Go to the documentation of this file.
45 #define PSY_3GPP_THR_SPREAD_HI 1.5f // spreading factor for low-to-hi threshold spreading (15 dB/Bark)
46 #define PSY_3GPP_THR_SPREAD_LOW 3.0f // spreading factor for hi-to-low threshold spreading (30 dB/Bark)
48 #define PSY_3GPP_EN_SPREAD_HI_L1 2.0f
50 #define PSY_3GPP_EN_SPREAD_HI_L2 1.5f
52 #define PSY_3GPP_EN_SPREAD_HI_S 1.5f
54 #define PSY_3GPP_EN_SPREAD_LOW_L 3.0f
56 #define PSY_3GPP_EN_SPREAD_LOW_S 2.0f
58 #define PSY_3GPP_RPEMIN 0.01f
59 #define PSY_3GPP_RPELEV 2.0f
61 #define PSY_3GPP_C1 3.0f
62 #define PSY_3GPP_C2 1.3219281f
63 #define PSY_3GPP_C3 0.55935729f
65 #define PSY_SNR_1DB 7.9432821e-1f
66 #define PSY_SNR_25DB 3.1622776e-3f
68 #define PSY_3GPP_SAVE_SLOPE_L -0.46666667f
69 #define PSY_3GPP_SAVE_SLOPE_S -0.36363637f
70 #define PSY_3GPP_SAVE_ADD_L -0.84285712f
71 #define PSY_3GPP_SAVE_ADD_S -0.75f
72 #define PSY_3GPP_SPEND_SLOPE_L 0.66666669f
73 #define PSY_3GPP_SPEND_SLOPE_S 0.81818181f
74 #define PSY_3GPP_SPEND_ADD_L -0.35f
75 #define PSY_3GPP_SPEND_ADD_S -0.26111111f
76 #define PSY_3GPP_CLIP_LO_L 0.2f
77 #define PSY_3GPP_CLIP_LO_S 0.2f
78 #define PSY_3GPP_CLIP_HI_L 0.95f
79 #define PSY_3GPP_CLIP_HI_S 0.75f
81 #define PSY_3GPP_AH_THR_LONG 0.5f
82 #define PSY_3GPP_AH_THR_SHORT 0.63f
84 #define PSY_PE_FORGET_SLOPE 511
92 #define PSY_3GPP_BITS_TO_PE(bits) ((bits) * 1.18f)
93 #define PSY_3GPP_PE_TO_BITS(bits) ((bits) / 1.18f)
96 #define PSY_LAME_FIR_LEN 21
97 #define AAC_BLOCK_SIZE_LONG 1024
98 #define AAC_BLOCK_SIZE_SHORT 128
99 #define AAC_NUM_BLOCKS_SHORT 8
100 #define PSY_LAME_NUM_SUBBLOCKS 2
222 -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2,
223 -3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2,
224 -5.52212e-17 * 2, -0.313819 * 2
233 int lower_range = 12, upper_range = 12;
241 for (
i = 1;
i < 13;
i++) {
282 return 13.3f *
atanf(0.00076
f *
f) + 3.5f *
atanf((
f / 7500.0
f) * (
f / 7500.0
f));
293 return 3.64 * pow(
f, -0.8)
294 - 6.8 *
exp(-0.6 * (
f - 3.4) * (
f - 3.4))
295 + 6.0 *
exp(-0.15 * (
f - 8.7) * (
f - 8.7))
296 + (0.6 + 0.04 * add) * 0.001 *
f *
f *
f *
f;
303 float prev, minscale, minath, minsnr, pe_min;
307 const float num_bark =
calc_bark((
float)bandwidth);
313 if (!
ctx->model_priv_data)
315 pctx =
ctx->model_priv_data;
320 chan_bitrate = (int)(chan_bitrate / 120.0 * (
ctx->avctx->global_quality ?
ctx->avctx->global_quality : 120));
328 ctx->bitres.size -=
ctx->bitres.size % 8;
331 for (j = 0; j < 2; j++) {
333 const uint8_t *band_sizes =
ctx->bands[j];
334 float line_to_frequency =
ctx->avctx->sample_rate / (j ? 256.f : 2048.0f);
335 float avg_chan_bits = chan_bitrate * (j ? 128.0f : 1024.0f) /
ctx->avctx->sample_rate;
344 for (
g = 0;
g <
ctx->num_bands[j];
g++) {
347 coeffs[
g].
