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/libm.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 83enum { 84 PSY_3GPP_AH_NONE, 85 PSY_3GPP_AH_INACTIVE, 86 PSY_3GPP_AH_ACTIVE 87}; 88 89#define PSY_3GPP_BITS_TO_PE(bits) ((bits) * 1.18f) 90 91/* LAME psy model constants */ 92#define PSY_LAME_FIR_LEN 21 ///< LAME psy model FIR order 93#define AAC_BLOCK_SIZE_LONG 1024 ///< long block size 94#define AAC_BLOCK_SIZE_SHORT 128 ///< short block size 95#define AAC_NUM_BLOCKS_SHORT 8 ///< number of blocks in a short sequence 96#define PSY_LAME_NUM_SUBBLOCKS 3 ///< Number of sub-blocks in each short block 97 98/** 99 * @} 100 */ 101 102/** 103 * information for single band used by 3GPP TS26.403-inspired psychoacoustic model 104 */ 105typedef struct AacPsyBand{ 106 float energy; ///< band energy 107 float thr; ///< energy threshold 108 float thr_quiet; ///< threshold in quiet 109 float nz_lines; ///< number of non-zero spectral lines 110 float active_lines; ///< number of active spectral lines 111 float pe; ///< perceptual entropy 112 float pe_const; ///< constant part of the PE calculation 113 float norm_fac; ///< normalization factor for linearization 114 int avoid_holes; ///< hole avoidance flag 115}AacPsyBand; 116 117/** 118 * single/pair channel context for psychoacoustic model 119 */ 120typedef struct AacPsyChannel{ 121 AacPsyBand band[128]; ///< bands information 122 AacPsyBand prev_band[128]; ///< bands information from the previous frame 123 124 float win_energy; ///< sliding average of channel energy 125 float iir_state[2]; ///< hi-pass IIR filter state 126 uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence) 127 enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame 128 /* LAME psy model specific members */ 129 float attack_threshold; ///< attack threshold for this channel 130 float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS]; 131 int prev_attack; ///< attack value for the last short block in the previous sequence 132}AacPsyChannel; 133 134/** 135 * psychoacoustic model frame type-dependent coefficients 136 */ 137typedef struct AacPsyCoeffs{ 138 float ath; ///< absolute threshold of hearing per bands 139 float barks; ///< Bark value for each spectral band in long frame 140 float spread_low[2]; ///< spreading factor for low-to-high threshold spreading in long frame 141 float spread_hi [2]; ///< spreading factor for high-to-low threshold spreading in long frame 142 float min_snr; ///< minimal SNR 143}AacPsyCoeffs; 144 145/** 146 * 3GPP TS26.403-inspired psychoacoustic model specific data 147 */ 148typedef struct AacPsyContext{ 149 int chan_bitrate; ///< bitrate per channel 150 int frame_bits; ///< average bits per frame 151 int fill_level; ///< bit reservoir fill level 152 struct { 153 float min; ///< minimum allowed PE for bit factor calculation 154 float max; ///< maximum allowed PE for bit factor calculation 155 float previous; ///< allowed PE of the previous frame 156 float correction; ///< PE correction factor 157 } pe; 158 AacPsyCoeffs psy_coef[2][64]; 159 AacPsyChannel *ch; 160}AacPsyContext; 161 162/** 163 * LAME psy model preset struct 164 */ 165typedef struct { 166 int quality; ///< Quality to map the rest of the vaules to. 167 /* This is overloaded to be both kbps per channel in ABR mode, and 168 * requested quality in constant quality mode. 169 */ 170 float st_lrm; ///< short threshold for L, R, and M channels 171} PsyLamePreset; 172 173/** 174 * LAME psy model preset table for ABR 175 */ 176static const PsyLamePreset psy_abr_map[] = { 177/* TODO: Tuning. These were taken from LAME. */ 178/* kbps/ch st_lrm */ 179 { 8, 6.60}, 180 { 16, 6.60}, 181 { 24, 6.60}, 182 { 32, 6.60}, 183 { 40, 6.60}, 184 { 48, 6.60}, 185 { 56, 6.60}, 186 { 64, 6.40}, 187 { 80, 6.00}, 188 { 96, 5.60}, 189 {112, 5.20}, 190 {128, 5.20}, 191 {160, 5.20} 