1/* 2 * AAC encoder psychoacoustic model 3 * Copyright (C) 2008 Konstantin Shishkov 4 * 5 * This file is part of FFmpeg. 6 * 7 * FFmpeg is free software; you can redistribute it and/or 8 * modify it under the terms of the GNU Lesser General Public 9 * License as published by the Free Software Foundation; either 10 * version 2.1 of the License, or (at your option) any later version. 11 * 12 * FFmpeg is distributed in the hope that it will be useful, 13 * but WITHOUT ANY WARRANTY; without even the implied warranty of 14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 15 * Lesser General Public License for more details. 16 * 17 * You should have received a copy of the GNU Lesser General Public 18 * License along with FFmpeg; if not, write to the Free Software 19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA 20 */ 21 22/** 23 * @file 24 * AAC encoder psychoacoustic model 25 */ 26 27#include "avcodec.h" 28#include "aactab.h" 29#include "psymodel.h" 30 31/*********************************** 32 * TODOs: 33 * thresholds linearization after their modifications for attaining given bitrate 34 * try other bitrate controlling mechanism (maybe use ratecontrol.c?) 35 * control quality for quality-based output 36 **********************************/ 37 38/** 39 * constants for 3GPP AAC psychoacoustic model 40 * @{ 41 */ 42#define PSY_3GPP_SPREAD_LOW 1.5f // spreading factor for ascending threshold spreading (15 dB/Bark) 43#define PSY_3GPP_SPREAD_HI 3.0f // spreading factor for descending threshold spreading (30 dB/Bark) 44 45#define PSY_3GPP_RPEMIN 0.01f 46#define PSY_3GPP_RPELEV 2.0f 47/** 48 * @} 49 */ 50 51/** 52 * information for single band used by 3GPP TS26.403-inspired psychoacoustic model 53 */ 54typedef struct Psy3gppBand{ 55 float energy; ///< band energy 56 float ffac; ///< form factor 57 float thr; ///< energy threshold 58 float min_snr; ///< minimal SNR 59 float thr_quiet; ///< threshold in quiet 60}Psy3gppBand; 61 62/** 63 * single/pair channel context for psychoacoustic model 64 */ 65typedef struct Psy3gppChannel{ 66 Psy3gppBand band[128]; ///< bands information 67 Psy3gppBand prev_band[128]; ///< bands information from the previous frame 68 69 float win_energy; ///< sliding average of channel energy 70 float iir_state[2]; ///< hi-pass IIR filter state 71 uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence) 72 enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame 73}Psy3gppChannel; 74 75/** 76 * psychoacoustic model frame type-dependent coefficients 77 */ 78typedef struct Psy3gppCoeffs{ 79 float ath [64]; ///< absolute threshold of hearing per bands 80 float barks [64]; ///< Bark value for each spectral band in long frame 81 float spread_low[64]; ///< spreading factor for low-to-high threshold spreading in long frame 82 float spread_hi [64]; ///< spreading factor for high-to-low threshold spreading in long frame 83}Psy3gppCoeffs; 84 85/** 86 * 3GPP TS26.403-inspired psychoacoustic model specific data 87 */ 88typedef struct Psy3gppContext{ 89 Psy3gppCoeffs psy_coef[2]; 90 Psy3gppChannel *ch; 91}Psy3gppContext; 92 93/** 94 * Calculate Bark value for given line. 95 */ 96static av_cold float calc_bark(float f) 97{ 98 return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f)); 99} 100 101#define ATH_ADD 4 102/** 103 * Calculate ATH value for given frequency. 