barks = (bark + prev) / 2.0;
350 for (
g = 0;
g <
ctx->num_bands[j] - 1;
g++) {
352 float bark_width = coeffs[
g+1].
barks - coeffs->
barks;
355 coeff->spread_low[1] =
ff_exp10(-bark_width * en_spread_low);
357 pe_min = bark_pe * bark_width;
358 minsnr =
exp2(pe_min / band_sizes[
g]) - 1.5f;
362 for (
g = 0;
g <
ctx->num_bands[j];
g++) {
363 minscale =
ath(start * line_to_frequency,
ATH_ADD);
364 for (
i = 1;
i < band_sizes[
g];
i++)
366 coeffs[
g].
ath = minscale - minath;
367 start += band_sizes[
g];
399 0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36
407 const int16_t *audio,
413 int attack_ratio = br <= 16000 ? 18 : 10;
416 uint8_t grouping = 0;
422 int switch_to_eight = 0;
423 float sum = 0.0, sum2 = 0.0;
426 for (
i = 0;
i < 8;
i++) {
427 for (j = 0; j < 128; j++) {
434 for (
i = 0;
i < 8;
i++) {
435 if (
s[
i] > pch->win_energy * attack_ratio) {
441 pch->win_energy = pch->win_energy*7/8 + sum2/64;
443 wi.window_type[1] = prev_type;
451 grouping = pch->next_grouping;
467 pch->next_window_seq = next_type;
469 for (
i = 0;
i < 3;
i++)
470 wi.window_type[
i] = prev_type;
481 for (
i = 0;
i < 8;
i++) {
482 if (!((grouping >>
i) & 1))
484 wi.grouping[lastgrp]++;
501 float clipped_pe, bit_save, bit_spend, bit_factor, fill_level, forgetful_min_pe;
505 fill_level =
av_clipf((
float)
ctx->fill_level /
size, clip_low, clip_high);
507 bit_save = (fill_level + bitsave_add) * bitsave_slope;
508 assert(bit_save <= 0.3f && bit_save >= -0.05000001
f);
509 bit_spend = (fill_level + bitspend_add) * bitspend_slope;
510 assert(bit_spend <= 0.5f && bit_spend >= -0.1
f);
517 bit_factor = 1.0f - bit_save + ((bit_spend - bit_save) / (
ctx->pe.max -
ctx->pe.min)) * (clipped_pe -
ctx->pe.min);
525 ctx->pe.min =
FFMIN(pe, forgetful_min_pe);
531 ctx->frame_bits * bit_factor,
561 float thr_avg, reduction;
563 if(active_lines == 0.0)
566 thr_avg =
exp2f((
a - pe) / (4.0
f * active_lines));
567 reduction =
exp2f((
a - desired_pe) / (4.0
f * active_lines)) - thr_avg;
569 return FFMAX(reduction, 0.0
f);
575 float thr = band->
thr;
579 thr =
sqrtf(thr) + reduction;
598 const uint8_t *band_sizes,
const float *coefs,
const int cutoff)
601 int start = 0, wstart = 0;
604 for (
g = 0;
g < num_bands;
g++) {
607 float form_factor = 0.0f;
610 if (wstart < cutoff) {
611 for (
i = 0;
i < band_sizes[
g];
i++) {
612 band->
energy += coefs[start+
i] * coefs[start+
i];
620 start += band_sizes[
g];
621 wstart += band_sizes[
g];
639 hpfsmpl[
i] = (sum1 + sum2) * 32768.0
f;
652 float desired_bits, desired_pe, delta_pe, reduction=
NAN, spread_en[128] = {0};
653 float a = 0.0f, active_lines = 0.0f, norm_fac = 0.0f;
654 float pe = pctx->chan_bitrate > 32000 ? 0.0f :
FFMAX(50.0
f, 100.0
f - pctx->chan_bitrate * 100.0f / 32000.0f);
655 const int num_bands =
ctx->num_bands[wi->num_windows == 8];
656 const uint8_t *band_sizes =
ctx->bands[wi->num_windows == 8];
657 AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8];
660 const int cutoff = bandwidth * 2048 / wi->num_windows /
ctx->avctx->sample_rate;
663 calc_thr_3gpp(wi, num_bands, pch, band_sizes, coefs, cutoff);
666 for (
w = 0;
w < wi->num_windows*16;
w += 16) {
670 spread_en[0] =
bands[0].energy;
671 for (
g = 1;
g < num_bands;
g++) {
673 spread_en[
w+
g] =
FFMAX(
bands[
g].energy, spread_en[
w+
g-1] * coeffs[
g].spread_hi[1]);
675 for (
g = num_bands - 2;
g >= 0;
g--) {
677 spread_en[
w+
g] =
FFMAX(spread_en[
w+
g], spread_en[
w+
g+1] * coeffs[
g].spread_low[1]);
680 for (
g = 0;
g < num_bands;
g++) {
695 if (spread_en[
w+
g] * avoid_hole_thr > band->
energy || coeffs[
g].min_snr > 1.0f)
708 desired_pe = pe * (
ctx->avctx->global_quality ?