192}; 193 194/** 195* LAME psy model preset table for constant quality 196*/ 197static const PsyLamePreset psy_vbr_map[] = { 198/* vbr_q st_lrm */ 199 { 0, 4.20}, 200 { 1, 4.20}, 201 { 2, 4.20}, 202 { 3, 4.20}, 203 { 4, 4.20}, 204 { 5, 4.20}, 205 { 6, 4.20}, 206 { 7, 4.20}, 207 { 8, 4.20}, 208 { 9, 4.20}, 209 {10, 4.20} 210}; 211 212/** 213 * LAME psy model FIR coefficient table 214 */ 215static const float psy_fir_coeffs[] = { 216 -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2, 217 -3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2, 218 -5.52212e-17 * 2, -0.313819 * 2 219}; 220 221#if ARCH_MIPS 222# include "mips/aacpsy_mips.h" 223#endif /* ARCH_MIPS */ 224 225/** 226 * Calculate the ABR attack threshold from the above LAME psymodel table. 227 */ 228static float lame_calc_attack_threshold(int bitrate) 229{ 230 /* Assume max bitrate to start with */ 231 int lower_range = 12, upper_range = 12; 232 int lower_range_kbps = psy_abr_map[12].quality; 233 int upper_range_kbps = psy_abr_map[12].quality; 234 int i; 235 236 /* Determine which bitrates the value specified falls between. 237 * If the loop ends without breaking our above assumption of 320kbps was correct. 238 */ 239 for (i = 1; i < 13; i++) { 240 if (FFMAX(bitrate, psy_abr_map[i].quality) != bitrate) { 241 upper_range = i; 242 upper_range_kbps = psy_abr_map[i ].quality; 243 lower_range = i - 1; 244 lower_range_kbps = psy_abr_map[i - 1].quality; 245 break; /* Upper range found */ 246 } 247 } 248 249 /* Determine which range the value specified is closer to */ 250 if ((upper_range_kbps - bitrate) > (bitrate - lower_range_kbps)) 251 return psy_abr_map[lower_range].st_lrm; 252 return psy_abr_map[upper_range].st_lrm; 253} 254 255/** 256 * LAME psy model specific initialization 257 */ 258static av_cold void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx) 259{ 260 int i, j; 261 262 for (i = 0; i < avctx->channels; i++) { 263 AacPsyChannel *pch = &ctx->ch[i]; 264 265 if (avctx->flags & CODEC_FLAG_QSCALE) 266 pch->attack_threshold = psy_vbr_map[avctx->global_quality / FF_QP2LAMBDA].st_lrm; 267 else 268 pch->attack_threshold = lame_calc_attack_threshold(avctx->bit_rate / avctx->channels / 1000); 269 270 for (j = 0; j < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; j++) 271 pch->prev_energy_subshort[j] = 10.0f; 272 } 273} 274 275/** 276 * Calculate Bark value for given line. 277 */ 278static av_cold float calc_bark(float f) 279{ 280 return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f)); 281} 282 283#define ATH_ADD 4 284/** 285 * Calculate ATH value for given frequency. 286 * Borrowed from Lame. 287 */ 288static av_cold float ath(float f, float add) 289{ 290 f /= 1000.0f; 291 return 3.64 * pow(f, -0.8) 292 - 6.8 * exp(-0.6 * (f - 3.4) * (f - 3.4)) 293 + 6.0 * exp(-0.15 * (f - 8.7) * (f - 8.7)) 294 + (0.6 + 0.04 * add) * 0.001 * f * f * f * f; 295} 296 297static av_cold int psy_3gpp_init(FFPsyContext *ctx) { 298 AacPsyContext *pctx; 299 float bark; 300 int i, j, g, start; 301 float prev, minscale, minath, minsnr, pe_min; 302 const int chan_bitrate = ctx->avctx->bit_rate / ctx->avctx->channels; 303 const int bandwidth = ctx->avctx->cutoff ? ctx->avctx->cutoff : AAC_CUTOFF(ctx->avctx); 304 const float num_bark = calc_bark((float)bandwidth); 305 306 ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext)); 307 pctx = (AacPsyContext*) ctx->model_priv_data; 308 309 pctx->chan_bitrate = chan_bitrate; 310 pctx->frame_bits = chan_bitrate * AAC_BLOCK_SIZE_LONG / ctx->avctx->sample_rate; 311 pctx->pe.min = 8.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f); 312 pctx->pe.max = 12.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f); 313 ctx->bitres.size = 6144 - pctx->frame_bits; 314 ctx->bitres.size -= ctx->bitres.size % 8; 315 pctx->fill_level = ctx->bitres.size; 316 minath = ath(3410, ATH_ADD); 317 for (j = 0; j < 2; j++) { 318 AacPsyCoeffs *coeffs = pctx->psy_coef[j]; 319 const uint8_t *band_sizes = ctx->bands[j]; 320 float line_to_frequency = ctx->avctx->sample_rate / (j ? 