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 "libavutil/attributes.h" 28#include "libavutil/ffmath.h" 29 30#include "avcodec.h" 31#include "aactab.h" 32#include "psymodel.h" 33 34/*********************************** 35 * TODOs: 36 * try other bitrate controlling mechanism (maybe use ratecontrol.c?) 37 * control quality for quality-based output 38 **********************************/ 39 40/** 41 * constants for 3GPP AAC psychoacoustic model 42 * @{ 43 */ 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) 46/* spreading factor for low-to-hi energy spreading, long block, > 22kbps/channel (20dB/Bark) */ 47#define PSY_3GPP_EN_SPREAD_HI_L1 2.0f 48/* spreading factor for low-to-hi energy spreading, long block, <= 22kbps/channel (15dB/Bark) */ 49#define PSY_3GPP_EN_SPREAD_HI_L2 1.5f 50/* spreading factor for low-to-hi energy spreading, short block (15 dB/Bark) */ 51#define PSY_3GPP_EN_SPREAD_HI_S 1.5f 52/* spreading factor for hi-to-low energy spreading, long block (30dB/Bark) */ 53#define PSY_3GPP_EN_SPREAD_LOW_L 3.0f 54/* spreading factor for hi-to-low energy spreading, short block (20dB/Bark) */ 55#define PSY_3GPP_EN_SPREAD_LOW_S 2.0f 56 57#define PSY_3GPP_RPEMIN 0.01f 58#define PSY_3GPP_RPELEV 2.0f 59 60#define PSY_3GPP_C1 3.0f /* log2(8) */ 61#define PSY_3GPP_C2 1.3219281f /* log2(2.5) */ 62#define PSY_3GPP_C3 0.55935729f /* 1 - C2 / C1 */ 63 64#define PSY_SNR_1DB 7.9432821e-1f /* -1dB */ 65#define PSY_SNR_25DB 3.1622776e-3f /* -25dB */ 66 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 79 80#define PSY_3GPP_AH_THR_LONG 0.5f 81#define PSY_3GPP_AH_THR_SHORT 0.63f 82 83#define PSY_PE_FORGET_SLOPE 511 84 85enum { 86 PSY_3GPP_AH_NONE, 87 PSY_3GPP_AH_INACTIVE, 88 PSY_3GPP_AH_ACTIVE 89}; 90 91#define PSY_3GPP_BITS_TO_PE(bits) ((bits) * 1.18f) 92#define PSY_3GPP_PE_TO_BITS(bits) ((bits) / 1.18f) 93 94/* LAME psy model constants */ 95#define PSY_LAME_FIR_LEN 21 ///< LAME psy model FIR order 96#define AAC_BLOCK_SIZE_LONG 1024 ///< long block size 97#define AAC_BLOCK_SIZE_SHORT 128 ///< short block size 98#define AAC_NUM_BLOCKS_SHORT 8 ///< number of blocks in a short sequence 99#define PSY_LAME_NUM_SUBBLOCKS 3 ///< Number of sub-blocks in each short block 100 101/** 102 * @} 103 */ 104 105/** 106 * information for single band used by 3GPP TS26.403-inspired psychoacoustic model 107 */ 108typedef struct AacPsyBand{ 109 float energy; ///< band energy 110 float thr; ///< energy threshold 111 float thr_quiet; ///< threshold in quiet 112 float nz_lines; ///< number of non-zero spectral lines 113 float active_lines; ///< number of active spectral lines 114 float pe; ///< perceptual entropy 115 float pe_const; ///< constant part of the PE calculation 116 float norm_fac; ///< normalization factor for linearization 117 int avoid_holes; ///< hole avoidance flag 118}AacPsyBand; 119 120/** 121 * single/pair channel context for psychoacoustic model 122 */ 123typedef struct AacPsyChannel{ 124 AacPsyBand band[128]; ///< bands information 125 AacPsyBand prev_band[128]; ///< bands information from the previous frame 126 127 float win_energy; ///< sliding average of channel energy 128 float iir_state[2]; ///< hi-pass IIR filter state 129 uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence) 130 enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame 131 /* LAME psy model specific members */ 132 float attack_threshold; ///< attack threshold for this channel 133 float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS]; 134 int prev_attack; ///< attack value for the last short block in the previous sequence 135}AacPsyChannel; 136 137/** 138 * psychoacoustic model frame type-dependent coefficients 139 */ 140typedef struct AacPsyCoeffs{ 141 float ath; ///< absolute threshold of hearing per bands 142 float barks; ///< Bark value for each spectral band in long frame 143 float spread_low[2]; ///< spreading factor for low-to-high threshold spreading in long frame 144 float spread_hi [2]; ///< spreading factor for high-to-low threshold spreading in long frame 145 float min_snr; ///< minimal SNR 146}AacPsyCoeffs; 147 148/** 149 * 3GPP TS26.403-inspired psychoacoustic model specific data 150 */ 151typedef struct AacPsyContext{ 152 int chan_bitrate; ///< bitrate