, const float *audio, const float *la, int channel, int prev_type)
float thr
energy threshold
static void calc_thr_3gpp(const FFPsyWindowInfo *wi, const int num_bands, AacPsyChannel *pch, const uint8_t *band_sizes, const float *coefs, const int cutoff)
#define PSY_3GPP_PE_TO_BITS(bits)
#define AV_CODEC_FLAG_QSCALE
Use fixed qscale.
static av_cold float calc_bark(float f)
Calculate Bark value for given line.
float nz_lines
number of non-zero spectral lines
#define PSY_3GPP_CLIP_LO_S
#define PSY_3GPP_AH_THR_LONG
int window_shape
window shape (sine/KBD/whatever)
static float calc_pe_3gpp(AacPsyBand *band)
float min
minimum allowed PE for bit factor calculation
#define PSY_3GPP_SPEND_SLOPE_L
#define PSY_3GPP_THR_SPREAD_HI
constants for 3GPP AAC psychoacoustic model
int fill_level
bit reservoir fill level
int nb_channels
Number of channels in this layout.
float spread_hi[2]
spreading factor for high-to-low threshold spreading in long frame
trying all byte sequences megabyte in length and selecting the best looking sequence will yield cases to try But a word about quality
static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock)
psychoacoustic model frame type-dependent coefficients
AVChannelLayout ch_layout
Audio channel layout.
static av_cold void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx)
LAME psy model specific initialization.
float st_lrm
short threshold for L, R, and M channels
#define PSY_3GPP_EN_SPREAD_HI_S
#define PSY_3GPP_SPEND_ADD_L
int flags
AV_CODEC_FLAG_*.
float barks
Bark value for each spectral band in long frame.
float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT *PSY_LAME_NUM_SUBBLOCKS]
static __device__ float fabsf(float a)
windowing related information
int64_t bit_rate
Total stream bitrate in bit/s, 0 if not available.
float previous
allowed PE of the previous frame
const FFPsyModel ff_aac_psy_model
uint8_t num_ch
number of channels in this group
LAME psy model preset struct.
struct AacPsyContext::@15 pe
#define PSY_3GPP_CLIP_HI_S
information for single band used by 3GPP TS26.403-inspired psychoacoustic model
int global_quality
Global quality for codecs which cannot change it per frame.
int flags
Flags modifying the (de)muxer behaviour.
int quality
Quality to map the rest of the vaules to.
float pe_const
constant part of the PE calculation
3GPP TS26.403-inspired psychoacoustic model specific data
static float calc_reduction_3gpp(float a, float desired_pe, float pe, float active_lines)
static const uint8_t window_grouping[9]
window grouping information stored as bits (0 - new group, 1 - group continues)
#define AAC_BLOCK_SIZE_SHORT
short block size
static const float bands[]
static av_cold float ath(float f, float add)
Calculate ATH value for given frequency.
static int calc_bit_demand(AacPsyContext *ctx, float pe, int bits, int size, int short_window)
#define PSY_3GPP_AH_THR_SHORT
static void psy_hp_filter(const float *firbuf, float *hpfsmpl, const float *psy_fir_coeffs)
static float iir_filter(int in, float state[2])
IIR filter used in block switching decision.
static const PsyLamePreset psy_vbr_map[]
LAME psy model preset table for constant quality.
int window_type[3]
window type (short/long/transitional, etc.) - current, previous and next
static __device__ float fabs(float a)
static const PsyLamePreset psy_abr_map[]
LAME psy model preset table for ABR.
int64_t bit_rate
the average bitrate
static av_cold void psy_3gpp_end(FFPsyContext *apc)
#define PSY_3GPP_BITS_TO_PE(bits)
single band psychoacoustic information
static __device__ float sqrtf(float a)
int grouping[8]
window grouping (for e.g. AAC)
float max
maximum allowed PE for bit factor calculation
float iir_state[2]
hi-pass IIR filter state
AacPsyCoeffs psy_coef[2][64]
float thr_quiet
threshold in quiet
#define AAC_BLOCK_SIZE_LONG
long block size
AacPsyBand band[128]
bands information
static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr, float reduction)
float ath
absolute threshold of hearing per bands
float active_lines
number of active spectral lines
#define AAC_NUM_BLOCKS_SHORT
number of blocks in a short sequence
#define PSY_LAME_FIR_LEN
LAME psy model FIR order.
The reader does not expect b to be semantically here and if the code is changed by maybe adding a a division or other the signedness will almost certainly be mistaken To avoid this confusion a new type was SUINT is the C unsigned type but it holds a signed int to use the same example SUINT a
#define PSY_3GPP_CLIP_LO_L
int avoid_holes
hole avoidance flag
#define PSY_3GPP_THR_SPREAD_LOW
#define PSY_3GPP_SAVE_ADD_S
#define PSY_3GPP_SPEND_ADD_S
static const float psy_fir_coeffs[]
LAME psy model FIR coefficient table.
float attack_threshold
attack threshold for this channel
#define i(width, name, range_min, range_max)
float norm_fac
normalization factor for linearization
#define PSY_3GPP_CLIP_HI_L
float pe
perceptual entropy
void * av_mallocz(size_t size)
Allocate a memory block with alignment suitable for all memory accesses (including vectors if availab...
void * av_calloc(size_t nmemb, size_t size)
psychoacoustic information for an arbitrary group of channels
enum WindowSequence next_window_seq
window sequence to be used in the next frame
float win_energy
sliding average of channel energy
single/pair channel context for psychoacoustic model
float correction
PE correction factor.
void * model_priv_data
psychoacoustic model implementation private data
#define PSY_3GPP_SAVE_SLOPE_S
#define PSY_3GPP_EN_SPREAD_HI_L1
uint8_t next_grouping
stored grouping scheme for the next frame (in case of 8 short window sequence)
main external API structure.
#define PSY_LAME_NUM_SUBBLOCKS
Number of sub-blocks in each short block.
float global_quality
normalized global quality taken from avctx
static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel, const float *coefs, const FFPsyWindowInfo *wi)
Calculate band thresholds as suggested in 3GPP TS26.403.
codec-specific psychoacoustic model implementation
int frame_bits
average bits per frame
FFPsyChannelGroup * ff_psy_find_group(FFPsyContext *ctx, int channel)
Determine what group a channel belongs to.
static void psy_3gpp_analyze(FFPsyContext *ctx, int channel, const float **coeffs, const FFPsyWindowInfo *wi)
#define PSY_3GPP_EN_SPREAD_LOW_L
int chan_bitrate
bitrate per channel
#define PSY_3GPP_SAVE_SLOPE_L
static const double coeff[2][5]
#define PSY_3GPP_SPEND_SLOPE_S
#define FF_QP2LAMBDA
factor to convert from H.263 QP to lambda
#define PSY_3GPP_EN_SPREAD_LOW_S
int prev_attack
attack value for the last short block in the previous sequence
context used by psychoacoustic model
AacPsyBand prev_band[128]
bands information from the previous frame
int num_windows
number of windows in a frame
#define PSY_3GPP_SAVE_ADD_L