256.f : 2048.0f); 321 float avg_chan_bits = chan_bitrate * (j ? 128.0f : 1024.0f) / ctx->avctx->sample_rate; 322 /* reference encoder uses 2.4% here instead of 60% like the spec says */ 323 float bark_pe = 0.024f * PSY_3GPP_BITS_TO_PE(avg_chan_bits) / num_bark; 324 float en_spread_low = j ? PSY_3GPP_EN_SPREAD_LOW_S : PSY_3GPP_EN_SPREAD_LOW_L; 325 /* High energy spreading for long blocks <= 22kbps/channel and short blocks are the same. */ 326 float en_spread_hi = (j || (chan_bitrate <= 22.0f)) ? PSY_3GPP_EN_SPREAD_HI_S : PSY_3GPP_EN_SPREAD_HI_L1; 327 328 i = 0; 329 prev = 0.0; 330 for (g = 0; g < ctx->num_bands[j]; g++) { 331 i += band_sizes[g]; 332 bark = calc_bark((i-1) * line_to_frequency); 333 coeffs[g].barks = (bark + prev) / 2.0; 334 prev = bark; 335 } 336 for (g = 0; g < ctx->num_bands[j] - 1; g++) { 337 AacPsyCoeffs *coeff = &coeffs[g]; 338 float bark_width = coeffs[g+1].barks - coeffs->barks; 339 coeff->spread_low[0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_LOW); 340 coeff->spread_hi [0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_HI); 341 coeff->spread_low[1] = pow(10.0, -bark_width * en_spread_low); 342 coeff->spread_hi [1] = pow(10.0, -bark_width * en_spread_hi); 343 pe_min = bark_pe * bark_width; 344 minsnr = exp2(pe_min / band_sizes[g]) - 1.5f; 345 coeff->min_snr = av_clipf(1.0f / minsnr, PSY_SNR_25DB, PSY_SNR_1DB); 346 } 347 start = 0; 348 for (g = 0; g < ctx->num_bands[j]; g++) { 349 minscale = ath(start * line_to_frequency, ATH_ADD); 350 for (i = 1; i < band_sizes[g]; i++) 351 minscale = FFMIN(minscale, ath((start + i) * line_to_frequency, ATH_ADD)); 352 coeffs[g].ath = minscale - minath; 353 start += band_sizes[g]; 354 } 355 } 356 357 pctx->ch = av_mallocz_array(ctx->avctx->channels, sizeof(AacPsyChannel)); 358 359 lame_window_init(pctx, ctx->avctx); 360 361 return 0; 362} 363 364/** 365 * IIR filter used in block switching decision 366 */ 367static float iir_filter(int in, float state[2]) 368{ 369 float ret; 370 371 ret = 0.7548f * (in - state[0]) + 0.5095f * state[1]; 372 state[0] = in; 373 state[1] = ret; 374 return ret; 375} 376 377/** 378 * window grouping information stored as bits (0 - new group, 1 - group continues) 379 */ 380static const uint8_t window_grouping[9] = { 381 0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36 382}; 383 384/** 385 * Tell encoder which window types to use. 386 * @see 3GPP TS26.403 5.4.1 "Blockswitching" 387 */ 388static av_unused FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx, 389 const int16_t *audio, 390 const int16_t *la, 391 int channel, int prev_type) 392{ 393 int i, j; 394 int br = ctx->avctx->bit_rate / ctx->avctx->channels; 395 int attack_ratio = br <= 16000 ? 18 : 10; 396 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data; 397 AacPsyChannel *pch = &pctx->ch[channel]; 398 uint8_t grouping = 0; 399 int next_type = pch->next_window_seq; 400 FFPsyWindowInfo wi = { { 0 } }; 401 402 if (la) { 403 float s[8], v; 404 int switch_to_eight = 0; 405 float sum = 0.0, sum2 = 0.0; 406 int attack_n = 0; 407 int stay_short = 0; 408 for (i = 0; i < 8; i++) { 409 for (j = 0; j < 128; j++) { 410 v = iir_filter(la[i*128+j], pch->iir_state); 411 sum += v*v; 412 } 413 s[i] = sum; 414 sum2 += sum; 415 } 416 for (i = 0; i < 8; i++) { 417 if (s[i] > pch->win_energy * attack_ratio) { 418 attack_n = i + 1; 419 switch_to_eight = 1; 420 break; 421 } 422 } 423 pch->win_energy = pch->win_energy*7/8 + sum2/64; 424 425 wi.window_type[1] = prev_type; 426 switch (prev_type) { 427 case ONLY_LONG_SEQUENCE: 428 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE; 429 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE; 