per channel 153 int frame_bits; ///< average bits per frame 154 int fill_level; ///< bit reservoir fill level 155 struct { 156 float min; ///< minimum allowed PE for bit factor calculation 157 float max; ///< maximum allowed PE for bit factor calculation 158 float previous; ///< allowed PE of the previous frame 159 float correction; ///< PE correction factor 160 } pe; 161 AacPsyCoeffs psy_coef[2][64]; 162 AacPsyChannel *ch; 163 float global_quality; ///< normalized global quality taken from avctx 164}AacPsyContext; 165 166/** 167 * LAME psy model preset struct 168 */ 169typedef struct PsyLamePreset { 170 int quality; ///< Quality to map the rest of the vaules to. 171 /* This is overloaded to be both kbps per channel in ABR mode, and 172 * requested quality in constant quality mode. 173 */ 174 float st_lrm; ///< short threshold for L, R, and M channels 175} PsyLamePreset; 176 177/** 178 * LAME psy model preset table for ABR 179 */ 180static const PsyLamePreset psy_abr_map[] = { 181/* TODO: Tuning. These were taken from LAME. */ 182/* kbps/ch st_lrm */ 183 { 8, 6.60}, 184 { 16, 6.60}, 185 { 24, 6.60}, 186 { 32, 6.60}, 187 { 40, 6.60}, 188 { 48, 6.60}, 189 { 56, 6.60}, 190 { 64, 6.40}, 191 { 80, 6.00}, 192 { 96, 5.60}, 193 {112, 5.20}, 194 {128, 5.20}, 195 {160, 5.20} 196}; 197 198/** 199* LAME psy model preset table for constant quality 200*/ 201static const PsyLamePreset psy_vbr_map[] = { 202/* vbr_q st_lrm */ 203 { 0, 4.20}, 204 { 1, 4.20}, 205 { 2, 4.20}, 206 { 3, 4.20}, 207 { 4, 4.20}, 208 { 5, 4.20}, 209 { 6, 4.20}, 210 { 7, 4.20}, 211 { 8, 4.20}, 212 { 9, 4.20}, 213 {10, 4.20} 214}; 215 216/** 217 * LAME psy model FIR coefficient table 218 */ 219static const float psy_fir_coeffs[] = { 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 * 2 223}; 224 225#if ARCH_MIPS 226# include "mips/aacpsy_mips.h" 227#endif /* ARCH_MIPS */ 228 229/** 230 * Calculate the ABR attack threshold from the above LAME psymodel table. 231 */ 232static float lame_calc_attack_threshold(int bitrate) 233{ 234 /* Assume max bitrate to start with */ 235 int lower_range = 12, upper_range = 12; 236 int lower_range_kbps = psy_abr_map[12].quality; 237 int upper_range_kbps = psy_abr_map[12].quality; 238 int i; 239 240 /* Determine which bitrates the value specified falls between. 241 * If the loop ends without breaking our above assumption of 320kbps was correct. 242 */ 243 for (i = 1; i < 13; i++) { 244 if (FFMAX(bitrate, psy_abr_map[i].quality) != bitrate) { 245 upper_range = i; 246 upper_range_kbps = psy_abr_map[i ].quality; 247 lower_range = i - 1; 248 lower_range_kbps = psy_abr_map[i - 1].quality; 249 break; /* Upper range found */ 250 } 251 } 252 253 /* Determine which range the value specified is closer to */ 254 if ((upper_range_kbps - bitrate) > (bitrate - lower_range_kbps)) 255 return psy_abr_map[lower_range].st_lrm; 256 return psy_abr_map[upper_range].st_lrm; 257} 258 259/** 260 * LAME psy model specific initialization 261 */ 262static av_cold void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx) 263{ 264 int i, j; 265 266 for (i = 0; i < avctx->ch_layout.nb_channels; i++) { 267 AacPsyChannel *pch = &ctx->ch[i]; 268 269 if (avctx->flags & AV_CODEC_FLAG_QSCALE) 270 pch->attack_threshold = psy_vbr_map[av_clip(avctx->global_quality / FF_QP2LAMBDA, 0, 10)].st_lrm; 271 else 272 pch->attack_threshold = lame_calc_attack_threshold(avctx->bit_rate / avctx->ch_layout.nb_channels / 1000); 273 274 for (j = 0; j < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; j++) 275 pch->prev_energy_subshort[j] = 10.0f; 276 } 277} 278 279/** 280 * Calculate Bark value for given line. 281 */ 282static av_cold float calc_bark(float f) 283{ 284 return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f)); 285} 286 287#define ATH_ADD 4 288/** 289 * Calculate ATH value for given frequency. 290 * Borrowed from Lame. 291 */ 292static av_cold float ath(float f, float add) 293{ 294 f /= 1000.0f; 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; 299} 300 301static av_cold int psy_3gpp_init(FFPsyContext *ctx) { 302 AacPsyContext *pctx; 303 float bark; 304 int i, j, g, start; 305 float prev, minscale, minath, minsnr, pe_min; 306 int chan_bitrate = ctx->avctx->bit_rate / ((ctx->avctx->flags & AV_CODEC_FLAG_QSCALE) ? 