104 * Borrowed from Lame. 105 */ 106static av_cold float ath(float f, float add) 107{ 108 f /= 1000.0f; 109 return 3.64 * pow(f, -0.8) 110 - 6.8 * exp(-0.6 * (f - 3.4) * (f - 3.4)) 111 + 6.0 * exp(-0.15 * (f - 8.7) * (f - 8.7)) 112 + (0.6 + 0.04 * add) * 0.001 * f * f * f * f; 113} 114 115static av_cold int psy_3gpp_init(FFPsyContext *ctx) { 116 Psy3gppContext *pctx; 117 float barks[1024]; 118 int i, j, g, start; 119 float prev, minscale, minath; 120 121 ctx->model_priv_data = av_mallocz(sizeof(Psy3gppContext)); 122 pctx = (Psy3gppContext*) ctx->model_priv_data; 123 124 for (i = 0; i < 1024; i++) 125 barks[i] = calc_bark(i * ctx->avctx->sample_rate / 2048.0); 126 minath = ath(3410, ATH_ADD); 127 for (j = 0; j < 2; j++) { 128 Psy3gppCoeffs *coeffs = &pctx->psy_coef[j]; 129 i = 0; 130 prev = 0.0; 131 for (g = 0; g < ctx->num_bands[j]; g++) { 132 i += ctx->bands[j][g]; 133 coeffs->barks[g] = (barks[i - 1] + prev) / 2.0; 134 prev = barks[i - 1]; 135 } 136 for (g = 0; g < ctx->num_bands[j] - 1; g++) { 137 coeffs->spread_low[g] = pow(10.0, -(coeffs->barks[g+1] - coeffs->barks[g]) * PSY_3GPP_SPREAD_LOW); 138 coeffs->spread_hi [g] = pow(10.0, -(coeffs->barks[g+1] - coeffs->barks[g]) * PSY_3GPP_SPREAD_HI); 139 } 140 start = 0; 141 for (g = 0; g < ctx->num_bands[j]; g++) { 142 minscale = ath(ctx->avctx->sample_rate * start / 1024.0, ATH_ADD); 143 for (i = 1; i < ctx->bands[j][g]; i++) 144 minscale = FFMIN(minscale, ath(ctx->avctx->sample_rate * (start + i) / 1024.0 / 2.0, ATH_ADD)); 145 coeffs->ath[g] = minscale - minath; 146 start += ctx->bands[j][g]; 147 } 148 } 149 150 pctx->ch = av_mallocz(sizeof(Psy3gppChannel) * ctx->avctx->channels); 151 return 0; 152} 153 154/** 155 * IIR filter used in block switching decision 156 */ 157static float iir_filter(int in, float state[2]) 158{ 159 float ret; 160 161 ret = 0.7548f * (in - state[0]) + 0.5095f * state[1]; 162 state[0] = in; 163 state[1] = ret; 164 return ret; 165} 166 167/** 168 * window grouping information stored as bits (0 - new group, 1 - group continues) 169 */ 170static const uint8_t window_grouping[9] = { 171 0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36 172}; 173 174/** 175 * Tell encoder which window types to use. 176 * @see 3GPP TS26.403 5.4.1 "Blockswitching" 177 */ 178static FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx, 179 const int16_t *audio, const int16_t *la, 180 int channel, int prev_type) 181{ 182 int i, j; 183 int br = ctx->avctx->bit_rate / ctx->avctx->channels; 184 int attack_ratio = br <= 16000 ? 18 : 10; 185 Psy3gppContext *pctx = (Psy3gppContext*) ctx->model_priv_data; 186 Psy3gppChannel *pch = &pctx->ch[channel]; 187 uint8_t grouping = 0; 188 FFPsyWindowInfo wi; 189 190 memset(&wi, 0, sizeof(wi)); 191 if (la) { 192 float s[8], v; 193 int switch_to_eight = 0; 194 float sum = 0.0, sum2 = 0.0; 195 int attack_n = 0; 196 for (i = 0; i < 8; i++) { 197 for (j = 0; j < 128; j++) { 198 v = iir_filter(audio[(i*128+j)*ctx->avctx->channels], pch->iir_state); 199 sum += v*v; 200 } 201 s[i] = sum; 202 sum2 += sum; 203 } 204 for (i = 0; i < 8; i++) { 205 if (s[i] > pch->win_energy * attack_ratio) { 206 attack_n = i + 1; 207 switch_to_eight = 1; 208 break; 209 } 210 } 211 pch->win_energy = pch->win_energy*7/8 + sum2/64; 212 213 wi.window_type[1] = prev_type; 214 switch (prev_type) { 215 case ONLY_LONG_SEQUENCE: 216 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE; 