430 break; 431 case LONG_START_SEQUENCE: 432 wi.window_type[0] = EIGHT_SHORT_SEQUENCE; 433 grouping = pch->next_grouping; 434 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE; 435 break; 436 case LONG_STOP_SEQUENCE: 437 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE; 438 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE; 439 break; 440 case EIGHT_SHORT_SEQUENCE: 441 stay_short = next_type == EIGHT_SHORT_SEQUENCE || switch_to_eight; 442 wi.window_type[0] = stay_short ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE; 443 grouping = next_type == EIGHT_SHORT_SEQUENCE ? pch->next_grouping : 0; 444 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE; 445 break; 446 } 447 448 pch->next_grouping = window_grouping[attack_n]; 449 pch->next_window_seq = next_type; 450 } else { 451 for (i = 0; i < 3; i++) 452 wi.window_type[i] = prev_type; 453 grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0; 454 } 455 456 wi.window_shape = 1; 457 if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) { 458 wi.num_windows = 1; 459 wi.grouping[0] = 1; 460 } else { 461 int lastgrp = 0; 462 wi.num_windows = 8; 463 for (i = 0; i < 8; i++) { 464 if (!((grouping >> i) & 1)) 465 lastgrp = i; 466 wi.grouping[lastgrp]++; 467 } 468 } 469 470 return wi; 471} 472 473/* 5.6.1.2 "Calculation of Bit Demand" */ 474static int calc_bit_demand(AacPsyContext *ctx, float pe, int bits, int size, 475 int short_window) 476{ 477 const float bitsave_slope = short_window ? PSY_3GPP_SAVE_SLOPE_S : PSY_3GPP_SAVE_SLOPE_L; 478 const float bitsave_add = short_window ? PSY_3GPP_SAVE_ADD_S : PSY_3GPP_SAVE_ADD_L; 479 const float bitspend_slope = short_window ? PSY_3GPP_SPEND_SLOPE_S : PSY_3GPP_SPEND_SLOPE_L; 480 const float bitspend_add = short_window ? PSY_3GPP_SPEND_ADD_S : PSY_3GPP_SPEND_ADD_L; 481 const float clip_low = short_window ? PSY_3GPP_CLIP_LO_S : PSY_3GPP_CLIP_LO_L; 482 const float clip_high = short_window ? PSY_3GPP_CLIP_HI_S : PSY_3GPP_CLIP_HI_L; 483 float clipped_pe, bit_save, bit_spend, bit_factor, fill_level; 484 485 ctx->fill_level += ctx->frame_bits - bits; 486 ctx->fill_level = av_clip(ctx->fill_level, 0, size); 487 fill_level = av_clipf((float)ctx->fill_level / size, clip_low, clip_high); 488 clipped_pe = av_clipf(pe, ctx->pe.min, ctx->pe.max); 489 bit_save = (fill_level + bitsave_add) * bitsave_slope; 490 assert(bit_save <= 0.3f && bit_save >= -0.05000001f); 491 bit_spend = (fill_level + bitspend_add) * bitspend_slope; 492 assert(bit_spend <= 0.5f && bit_spend >= -0.1f); 493 /* The bit factor graph in the spec is obviously incorrect. 494 * bit_spend + ((bit_spend - bit_spend))... 495 * The reference encoder subtracts everything from 1, but also seems incorrect. 496 * 1 - bit_save + ((bit_spend + bit_save))... 497 * Hopefully below is correct. 498 */ 499 bit_factor = 1.0f - bit_save + ((bit_spend - bit_save) / (ctx->pe.max - ctx->pe.min)) * (clipped_pe - ctx->pe.min); 500 /* NOTE: The reference encoder attempts to center pe max/min around the current pe. */ 501 ctx->pe.max = FFMAX(pe, ctx->pe.max); 502 ctx->pe.min = FFMIN(pe, ctx->pe.min); 503 504 return FFMIN(ctx->frame_bits * bit_factor, ctx->frame_bits + size - bits); 505} 506 507static float calc_pe_3gpp(AacPsyBand *band) 508{ 509 float pe, a; 510 511 band->pe = 0.0f; 512 band->pe_const = 0.0f; 513 band->active_lines = 0.0f; 514 if (band->energy > band->thr) { 515 a = log2f(band->energy); 516 pe = a - log2f(band->thr); 517 band->active_lines = band->nz_lines; 518 if (pe < PSY_3GPP_C1) { 519 pe = pe * PSY_3GPP_C3 + PSY_3GPP_C2; 520 a = a * PSY_3GPP_C3 + PSY_3GPP_C2; 521 band->active_lines *= PSY_3GPP_C3; 522 } 523 band->pe = pe * band->nz_lines; 524 band->pe_const = a * band->nz_lines; 525 } 526 527 return band->pe; 528} 529 530static float calc_reduction_3gpp(float a, float desired_pe, float pe, 531 