2.0f : ctx->avctx->ch_layout.nb_channels); 307 308 const int bandwidth = ctx->cutoff ? ctx->cutoff : AAC_CUTOFF(ctx->avctx); 309 const float num_bark = calc_bark((float)bandwidth); 310 311 if (bandwidth <= 0) 312 return AVERROR(EINVAL); 313 314 ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext)); 315 if (!ctx->model_priv_data) 316 return AVERROR(ENOMEM); 317 pctx = ctx->model_priv_data; 318 pctx->global_quality = (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120) * 0.01f; 319 320 if (ctx->avctx->flags & AV_CODEC_FLAG_QSCALE) { 321 /* Use the target average bitrate to compute spread parameters */ 322 chan_bitrate = (int)(chan_bitrate / 120.0 * (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120)); 323 } 324 325 pctx->chan_bitrate = chan_bitrate; 326 pctx->frame_bits = FFMIN(2560, chan_bitrate * AAC_BLOCK_SIZE_LONG / ctx->avctx->sample_rate); 327 pctx->pe.min = 8.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f); 328 pctx->pe.max = 12.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f); 329 ctx->bitres.size = 6144 - pctx->frame_bits; 330 ctx->bitres.size -= ctx->bitres.size % 8; 331 pctx->fill_level = ctx->bitres.size; 332 minath = ath(3410 - 0.733 * ATH_ADD, ATH_ADD); 333 for (j = 0; j < 2; j++) { 334 AacPsyCoeffs *coeffs = pctx->psy_coef[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; 338 /* reference encoder uses 2.4% here instead of 60% like the spec says */ 339 float bark_pe = 0.024f * PSY_3GPP_BITS_TO_PE(avg_chan_bits) / num_bark; 340 float en_spread_low = j ? PSY_3GPP_EN_SPREAD_LOW_S : PSY_3GPP_EN_SPREAD_LOW_L; 341 /* High energy spreading for long blocks <= 22kbps/channel and short blocks are the same. */ 342 float en_spread_hi = (j || (chan_bitrate <= 22.0f)) ? PSY_3GPP_EN_SPREAD_HI_S : PSY_3GPP_EN_SPREAD_HI_L1; 343 344 i = 0; 345 prev = 0.0; 346 for (g = 0; g < ctx->num_bands[j]; g++) { 347 i += band_sizes[g]; 348 bark = calc_bark((i-1) * line_to_frequency); 349 coeffs[g].barks = (bark + prev) / 2.0; 350 prev = bark; 351 } 352 for (g = 0; g < ctx->num_bands[j] - 1; g++) { 353 AacPsyCoeffs *coeff = &coeffs[g]; 354 float bark_width = coeffs[g+1].barks - coeffs->barks; 355 coeff->spread_low[0] = ff_exp10(-bark_width * PSY_3GPP_THR_SPREAD_LOW); 356 coeff->spread_hi [0] = ff_exp10(-bark_width * PSY_3GPP_THR_SPREAD_HI); 357 coeff->spread_low[1] = ff_exp10(-bark_width * en_spread_low); 358 coeff->spread_hi [1] = ff_exp10(-bark_width * en_spread_hi); 359 pe_min = bark_pe * bark_width; 360 minsnr = exp2(pe_min / band_sizes[g]) - 1.5f; 361 coeff->min_snr = av_clipf(1.0f / minsnr, PSY_SNR_25DB, PSY_SNR_1DB); 362 } 363 start = 0; 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++) 367 minscale = FFMIN(minscale, ath((start + i) * line_to_frequency, ATH_ADD)); 368 coeffs[g].ath = minscale - minath; 369 start += band_sizes[g]; 370 } 371 } 372 373 pctx->ch = av_calloc(ctx->avctx->ch_layout.nb_channels, sizeof(*pctx->ch)); 374 if (!pctx->ch) { 375 av_freep(&ctx->model_priv_data); 376 return AVERROR(ENOMEM); 377 } 378 379 lame_window_init(pctx, ctx->avctx); 380 381 return 0; 382} 383 384/** 385 * IIR filter used in block switching decision 386 */ 387static float iir_filter(int in, float state[2]) 388{ 389 float ret; 390 391 ret = 0.7548f * (in - state[0]) + 0.5095f * state[1]; 392 state[0] = in; 393 state[1] = ret; 394 return ret; 395} 396 397/** 398 * window grouping information stored as bits (0 - new group, 1 - group continues) 399 */ 400static const uint8_t window_grouping[9] = { 401 0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36 402}; 403 404/** 405 * Tell encoder which window types to use. 406 * @see 3GPP TS26.403 5.4.1 "Blockswitching" 407 */ 408static av_unused FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx, 409 const int16_t *audio, 410 const int16_t *la, 411 int channel, int prev_type) 412{ 413 int i, j; 414 int br = ((AacPsyContext*)ctx->model_priv_data)->chan_bitrate; 415 int attack_ratio = br <= 16000 ? 