You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

1006 lines
38KB

  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. * @file
  23. * AAC encoder psychoacoustic model
  24. */
  25. #include "libavutil/attributes.h"
  26. #include "libavutil/libm.h"
  27. #include "avcodec.h"
  28. #include "aactab.h"
  29. #include "psymodel.h"
  30. /***********************************
  31. * TODOs:
  32. * try other bitrate controlling mechanism (maybe use ratecontrol.c?)
  33. * control quality for quality-based output
  34. **********************************/
  35. /**
  36. * constants for 3GPP AAC psychoacoustic model
  37. * @{
  38. */
  39. #define PSY_3GPP_THR_SPREAD_HI 1.5f // spreading factor for low-to-hi threshold spreading (15 dB/Bark)
  40. #define PSY_3GPP_THR_SPREAD_LOW 3.0f // spreading factor for hi-to-low threshold spreading (30 dB/Bark)
  41. /* spreading factor for low-to-hi energy spreading, long block, > 22kbps/channel (20dB/Bark) */
  42. #define PSY_3GPP_EN_SPREAD_HI_L1 2.0f
  43. /* spreading factor for low-to-hi energy spreading, long block, <= 22kbps/channel (15dB/Bark) */
  44. #define PSY_3GPP_EN_SPREAD_HI_L2 1.5f
  45. /* spreading factor for low-to-hi energy spreading, short block (15 dB/Bark) */
  46. #define PSY_3GPP_EN_SPREAD_HI_S 1.5f
  47. /* spreading factor for hi-to-low energy spreading, long block (30dB/Bark) */
  48. #define PSY_3GPP_EN_SPREAD_LOW_L 3.0f
  49. /* spreading factor for hi-to-low energy spreading, short block (20dB/Bark) */
  50. #define PSY_3GPP_EN_SPREAD_LOW_S 2.0f
  51. #define PSY_3GPP_RPEMIN 0.01f
  52. #define PSY_3GPP_RPELEV 2.0f
  53. #define PSY_3GPP_C1 3.0f /* log2(8) */
  54. #define PSY_3GPP_C2 1.3219281f /* log2(2.5) */
  55. #define PSY_3GPP_C3 0.55935729f /* 1 - C2 / C1 */
  56. #define PSY_SNR_1DB 7.9432821e-1f /* -1dB */
  57. #define PSY_SNR_25DB 3.1622776e-3f /* -25dB */
  58. #define PSY_3GPP_SAVE_SLOPE_L -0.46666667f
  59. #define PSY_3GPP_SAVE_SLOPE_S -0.36363637f
  60. #define PSY_3GPP_SAVE_ADD_L -0.84285712f
  61. #define PSY_3GPP_SAVE_ADD_S -0.75f
  62. #define PSY_3GPP_SPEND_SLOPE_L 0.66666669f
  63. #define PSY_3GPP_SPEND_SLOPE_S 0.81818181f
  64. #define PSY_3GPP_SPEND_ADD_L -0.35f
  65. #define PSY_3GPP_SPEND_ADD_S -0.26111111f
  66. #define PSY_3GPP_CLIP_LO_L 0.2f
  67. #define PSY_3GPP_CLIP_LO_S 0.2f
  68. #define PSY_3GPP_CLIP_HI_L 0.95f
  69. #define PSY_3GPP_CLIP_HI_S 0.75f
  70. #define PSY_3GPP_AH_THR_LONG 0.5f
  71. #define PSY_3GPP_AH_THR_SHORT 0.63f
  72. enum {
  73. PSY_3GPP_AH_NONE,
  74. PSY_3GPP_AH_INACTIVE,
  75. PSY_3GPP_AH_ACTIVE
  76. };
  77. #define PSY_3GPP_BITS_TO_PE(bits) ((bits) * 1.18f)
  78. /* LAME psy model constants */
  79. #define PSY_LAME_FIR_LEN 21 ///< LAME psy model FIR order
  80. #define AAC_BLOCK_SIZE_LONG 1024 ///< long block size
  81. #define AAC_BLOCK_SIZE_SHORT 128 ///< short block size
  82. #define AAC_NUM_BLOCKS_SHORT 8 ///< number of blocks in a short sequence
  83. #define PSY_LAME_NUM_SUBBLOCKS 3 ///< Number of sub-blocks in each short block
  84. /**
  85. * @}
  86. */
  87. /**
  88. * information for single band used by 3GPP TS26.403-inspired psychoacoustic model
  89. */
  90. typedef struct AacPsyBand{
  91. float energy; ///< band energy
  92. float thr; ///< energy threshold
  93. float thr_quiet; ///< threshold in quiet
  94. float nz_lines; ///< number of non-zero spectral lines
  95. float active_lines; ///< number of active spectral lines
  96. float pe; ///< perceptual entropy
  97. float pe_const; ///< constant part of the PE calculation
  98. float norm_fac; ///< normalization factor for linearization
  99. int avoid_holes; ///< hole avoidance flag
  100. }AacPsyBand;
  101. /**
  102. * single/pair channel context for psychoacoustic model
  103. */