float active_lines) 532{ 533 float thr_avg, reduction; 534 535 if(active_lines == 0.0) 536 return 0; 537 538 thr_avg = exp2f((a - pe) / (4.0f * active_lines)); 539 reduction = exp2f((a - desired_pe) / (4.0f * active_lines)) - thr_avg; 540 541 return FFMAX(reduction, 0.0f); 542} 543 544static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr, 545 float reduction) 546{ 547 float thr = band->thr; 548 549 if (band->energy > thr) { 550 thr = sqrtf(thr); 551 thr = sqrtf(thr) + reduction; 552 thr *= thr; 553 thr *= thr; 554 555 /* This deviates from the 3GPP spec to match the reference encoder. 556 * It performs min(thr_reduced, max(thr, energy/min_snr)) only for bands 557 * that have hole avoidance on (active or inactive). It always reduces the 558 * threshold of bands with hole avoidance off. 559 */ 560 if (thr > band->energy * min_snr && band->avoid_holes != PSY_3GPP_AH_NONE) { 561 thr = FFMAX(band->thr, band->energy * min_snr); 562 band->avoid_holes = PSY_3GPP_AH_ACTIVE; 563 } 564 } 565 566 return thr; 567} 568 569#ifndef calc_thr_3gpp 570static void calc_thr_3gpp(const FFPsyWindowInfo *wi, const int num_bands, AacPsyChannel *pch, 571 const uint8_t *band_sizes, const float *coefs) 572{ 573 int i, w, g; 574 int start = 0; 575 for (w = 0; w < wi->num_windows*16; w += 16) { 576 for (g = 0; g < num_bands; g++) { 577 AacPsyBand *band = &pch->band[w+g]; 578 579 float form_factor = 0.0f; 580 float Temp; 581 band->energy = 0.0f; 582 for (i = 0; i < band_sizes[g]; i++) { 583 band->energy += coefs[start+i] * coefs[start+i]; 584 form_factor += sqrtf(fabs(coefs[start+i])); 585 } 586 Temp = band->energy > 0 ? sqrtf((float)band_sizes[g] / band->energy) : 0; 587 band->thr = band->energy * 0.001258925f; 588 band->nz_lines = form_factor * sqrtf(Temp); 589 590 start += band_sizes[g]; 591 } 592 } 593} 594#endif /* calc_thr_3gpp */ 595 596#ifndef psy_hp_filter 597static void psy_hp_filter(const float *firbuf, float *hpfsmpl, const float *psy_fir_coeffs) 598{ 599 int i, j; 600 for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) { 601 float sum1, sum2; 602 sum1 = firbuf[i + (PSY_LAME_FIR_LEN - 1) / 2]; 603 sum2 = 0.0; 604 for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) { 605 sum1 += psy_fir_coeffs[j] * (firbuf[i + j] + firbuf[i + PSY_LAME_FIR_LEN - j]); 606 sum2 += psy_fir_coeffs[j + 1] * (firbuf[i + j + 1] + firbuf[i + PSY_LAME_FIR_LEN - j - 1]); 607 } 608 /* NOTE: The LAME psymodel expects it's input in the range -32768 to 32768. 609 * Tuning this for normalized floats would be difficult. */ 610 hpfsmpl[i] = (sum1 + sum2) * 32768.0f; 611 } 612} 613#endif /* psy_hp_filter */ 614 615/** 616 * Calculate band thresholds as suggested in 3GPP TS26.403 617 */ 618static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel, 619 const float *coefs, const FFPsyWindowInfo *wi) 620{ 621 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data; 622 AacPsyChannel *pch = &pctx->ch[channel]; 623 int i, w, g; 624 float desired_bits, desired_pe, delta_pe, reduction= NAN, spread_en[128] = {0}; 625 float a = 0.0f, active_lines = 0.0f, norm_fac = 0.0f; 626 float pe = pctx->chan_bitrate > 32000 ? 0.0f : FFMAX(50.0f, 100.0f - pctx->chan_bitrate * 100.0f / 32000.0f); 627 const int num_bands = ctx->num_bands[wi->num_windows == 8]; 628 const uint8_t *band_sizes = ctx->bands[wi->num_windows == 8]; 629 AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8]; 630 const float avoid_hole_thr = wi->num_windows == 8 ? PSY_3GPP_AH_THR_SHORT : PSY_3GPP_AH_THR_LONG; 631 632 //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation" 633 calc_thr_3gpp(wi, num_bands, pch, band_sizes, coefs); 634 635 //modify thresholds and energies - spread, threshold in quiet, pre-echo control 636 for (w = 0; w < wi->num_windows*16; w += 16) { 637 AacPsyBand *bands = &pch->band[w]; 638 639 /* 5.4.2.3 "Spreading" & 