18 : 10; 416 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data; 417 AacPsyChannel *pch = &pctx->ch[channel]; 418 uint8_t grouping = 0; 419 int next_type = pch->next_window_seq; 420 FFPsyWindowInfo wi = { { 0 } }; 421 422 if (la) { 423 float s[8], v; 424 int switch_to_eight = 0; 425 float sum = 0.0, sum2 = 0.0; 426 int attack_n = 0; 427 int stay_short = 0; 428 for (i = 0; i < 8; i++) { 429 for (j = 0; j < 128; j++) { 430 v = iir_filter(la[i*128+j], pch->iir_state); 431 sum += v*v; 432 } 433 s[i] = sum; 434 sum2 += sum; 435 } 436 for (i = 0; i < 8; i++) { 437 if (s[i] > pch->win_energy * attack_ratio) { 438 attack_n = i + 1; 439 switch_to_eight = 1; 440 break; 441 } 442 } 443 pch->win_energy = pch->win_energy*7/8 + sum2/64; 444 445 wi.window_type[1] = prev_type; 446 switch (prev_type) { 447 case ONLY_LONG_SEQUENCE: 448 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE; 449 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE; 450 break; 451 case LONG_START_SEQUENCE: 452 wi.window_type[0] = EIGHT_SHORT_SEQUENCE; 453 grouping = pch->next_grouping; 454 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE; 455 break; 456 case LONG_STOP_SEQUENCE: 457 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE; 458 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE; 459 break; 460 case EIGHT_SHORT_SEQUENCE: 461 stay_short = next_type == EIGHT_SHORT_SEQUENCE || switch_to_eight; 462 wi.window_type[0] = stay_short ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE; 463 grouping = next_type == EIGHT_SHORT_SEQUENCE ? pch->next_grouping : 0; 464 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE; 465 break; 466 } 467 468 pch->next_grouping = window_grouping[attack_n]; 469 pch->next_window_seq = next_type; 470 } else { 471 for (i = 0; i < 3; i++) 472 wi.window_type[i] = prev_type; 473 grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0; 474 } 475 476 wi.window_shape = 1; 477 if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) { 478 wi.num_windows = 1; 479 wi.grouping[0] = 1; 480 } else { 481 int lastgrp = 0; 482 wi.num_windows = 8; 483 for (i = 0; i < 8; i++) { 484 if (!((grouping >> i) & 1)) 485 lastgrp = i; 486 wi.grouping[lastgrp]++; 487 } 488 } 489 490 return wi; 491} 492 493/* 5.6.1.2 "Calculation of Bit Demand" */ 494static int calc_bit_demand(AacPsyContext *ctx, float pe, int bits, int size, 495 int short_window) 496{ 497 const float bitsave_slope = short_window ? PSY_3GPP_SAVE_SLOPE_S : PSY_3GPP_SAVE_SLOPE_L; 498 const float bitsave_add = short_window ? PSY_3GPP_SAVE_ADD_S : PSY_3GPP_SAVE_ADD_L; 499 const float bitspend_slope = short_window ? PSY_3GPP_SPEND_SLOPE_S : PSY_3GPP_SPEND_SLOPE_L; 500 const float bitspend_add = short_window ? PSY_3GPP_SPEND_ADD_S : PSY_3GPP_SPEND_ADD_L; 501 const float clip_low = short_window ? PSY_3GPP_CLIP_LO_S : PSY_3GPP_CLIP_LO_L; 502 const float clip_high = short_window ? PSY_3GPP_CLIP_HI_S : PSY_3GPP_CLIP_HI_L; 503 float clipped_pe, bit_save, bit_spend, bit_factor, fill_level, forgetful_min_pe; 504 505 ctx->fill_level += ctx->frame_bits - bits; 506 ctx->fill_level = av_clip(ctx->fill_level, 0, size); 507 fill_level = av_clipf((float)ctx->fill_level / size, clip_low, clip_high); 508 clipped_pe = av_clipf(pe, ctx->pe.min, ctx->pe.max); 509 bit_save = (fill_level + bitsave_add) * bitsave_slope; 510 assert(bit_save <= 0.3f && bit_save >= -0.05000001f); 511 bit_spend = (fill_level + bitspend_add) * bitspend_slope; 512 assert(bit_spend <= 0.5f && bit_spend >= -0.1f); 513 /* The bit factor graph in the spec is obviously incorrect. 514 * bit_spend + ((bit_spend - bit_spend))... 515 * The reference encoder subtracts everything from 1, but also seems incorrect. 516 * 1 - bit_save + ((bit_spend + bit_save))... 517 * Hopefully below is correct. 518 */ 519 bit_factor = 1.0f - bit_save + ((bit_spend - bit_save) / (ctx->pe.max - ctx->pe.min)) * (clipped_pe - ctx->pe.min); 520 /* NOTE: The reference encoder attempts to center pe max/min around the current pe. 521 * Here we do that by slowly forgetting pe.min when pe stays in a range that makes 522 * it unlikely (ie: above the mean) 523 */ 524 ctx->pe.max = FFMAX(pe, ctx->pe.max); 525 forgetful_min_pe = ((ctx->pe.min * PSY_PE_FORGET_SLOPE) 526 + FFMAX(ctx->pe.min, pe * (pe / ctx->pe.max))) / (PSY_PE_FORGET_SLOPE + 1); 527 ctx->pe.min = FFMIN(pe, forgetful_min_pe); 528 529 /* NOTE: allocate a minimum of 1/8th average frame bits, to avoid 530 * reservoir starvation from producing zero-bit frames 531 */ 532 return