ctx->avctx->global_quality : 120) / (2 * 2.5
f * 120.0
f);
713 if (
ctx->bitres.bits > 0) {
718 pctx->pe.max =
FFMAX(pe, pctx->pe.max);
719 pctx->pe.min =
FFMIN(pe, pctx->pe.min);
728 if (
ctx->bitres.bits > 0)
733 ctx->bitres.alloc = desired_bits;
735 if (desired_pe < pe) {
737 for (
w = 0;
w < wi->num_windows*16;
w += 16) {
742 for (
g = 0;
g < num_bands;
g++) {
754 for (
i = 0;
i < 2;
i++) {
755 float pe_no_ah = 0.0f, desired_pe_no_ah;
756 active_lines =
a = 0.0f;
757 for (
w = 0;
w < wi->num_windows*16;
w += 16) {
758 for (
g = 0;
g < num_bands;
g++) {
762 pe_no_ah += band->
pe;
768 desired_pe_no_ah =
FFMAX(desired_pe - (pe - pe_no_ah), 0.0
f);
769 if (active_lines > 0.0
f)
773 for (
w = 0;
w < wi->num_windows*16;
w += 16) {
774 for (
g = 0;
g < num_bands;
g++) {
777 if (active_lines > 0.0
f)
780 if (band->
thr > 0.0f)
787 delta_pe = desired_pe - pe;
788 if (
fabs(delta_pe) > 0.05
f * desired_pe)
792 if (pe < 1.15
f * desired_pe) {
794 norm_fac = norm_fac ? 1.0f / norm_fac : 0;
795 for (
w = 0;
w < wi->num_windows*16;
w += 16) {
796 for (
g = 0;
g < num_bands;
g++) {
800 float delta_sfb_pe = band->
norm_fac * norm_fac * delta_pe;
801 float thr = band->
thr;
813 while (pe > desired_pe &&
g--) {
814 for (
w = 0;
w < wi->num_windows*16;
w+= 16) {
827 for (
w = 0;
w < wi->num_windows*16;
w += 16) {
828 for (
g = 0;
g < num_bands;
g++) {
839 memcpy(pch->prev_band, pch->band,
sizeof(pch->band));
848 for (ch = 0; ch < group->
num_ch; ch++)
875 ctx->next_window_seq = blocktype;
879 const float *la,
int channel,
int prev_type)
884 int uselongblock = 1;
891 const float *pf = hpfsmpl;
906 energy_short[0] += energy_subshort[
i];
912 for (; pf < pfe; pf++)
931 if (attack_intensity[
i] > pch->attack_threshold)
939 const float u = energy_short[
i - 1];
940 const float v = energy_short[
i];
941 const float m =
FFMAX(
u, v);
943 if (
u < 2.3
f * v && v < 2.3
f *
u) {
944 if (
i == 1 && attacks[0] < attacks[
i])
949 att_sum += attacks[
i];
952 if (pch->next_attack0_zero)
956 if (attacks[0] <= pch->prev_attack)
959 att_sum += attacks[0];
967 if (attacks[
i] && attacks[
i-1])
992 for (
i = 0;
i < 8;
i++) {
993 if (!((pch->next_grouping >>
i) & 1))
1005 for (
i = 0;
i < 9;
i++) {
1020 .
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 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 u(width, name, range_min, range_max)
#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.
#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 values to.
float pe_const
constant part of the PE calculation
3GPP TS26.403-inspired psychoacoustic model specific data
int next_attack0_zero
whether attack[0] of the next frame is zero
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)
struct AacPsyContext::@34 pe
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
IDirect3DDxgiInterfaceAccess _COM_Outptr_ void ** p
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