5.4.3 "Spread Energy Calculation" */ 640 spread_en[0] = bands[0].energy; 641 for (g = 1; g < num_bands; g++) { 642 bands[g].thr = FFMAX(bands[g].thr, bands[g-1].thr * coeffs[g].spread_hi[0]); 643 spread_en[w+g] = FFMAX(bands[g].energy, spread_en[w+g-1] * coeffs[g].spread_hi[1]); 644 } 645 for (g = num_bands - 2; g >= 0; g--) { 646 bands[g].thr = FFMAX(bands[g].thr, bands[g+1].thr * coeffs[g].spread_low[0]); 647 spread_en[w+g] = FFMAX(spread_en[w+g], spread_en[w+g+1] * coeffs[g].spread_low[1]); 648 } 649 //5.4.2.4 "Threshold in quiet" 650 for (g = 0; g < num_bands; g++) { 651 AacPsyBand *band = &bands[g]; 652 653 band->thr_quiet = band->thr = FFMAX(band->thr, coeffs[g].ath); 654 //5.4.2.5 "Pre-echo control" 655 if (!(wi->window_type[0] == LONG_STOP_SEQUENCE || (wi->window_type[1] == LONG_START_SEQUENCE && !w))) 656 band->thr = FFMAX(PSY_3GPP_RPEMIN*band->thr, FFMIN(band->thr, 657 PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet)); 658 659 /* 5.6.1.3.1 "Preparatory steps of the perceptual entropy calculation" */ 660 pe += calc_pe_3gpp(band); 661 a += band->pe_const; 662 active_lines += band->active_lines; 663 664 /* 5.6.1.3.3 "Selection of the bands for avoidance of holes" */ 665 if (spread_en[w+g] * avoid_hole_thr > band->energy || coeffs[g].min_snr > 1.0f) 666 band->avoid_holes = PSY_3GPP_AH_NONE; 667 else 668 band->avoid_holes = PSY_3GPP_AH_INACTIVE; 669 } 670 } 671 672 /* 5.6.1.3.2 "Calculation of the desired perceptual entropy" */ 673 ctx->ch[channel].entropy = pe; 674 desired_bits = calc_bit_demand(pctx, pe, ctx->bitres.bits, ctx->bitres.size, wi->num_windows == 8); 675 desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); 676 /* NOTE: PE correction is kept simple. During initial testing it had very 677 * little effect on the final bitrate. Probably a good idea to come 678 * back and do more testing later. 679 */ 680 if (ctx->bitres.bits > 0) 681 desired_pe *= av_clipf(pctx->pe.previous / PSY_3GPP_BITS_TO_PE(ctx->bitres.bits), 682 0.85f, 1.15f); 683 pctx->pe.previous = PSY_3GPP_BITS_TO_PE(desired_bits); 684 685 if (desired_pe < pe) { 686 /* 5.6.1.3.4 "First Estimation of the reduction value" */ 687 for (w = 0; w < wi->num_windows*16; w += 16) { 688 reduction = calc_reduction_3gpp(a, desired_pe, pe, active_lines); 689 pe = 0.0f; 690 a = 0.0f; 691 active_lines = 0.0f; 692 for (g = 0; g < num_bands; g++) { 693 AacPsyBand *band = &pch->band[w+g]; 694 695 band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction); 696 /* recalculate PE */ 697 pe += calc_pe_3gpp(band); 698 a += band->pe_const; 699 active_lines += band->active_lines; 700 } 701 } 702 703 /* 5.6.1.3.5 "Second Estimation of the reduction value" */ 704 for (i = 0; i < 2; i++) { 705 float pe_no_ah = 0.0f, desired_pe_no_ah; 706 active_lines = a = 0.0f; 707 for (w = 0; w < wi->num_windows*16; w += 16) { 708 for (g = 0; g < num_bands; g++) { 709 AacPsyBand *band = &pch->band[w+g]; 710 711 if (band->avoid_holes != PSY_3GPP_AH_ACTIVE) { 712 pe_no_ah += band->pe; 713 a += band->pe_const; 714 active_lines += band->active_lines; 715 } 716 } 717 } 718 desired_pe_no_ah = FFMAX(desired_pe - (pe - pe_no_ah), 0.0f); 719 if (active_lines > 0.0f) 720 reduction += calc_reduction_3gpp(a, desired_pe_no_ah, pe_no_ah, active_lines); 721 722 pe = 0.0f; 723 for (w = 0; w < wi->num_windows*16; w += 16) { 724 for (g = 0; g < num_bands; g++) { 725 AacPsyBand *band = &pch->band[w+g]; 726 727 if (active_lines > 0.0f) 728 band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction); 729 pe += calc_pe_3gpp(band); 730 band->norm_fac = band->active_lines / band->thr; 731 norm_fac += band->norm_fac; 732 } 733 } 734 delta_pe = desired_pe - pe; 735 if (fabs(delta_pe) > 0.05f * desired_pe) 736 break; 737 } 738 739 if (pe < 1.15f * desired_pe) { 740 /* 6.6.1.3.6 "Final threshold modification by linearization" */ 741 norm_fac = 1.0f / norm_fac; 742 for (w = 0; w < wi->num_windows*16; w += 16) { 743 for (g = 0; g < num_bands; g++) { 744 AacPsyBand *band = &pch->band[w+g]; 745 746 if (band->active_lines > 0.5f) { 747 float delta_sfb_pe = band->norm_fac * norm_fac * delta_pe; 748 float thr = band->thr; 749 750 thr *= exp2f(delta_sfb_pe / band->active_lines); 751 if (thr > coeffs[g].min_snr * band->energy && band->avoid_holes == PSY_3GPP_AH_INACTIVE) 752 thr = FFMAX(band->thr, coeffs[g].min_snr * band->energy); 753 band->thr = thr; 754 } 755 } 756 } 757 } else { 758 /* 5.6.1.3.7 "Further perceptual entropy reduction" */ 759 g = num_bands; 760 while (pe > desired_pe && g--) { 761 for (w = 0; w < wi->num_windows*16; w+= 16) { 762 AacPsyBand *band = &pch->band[w+g]; 763 if (band->avoid_holes != PSY_3GPP_AH_NONE && coeffs[g].min_snr < PSY_SNR_1DB) { 764 coeffs[g].min_snr = PSY_SNR_1DB; 765 band->thr = band->energy * PSY_SNR_1DB; 766 pe += band->active_lines * 1.5f - band->pe; 767 } 768 } 769 } 770 /* TODO: allow more holes (unused without mid/side) */ 771 } 772 } 773 774 for (w = 0; w < wi->num_windows*16; w += 16) { 775 for (g = 0; g < num_bands; g++) { 776 AacPsyBand *band = &pch->band[w+g]; 777 FFPsyBand *psy_band = &ctx->ch[channel].psy_bands[w+g]; 778 779 psy_band->threshold = band->thr; 780 psy_band->energy = band->energy; 781 } 782 } 783 784 memcpy(pch->prev_band, pch->band, sizeof(pch->band)); 785} 786 787static void psy_3gpp_analyze(FFPsyContext *ctx, int channel, 788 const float **coeffs, const FFPsyWindowInfo *wi) 789{ 790 int ch; 791 FFPsyChannelGroup *group = ff_psy_find_group(ctx, channel); 792 793 for (ch = 0; ch < group->num_ch; ch++) 794 psy_3gpp_analyze_channel(ctx, channel + ch, coeffs[ch], &wi[ch]); 795} 796 797static av_cold void psy_3gpp_end(FFPsyContext *apc) 798{ 799 AacPsyContext *pctx = (AacPsyContext*) apc->model_priv_data; 800 av_freep(&pctx->ch); 801 av_freep(&apc->model_priv_data); 802} 803 804static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock) 805{ 806 int blocktype = ONLY_LONG_SEQUENCE; 807 if (uselongblock) { 808 if (ctx->next_window_seq == EIGHT_SHORT_SEQUENCE) 809 blocktype = LONG_STOP_SEQUENCE; 810 } else { 811 blocktype = EIGHT_SHORT_SEQUENCE; 812 if (ctx->next_window_seq == ONLY_LONG_SEQUENCE) 813 ctx->next_window_seq = LONG_START_SEQUENCE; 814 if (ctx->next_window_seq == LONG_STOP_SEQUENCE) 815 ctx->next_window_seq = EIGHT_SHORT_SEQUENCE; 816 } 817 818 wi->window_type[0] = ctx->next_window_seq; 819 ctx->next_window_seq = blocktype; 820} 821 822static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx, const float *audio, 823 const float *la, int channel, int prev_type) 824{ 825 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data; 826 AacPsyChannel *pch = &pctx->ch[channel]; 827 int grouping = 0; 828 int uselongblock = 1; 829 int attacks[AAC_NUM_BLOCKS_SHORT + 1] = { 0 }; 830 int i; 831 FFPsyWindowInfo wi = { { 0 } }; 832 833 if (la) { 834 float hpfsmpl[AAC_BLOCK_SIZE_LONG]; 835 float const *pf = hpfsmpl; 836 float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS]; 837 float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS]; 838 float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 }; 839 const float *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN); 840 int att_sum = 0; 841 842 /* LAME comment: apply high pass filter of fs/4 */ 843 psy_hp_filter(firbuf, hpfsmpl, psy_fir_coeffs); 844 845 /* Calculate the energies of each sub-shortblock */ 846 for (i = 0; i < PSY_LAME_NUM_SUBBLOCKS; i++) { 847 energy_subshort[i] = pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS)]; 848 assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)] > 0); 849 attack_intensity[i] = energy_subshort[i] / pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)]; 850 energy_short[0] += energy_subshort[i]; 851 } 852 853 for (i = 0; i < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; i++) { 854 float const *const pfe = pf + AAC_BLOCK_SIZE_LONG / (AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS); 855 float p = 1.0f; 856 for (; pf < pfe; pf++) 857 p = FFMAX(p, fabsf(*pf)); 858 pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p; 859 energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p; 860 /* NOTE: The indexes below are [i + 3 - 2] in the LAME source. 