FFMIN( 533 ctx->frame_bits * bit_factor, 534 FFMAX(ctx->frame_bits + size - bits, ctx->frame_bits / 8)); 535} 536 537static float calc_pe_3gpp(AacPsyBand *band) 538{ 539 float pe, a; 540 541 band->pe = 0.0f; 542 band->pe_const = 0.0f; 543 band->active_lines = 0.0f; 544 if (band->energy > band->thr) { 545 a = log2f(band->energy); 546 pe = a - log2f(band->thr); 547 band->active_lines = band->nz_lines; 548 if (pe < PSY_3GPP_C1) { 549 pe = pe * PSY_3GPP_C3 + PSY_3GPP_C2; 550 a = a * PSY_3GPP_C3 + PSY_3GPP_C2; 551 band->active_lines *= PSY_3GPP_C3; 552 } 553 band->pe = pe * band->nz_lines; 554 band->pe_const = a * band->nz_lines; 555 } 556 557 return band->pe; 558} 559 560static float calc_reduction_3gpp(float a, float desired_pe, float pe, 561 float active_lines) 562{ 563 float thr_avg, reduction; 564 565 if(active_lines == 0.0) 566 return 0; 567 568 thr_avg = exp2f((a - pe) / (4.0f * active_lines)); 569 reduction = exp2f((a - desired_pe) / (4.0f * active_lines)) - thr_avg; 570 571 return FFMAX(reduction, 0.0f); 572} 573 574static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr, 575 float reduction) 576{ 577 float thr = band->thr; 578 579 if (band->energy > thr) { 580 thr = sqrtf(thr); 581 thr = sqrtf(thr) + reduction; 582 thr *= thr; 583 thr *= thr; 584 585 /* This deviates from the 3GPP spec to match the reference encoder. 586 * It performs min(thr_reduced, max(thr, energy/min_snr)) only for bands 587 * that have hole avoidance on (active or inactive). It always reduces the 588 * threshold of bands with hole avoidance off. 589 */ 590 if (thr > band->energy * min_snr && band->avoid_holes != PSY_3GPP_AH_NONE) { 591 thr = FFMAX(band->thr, band->energy * min_snr); 592 band->avoid_holes = PSY_3GPP_AH_ACTIVE; 593 } 594 } 595 596 return thr; 597} 598 599#ifndef calc_thr_3gpp 600static void calc_thr_3gpp(const FFPsyWindowInfo *wi, const int num_bands, AacPsyChannel *pch, 601 const uint8_t *band_sizes, const float *coefs, const int cutoff) 602{ 603 int i, w, g; 604 int start = 0, wstart = 0; 605 for (w = 0; w < wi->num_windows*16; w += 16) { 606 wstart = 0; 607 for (g = 0; g < num_bands; g++) { 608 AacPsyBand *band = &pch->band[w+g]; 609 610 float form_factor = 0.0f; 611 float Temp; 612 band->energy = 0.0f; 613 if (wstart < cutoff) { 614 for (i = 0; i < band_sizes[g]; i++) { 615 band->energy += coefs[start+i] * coefs[start+i]; 616 form_factor += sqrtf(fabs(coefs[start+i])); 617 } 618 } 619 Temp = band->energy > 0 ? sqrtf((float)band_sizes[g] / band->energy) : 0; 620 band->thr = band->energy * 0.001258925f; 621 band->nz_lines = form_factor * sqrtf(Temp); 622 623 start += band_sizes[g]; 624 wstart += band_sizes[g]; 625 } 626 } 627} 628#endif /* calc_thr_3gpp */ 629 630#ifndef psy_hp_filter 631static void psy_hp_filter(const float *firbuf, float *hpfsmpl, const float *psy_fir_coeffs) 632{ 633 int i, j; 634 for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) { 635 float sum1, sum2; 636 sum1 = firbuf[i + (PSY_LAME_FIR_LEN - 1) / 2]; 637 sum2 = 0.0; 638 for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) { 639 sum1 += psy_fir_coeffs[j] * (firbuf[i + j] + firbuf[i + PSY_LAME_FIR_LEN - j]); 640 sum2 += psy_fir_coeffs[j + 1] * (firbuf[i + j + 1] + firbuf[i + PSY_LAME_FIR_LEN - j - 1]); 641 } 642 /* NOTE: The LAME psymodel expects it's input in the range -32768 to 32768. 643 * Tuning this for normalized floats would be difficult. */ 644 hpfsmpl[i] = (sum1 + sum2) * 32768.0f; 645 } 646} 647#endif /* psy_hp_filter */ 648 649/** 650 * Calculate band thresholds as suggested in 3GPP TS26.403 651 */ 652static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel, 653 const float *coefs, const FFPsyWindowInfo *wi) 654{ 655 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data; 656 AacPsyChannel *pch = &pctx->ch[channel]; 657 int i, w, g; 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(50.0f, 100.0f - 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]; 664 const float avoid_hole_thr = wi->num_windows == 8 ? PSY_3GPP_AH_THR_SHORT : PSY_3GPP_AH_THR_LONG; 665 const int bandwidth = ctx->cutoff ? ctx->cutoff : AAC_CUTOFF(ctx->avctx); 666 const int cutoff = bandwidth * 2048 / wi->num_windows / ctx->avctx->sample_rate; 667 668 //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation" 669 calc_thr_3gpp(wi, num_bands, pch, band_sizes, coefs, cutoff); 670 671 //modify thresholds and energies - spread, threshold in quiet, pre-echo control 672 for (w = 0; w < wi->num_windows*16; w += 16) { 673 AacPsyBand *bands = &pch->band[w]; 674 675 /* 5.4.2.3 "Spreading" & 5.4.3 "Spread Energy Calculation" */ 676 spread_en[0] = bands[0].energy; 677 for (g = 1; g < num_bands; g++) { 678 bands[g].thr = FFMAX(bands[g].thr, bands[g-1].thr * coeffs[g].spread_hi[0]); 679 spread_en[w+g] = FFMAX(bands[g].energy, spread_en[w+g-1] * coeffs[g].spread_hi[1]); 680 } 681 for (g = num_bands - 2; g >= 0; g--) { 682 bands[g].thr = FFMAX(bands[g].thr, bands[g+1].thr * coeffs[g].spread_low[0]); 683 spread_en[w+g] = FFMAX(spread_en[w+g], spread_en[w+g+1] * coeffs[g].spread_low[1]); 684 } 685 //5.4.2.4 "Threshold in quiet" 686 for (g = 0; g < num_bands; g++) { 687 AacPsyBand *band = &bands[g]; 688 689 band->thr_quiet = band->thr = FFMAX(band->thr, coeffs[g].ath); 690 //5.4.2.5 "Pre-echo control" 691 if (!(wi->window_type[0] == LONG_STOP_SEQUENCE || (!w && wi->window_type[1] == LONG_START_SEQUENCE))) 692 band->thr = FFMAX(PSY_3GPP_RPEMIN*band->thr, FFMIN(band->thr, 693 PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet)); 694 695 /* 5.6.1.3.1 "Preparatory steps of the perceptual entropy calculation" */ 696 pe += calc_pe_3gpp(band); 697 a += band->pe_const; 698 active_lines += band->active_lines; 699 700 /* 5.6.1.3.3 "Selection of the bands for avoidance of holes" */ 701 if (spread_en[w+g] * avoid_hole_thr > band->energy || coeffs[g].min_snr > 1.0f) 702 band->avoid_holes = PSY_3GPP_AH_NONE; 703 else 704 band->avoid_holes = PSY_3GPP_AH_INACTIVE; 705 } 706 } 707 708 /* 5.6.1.3.2 "Calculation of the desired perceptual entropy" */ 709 ctx->ch[channel].entropy = pe; 710 if (ctx->avctx->flags & AV_CODEC_FLAG_QSCALE) { 711 /* (2.5 * 120) achieves almost transparent rate, and we want to give 712 * ample room downwards, so we make that equivalent to QSCALE=2.4 713 */ 714 desired_pe = pe * (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120) / (2 * 2.5f * 120.0f); 715 desired_bits = FFMIN(2560, PSY_3GPP_PE_TO_BITS(desired_pe)); 716 desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); // reflect clipping 717 718 /* PE slope smoothing */ 719 if (ctx->bitres.bits > 0) { 720 desired_bits = FFMIN(2560, PSY_3GPP_PE_TO_BITS(desired_pe)); 721 desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); // reflect clipping 722 } 723 724 pctx->pe.max = FFMAX(pe, pctx->pe.max); 725 pctx->pe.min = FFMIN(pe, pctx->pe.min); 726 } else { 727 desired_bits = calc_bit_demand(pctx, pe, ctx->bitres.bits, ctx->bitres.size, wi->num_windows == 8); 728 desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); 729 730 /* NOTE: PE correction is kept simple. During initial testing it had very 731 * little effect on the final bitrate. Probably a good idea to come 732 * back and do more testing later. 733 */ 734 if (ctx->bitres.bits > 0) 735 desired_pe *= av_clipf(pctx->pe.previous / PSY_3GPP_BITS_TO_PE(ctx->bitres.bits), 736 0.85f, 1.15f); 737 } 738 pctx->pe.previous = PSY_3GPP_BITS_TO_PE(desired_bits); 739 ctx->bitres.alloc = desired_bits; 740 741 if (desired_pe < pe) { 742 /* 5.6.1.3.4 "First Estimation of the reduction value" */ 743 for (w = 0; w < wi->num_windows*16; w += 16) { 744 reduction = calc_reduction_3gpp(a, desired_pe, pe, active_lines); 745 pe = 0.0f; 746 a = 0.0f; 747 active_lines = 0.0f; 748 for (g = 0; g < num_bands; g++) { 749 AacPsyBand *band = &pch->band[w+g]; 750 751 band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction); 752 /* recalculate PE */ 753 pe += calc_pe_3gpp(band); 754 a += band->pe_const; 755 active_lines += band->active_lines; 756 } 757 } 758 759 /* 5.6.1.3.5 "Second Estimation of the reduction value" */ 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++) { 765 AacPsyBand *band = &pch->band[w+g]; 766 767 if (band->avoid_holes != PSY_3GPP_AH_ACTIVE) { 768 pe_no_ah += band->pe; 769 a += band->pe_const; 770 active_lines += band->active_lines; 771 } 772 } 773 } 774 desired_pe_no_ah = FFMAX(desired_pe - (pe - pe_no_ah), 0.0f); 775 if (active_lines > 0.0f) 776 reduction = calc_reduction_3gpp(a, desired_pe_no_ah, pe_no_ah, active_lines); 