217 break; 218 case LONG_START_SEQUENCE: 219 wi.window_type[0] = EIGHT_SHORT_SEQUENCE; 220 grouping = pch->next_grouping; 221 break; 222 case LONG_STOP_SEQUENCE: 223 wi.window_type[0] = ONLY_LONG_SEQUENCE; 224 break; 225 case EIGHT_SHORT_SEQUENCE: 226 wi.window_type[0] = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE; 227 grouping = switch_to_eight ? pch->next_grouping : 0; 228 break; 229 } 230 pch->next_grouping = window_grouping[attack_n]; 231 } else { 232 for (i = 0; i < 3; i++) 233 wi.window_type[i] = prev_type; 234 grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0; 235 } 236 237 wi.window_shape = 1; 238 if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) { 239 wi.num_windows = 1; 240 wi.grouping[0] = 1; 241 } else { 242 int lastgrp = 0; 243 wi.num_windows = 8; 244 for (i = 0; i < 8; i++) { 245 if (!((grouping >> i) & 1)) 246 lastgrp = i; 247 wi.grouping[lastgrp]++; 248 } 249 } 250 251 return wi; 252} 253 254/** 255 * Calculate band thresholds as suggested in 3GPP TS26.403 256 */ 257static void psy_3gpp_analyze(FFPsyContext *ctx, int channel, 258 const float *coefs, FFPsyWindowInfo *wi) 259{ 260 Psy3gppContext *pctx = (Psy3gppContext*) ctx->model_priv_data; 261 Psy3gppChannel *pch = &pctx->ch[channel]; 262 int start = 0; 263 int i, w, g; 264 const int num_bands = ctx->num_bands[wi->num_windows == 8]; 265 const uint8_t* band_sizes = ctx->bands[wi->num_windows == 8]; 266 Psy3gppCoeffs *coeffs = &pctx->psy_coef[wi->num_windows == 8]; 267 268 //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation" 269 for (w = 0; w < wi->num_windows*16; w += 16) { 270 for (g = 0; g < num_bands; g++) { 271 Psy3gppBand *band = &pch->band[w+g]; 272 band->energy = 0.0f; 273 for (i = 0; i < band_sizes[g]; i++) 274 band->energy += coefs[start+i] * coefs[start+i]; 275 band->energy *= 1.0f / (512*512); 276 band->thr = band->energy * 0.001258925f; 277 start += band_sizes[g]; 278 279 ctx->psy_bands[channel*PSY_MAX_BANDS+w+g].energy = band->energy; 280 } 281 } 282 //modify thresholds - spread, threshold in quiet - 5.4.3 "Spreaded Energy Calculation" 283 for (w = 0; w < wi->num_windows*16; w += 16) { 284 Psy3gppBand *band = &pch->band[w]; 285 for (g = 1; g < num_bands; g++) 286 band[g].thr = FFMAX(band[g].thr, band[g-1].thr * coeffs->spread_low[g-1]); 287 for (g = num_bands - 2; g >= 0; g--) 288 band[g].thr = FFMAX(band[g].thr, band[g+1].thr * coeffs->spread_hi [g]); 289 for (g = 0; g < num_bands; g++) { 290 band[g].thr_quiet = FFMAX(band[g].thr, coeffs->ath[g]); 291 if (wi->num_windows != 8 && wi->window_type[1] != EIGHT_SHORT_SEQUENCE) 292 band[g].thr_quiet = FFMAX(PSY_3GPP_RPEMIN*band[g].thr_quiet, 293 FFMIN(band[g].thr_quiet, 294 PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet)); 295 band[g].thr = FFMAX(band[g].thr, band[g].thr_quiet * 0.25); 296 297 ctx->psy_bands[channel*PSY_MAX_BANDS+w+g].threshold = band[g].thr; 298 } 299 } 300 memcpy(pch->prev_band, pch->band, sizeof(pch->band)); 301} 302 303static av_cold void psy_3gpp_end(FFPsyContext *apc) 304{ 305 Psy3gppContext *pctx = (Psy3gppContext*) apc->model_priv_data; 306 av_freep(&pctx->ch); 307 av_freep(&apc->model_priv_data); 308} 309 310 311const FFPsyModel ff_aac_psy_model = 312{ 313 .name = "3GPP TS 26.403-inspired model", 314 .init = psy_3gpp_init, 315 .window = psy_3gpp_window, 316 .analyze = psy_3gpp_analyze, 317 .end = psy_3gpp_end, 318}; 319