861 * Obviously the 3 and 2 have some significance, or this would be just [i + 1] 862 * (which is what we use here). What the 3 stands for is ambiguous, as it is both 863 * number of short blocks, and the number of sub-short blocks. 864 * It seems that LAME is comparing each sub-block to sub-block + 1 in the 865 * previous block. 866 */ 867 if (p > energy_subshort[i + 1]) 868 p = p / energy_subshort[i + 1]; 869 else if (energy_subshort[i + 1] > p * 10.0f) 870 p = energy_subshort[i + 1] / (p * 10.0f); 871 else 872 p = 0.0; 873 attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p; 874 } 875 876 /* compare energy between sub-short blocks */ 877 for (i = 0; i < (AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS; i++) 878 if (!attacks[i / PSY_LAME_NUM_SUBBLOCKS]) 879 if (attack_intensity[i] > pch->attack_threshold) 880 attacks[i / PSY_LAME_NUM_SUBBLOCKS] = (i % PSY_LAME_NUM_SUBBLOCKS) + 1; 881 882 /* should have energy change between short blocks, in order to avoid periodic signals */ 883 /* Good samples to show the effect are Trumpet test songs */ 884 /* GB: tuned (1) to avoid too many short blocks for test sample TRUMPET */ 885 /* RH: tuned (2) to let enough short blocks through for test sample FSOL and SNAPS */ 886 for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) { 887 float const u = energy_short[i - 1]; 888 float const v = energy_short[i]; 889 float const m = FFMAX(u, v); 890 if (m < 40000) { /* (2) */ 891 if (u < 1.7f * v && v < 1.7f * u) { /* (1) */ 892 if (i == 1 && attacks[0] < attacks[i]) 893 attacks[0] = 0; 894 attacks[i] = 0; 895 } 896 } 897 att_sum += attacks[i]; 898 } 899 900 if (attacks[0] <= pch->prev_attack) 901 attacks[0] = 0; 902 903 att_sum += attacks[0]; 904 /* 3 below indicates the previous attack happened in the last sub-block of the previous sequence */ 905 if (pch->prev_attack == 3 || att_sum) { 906 uselongblock = 0; 907 908 for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) 909 if (attacks[i] && attacks[i-1]) 910 attacks[i] = 0; 911 } 912 } else { 913 /* We have no lookahead info, so just use same type as the previous sequence. */ 914 uselongblock = !(prev_type == EIGHT_SHORT_SEQUENCE); 915 } 916 917 lame_apply_block_type(pch, &wi, uselongblock); 918 919 wi.window_type[1] = prev_type; 920 if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) { 921 wi.num_windows = 1; 922 wi.grouping[0] = 1; 923 if (wi.window_type[0] == LONG_START_SEQUENCE) 924 wi.window_shape = 0; 925 else 926 wi.window_shape = 1; 927 } else { 928 int lastgrp = 0; 929 930 wi.num_windows = 8; 931 wi.window_shape = 0; 932 for (i = 0; i < 8; i++) { 933 if (!((pch->next_grouping >> i) & 1)) 934 lastgrp = i; 935 wi.grouping[lastgrp]++; 936 } 937 } 938 939 /* Determine grouping, based on the location of the first attack, and save for 940 * the next frame. 941 * FIXME: Move this to analysis. 942 * TODO: Tune groupings depending on attack location 943 * TODO: Handle more than one attack in a group 944 */ 945 for (i = 0; i < 9; i++) { 946 if (attacks[i]) { 947 grouping = i; 948 break; 949 } 950 } 951 pch->next_grouping = window_grouping[grouping]; 952 953 pch->prev_attack = attacks[8]; 954 955 return wi; 956} 957 958const FFPsyModel ff_aac_psy_model = 959{ 960 .name = "3GPP TS 26.403-inspired model", 961 .init = psy_3gpp_init, 962 .window = psy_lame_window, 963 .analyze = psy_3gpp_analyze, 964 .end = psy_3gpp_end, 965}; 966