777 778 pe = 0.0f; 779 for (w = 0; w < wi->num_windows*16; w += 16) { 780 for (g = 0; g < num_bands; g++) { 781 AacPsyBand *band = &pch->band[w+g]; 782 783 if (active_lines > 0.0f) 784 band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction); 785 pe += calc_pe_3gpp(band); 786 if (band->thr > 0.0f) 787 band->norm_fac = band->active_lines / band->thr; 788 else 789 band->norm_fac = 0.0f; 790 norm_fac += band->norm_fac; 791 } 792 } 793 delta_pe = desired_pe - pe; 794 if (fabs(delta_pe) > 0.05f * desired_pe) 795 break; 796 } 797 798 if (pe < 1.15f * desired_pe) { 799 /* 6.6.1.3.6 "Final threshold modification by linearization" */ 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++) { 803 AacPsyBand *band = &pch->band[w+g]; 804 805 if (band->active_lines > 0.5f) { 806 float delta_sfb_pe = band->norm_fac * norm_fac * delta_pe; 807 float thr = band->thr; 808 809 thr *= exp2f(delta_sfb_pe / band->active_lines); 810 if (thr > coeffs[g].min_snr * band->energy && band->avoid_holes == PSY_3GPP_AH_INACTIVE) 811 thr = FFMAX(band->thr, coeffs[g].min_snr * band->energy); 812 band->thr = thr; 813 } 814 } 815 } 816 } else { 817 /* 5.6.1.3.7 "Further perceptual entropy reduction" */ 818 g = num_bands; 819 while (pe > desired_pe && g--) { 820 for (w = 0; w < wi->num_windows*16; w+= 16) { 821 AacPsyBand *band = &pch->band[w+g]; 822 if (band->avoid_holes != PSY_3GPP_AH_NONE && coeffs[g].min_snr < PSY_SNR_1DB) { 823 coeffs[g].min_snr = PSY_SNR_1DB; 824 band->thr = band->energy * PSY_SNR_1DB; 825 pe += band->active_lines * 1.5f - band->pe; 826 } 827 } 828 } 829 /* TODO: allow more holes (unused without mid/side) */ 830 } 831 } 832 833 for (w = 0; w < wi->num_windows*16; w += 16) { 834 for (g = 0; g < num_bands; g++) { 835 AacPsyBand *band = &pch->band[w+g]; 836 FFPsyBand *psy_band = &ctx->ch[channel].psy_bands[w+g]; 837 838 psy_band->threshold = band->thr; 839 psy_band->energy = band->energy; 840 psy_band->spread = band->active_lines * 2.0f / band_sizes[g]; 841 psy_band->bits = PSY_3GPP_PE_TO_BITS(band->pe); 842 } 843 } 844 845 memcpy(pch->prev_band, pch->band, sizeof(pch->band)); 846} 847 848static void psy_3gpp_analyze(FFPsyContext *ctx, int channel, 849 const float **coeffs, const FFPsyWindowInfo *wi) 850{ 851 int ch; 852 FFPsyChannelGroup *group = ff_psy_find_group(ctx, channel); 853 854 for (ch = 0; ch < group->num_ch; ch++) 855 psy_3gpp_analyze_channel(ctx, channel + ch, coeffs[ch], &wi[ch]); 856} 857 858static av_cold void psy_3gpp_end(FFPsyContext *apc) 859{ 860 AacPsyContext *pctx = (AacPsyContext*) apc->model_priv_data; 861 if (pctx) 862 av_freep(&pctx->ch); 863 av_freep(&apc->model_priv_data); 864} 865 866static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock) 867{ 868 int blocktype = ONLY_LONG_SEQUENCE; 869 if (uselongblock) { 870 if (ctx->next_window_seq == EIGHT_SHORT_SEQUENCE) 871 blocktype = LONG_STOP_SEQUENCE; 872 } else { 873 blocktype = EIGHT_SHORT_SEQUENCE; 874 if (ctx->next_window_seq == ONLY_LONG_SEQUENCE) 875 ctx->next_window_seq = LONG_START_SEQUENCE; 876 if (ctx->next_window_seq == LONG_STOP_SEQUENCE) 877 ctx->next_window_seq = EIGHT_SHORT_SEQUENCE; 878 } 879 880 wi->window_type[0] = ctx->next_window_seq; 881 ctx->next_window_seq = blocktype; 882} 883 884static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx, const float *audio, 885 const float *la, int channel, int prev_type) 886{ 887 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data; 888 AacPsyChannel *pch = &pctx->ch[channel]; 889 int grouping = 0; 890 int uselongblock = 1; 891 int attacks[AAC_NUM_BLOCKS_SHORT + 1] = { 0 }; 892 int i; 893 FFPsyWindowInfo wi = { { 0 } }; 894 895 if (la) { 896 float hpfsmpl[AAC_BLOCK_SIZE_LONG]; 897 const float *pf = hpfsmpl; 898 float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS]; 899 float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS]; 900 float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 }; 901 const float *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN); 902 int att_sum = 0; 903 904 /* LAME comment: apply high pass filter of fs/4 */ 905 psy_hp_filter(firbuf, hpfsmpl, psy_fir_coeffs); 906 907 /* Calculate the energies of each sub-shortblock */ 908 for (i = 0; i < PSY_LAME_NUM_SUBBLOCKS; i++) { 909 energy_subshort[i] = pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS)]; 910 assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)] > 0); 911 attack_intensity[i] = energy_subshort[i] / pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)]; 912 energy_short[0] += energy_subshort[i]; 913 } 914 915 for (i = 0; i < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; i++) { 916 const float *const pfe = pf + AAC_BLOCK_SIZE_LONG / (AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS); 917 float p = 1.0f; 918 for (; pf < pfe; pf++) 919 p = FFMAX(p, fabsf(*pf)); 920 pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p; 921 energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p; 922 /* NOTE: The indexes below are [i + 3 - 2] in the LAME source. 923 * Obviously the 3 and 2 have some significance, or this would be just [i + 1] 924 * (which is what we use here). What the 3 stands for is ambiguous, as it is both 925 * number of short blocks, and the number of sub-short blocks. 926 * It seems that LAME is comparing each sub-block to sub-block + 1 in the 927 * previous block. 928 */ 929 if (p > energy_subshort[i + 1]) 930 p = p / energy_subshort[i + 1]; 931 else if (energy_subshort[i + 1] > p * 10.0f) 932 p = energy_subshort[i + 1] / (p * 10.0f); 933 else 934 p = 0.0; 935 attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p; 936 } 937 938 /* compare energy between sub-short blocks */ 939 for (i = 0; i < (AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS; i++) 940 if (!attacks[i / PSY_LAME_NUM_SUBBLOCKS]) 941 if (attack_intensity[i] > pch->attack_threshold) 942 attacks[i / PSY_LAME_NUM_SUBBLOCKS] = (i % PSY_LAME_NUM_SUBBLOCKS) + 1; 943 944 /* should have energy change between short blocks, in order to avoid periodic signals */ 945 /* Good samples to show the effect are Trumpet test songs */ 946 /* GB: tuned (1) to avoid too many short blocks for test sample TRUMPET */ 947 /* RH: tuned (2) to let enough short blocks through for test sample FSOL and SNAPS */ 948 for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) { 949 const float u = energy_short[i - 1]; 950 const float v = energy_short[i]; 951 const float m = FFMAX(u, v); 952 if (m < 40000) { /* (2) */ 953 if (u < 1.7f * v && v < 1.7f * u) { /* (1) */ 954 if (i == 1 && attacks[0] < attacks[i]) 955 attacks[0] = 0; 956 attacks[i] = 0; 957 } 958 } 959 att_sum += attacks[i]; 960 } 961 962 if (attacks[0] <= pch->prev_attack) 963 attacks[0] = 0; 964 965 att_sum += attacks[0]; 966 /* 3 below indicates the previous attack happened in the last sub-block of the previous sequence */ 967 if (pch->prev_attack == 3 || att_sum) { 968 uselongblock = 0; 969 970 for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) 971 if (attacks[i] && attacks[i-1]) 972 attacks[i] = 0; 973 } 974 } else { 975 /* We have no lookahead info, so just use same type as the previous sequence. */ 976 uselongblock = !(prev_type == EIGHT_SHORT_SEQUENCE); 977 } 978 979 lame_apply_block_type(pch, &wi, uselongblock); 980 981 wi.window_type[1] = prev_type; 982 if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) { 983 984 wi.num_windows = 1; 985 wi.grouping[0] = 1; 986 if (wi.window_type[0] == LONG_START_SEQUENCE) 987 wi.window_shape = 0; 988 else 989 wi.window_shape = 1; 990 991 } else { 992 int lastgrp = 0; 993 994 wi.num_windows = 8; 995 wi.window_shape = 0; 996 for (i = 0; i < 8; i++) { 997 if (!((pch->next_grouping >> i) & 1)) 998 lastgrp = i; 999 wi.grouping[lastgrp]++; 1000 } 1001 } 1002 1003 /* Determine grouping, based on the location of the first attack, and save for 1004 * the next frame. 1005 * FIXME: Move this to analysis. 1006 * TODO: Tune groupings depending on attack location 1007 * TODO: Handle more than one attack in a group 1008 */ 1009 for (i = 0; i < 9; i++) { 1010 if (attacks[i]) { 1011 grouping = i; 1012 break; 1013 } 1014 } 1015 pch->next_grouping = window_grouping[grouping]; 1016 1017 pch->prev_attack = attacks[8]; 1018 1019 return wi; 1020} 1021 1022const FFPsyModel ff_aac_psy_model = 1023{ 1024 .name = "3GPP TS 26.403-inspired model", 1025 .init = psy_3gpp_init, 1026 .window = psy_lame_window, 1027 .analyze = psy_3gpp_analyze, 1028 .end = psy_3gpp_end, 1029}; 1030