  104. typedef struct AacPsyChannel{
  105. AacPsyBand band[128]; ///< bands information
  106. AacPsyBand prev_band[128]; ///< bands information from the previous frame
  107. float win_energy; ///< sliding average of channel energy
  108. float iir_state[2]; ///< hi-pass IIR filter state
  109. uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence)
  110. enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame
  111. /* LAME psy model specific members */
  112. float attack_threshold; ///< attack threshold for this channel
  113. float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS];
  114. int prev_attack; ///< attack value for the last short block in the previous sequence
  115. }AacPsyChannel;
  116. /**
  117. * psychoacoustic model frame type-dependent coefficients
  118. */
  119. typedef struct AacPsyCoeffs{
  120. float ath; ///< absolute threshold of hearing per bands
  121. float barks; ///< Bark value for each spectral band in long frame
  122. float spread_low[2]; ///< spreading factor for low-to-high threshold spreading in long frame
  123. float spread_hi [2]; ///< spreading factor for high-to-low threshold spreading in long frame
  124. float min_snr; ///< minimal SNR
  125. }AacPsyCoeffs;
  126. /**
  127. * 3GPP TS26.403-inspired psychoacoustic model specific data
  128. */
  129. typedef struct AacPsyContext{
  130. int chan_bitrate; ///< bitrate per channel
  131. int frame_bits; ///< average bits per frame
  132. int fill_level; ///< bit reservoir fill level
  133. struct {
  134. float min; ///< minimum allowed PE for bit factor calculation
  135. float max; ///< maximum allowed PE for bit factor calculation
  136. float previous; ///< allowed PE of the previous frame
  137. float correction; ///< PE correction factor
  138. } pe;
  139. AacPsyCoeffs psy_coef[2][64];
  140. AacPsyChannel *ch;
  141. }AacPsyContext;
  142. /**
  143. * LAME psy model preset struct
  144. */
  145. typedef struct PsyLamePreset {
  146. int quality; ///< Quality to map the rest of the vaules to.
  147. /* This is overloaded to be both kbps per channel in ABR mode, and
  148. * requested quality in constant quality mode.
  149. */
  150. float st_lrm; ///< short threshold for L, R, and M channels
  151. } PsyLamePreset;
  152. /**
  153. * LAME psy model preset table for ABR
  154. */
  155. static const PsyLamePreset psy_abr_map[] = {
  156. /* TODO: Tuning. These were taken from LAME. */
  157. /* kbps/ch st_lrm */
  158. { 8, 6.60},
  159. { 16, 6.60},
  160. { 24, 6.60},
  161. { 32, 6.60},
  162. { 40, 6.60},
  163. { 48, 6.60},
  164. { 56, 6.60},
  165. { 64, 6.40},
  166. { 80, 6.00},
  167. { 96, 5.60},
  168. {112, 5.20},
  169. {128, 5.20},
  170. {160, 5.20}
  171. };
  172. /**
  173. * LAME psy model preset table for constant quality
  174. */
  175. static const PsyLamePreset psy_vbr_map[] = {
  176. /* vbr_q st_lrm */
  177. { 0, 4.20},
  178. { 1, 4.20},
  179. { 2, 4.20},
  180. { 3, 4.20},
  181. { 4, 4.20},
  182. { 5, 4.20},
  183. { 6, 4.20},
  184. { 7, 4.20},
  185. { 8, 4.20},
  186. { 9, 4.20},
  187. {10, 4.20}
  188. };
  189. /**
  190. * LAME psy model FIR coefficient table
  191. */
  192. static const float psy_fir_coeffs[] = {
  193. -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2,
  194. -3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2,
  195. -5.52212e-17 * 2, -0.313819 * 2
  196. };
  197. #if ARCH_MIPS
  198. # include "mips/aacpsy_mips.h"
  199. #endif /* ARCH_MIPS */
  200. /**
  201. * Calculate the ABR attack threshold from the above LAME psymodel table.
  202. */
  203. static float lame_calc_attack_threshold(int bitrate)
  204. {
  205. /* Assume max bitrate to start with */
  206. int lower_range = 12, upper_range = 12;
  207. int lower_range_kbps = psy_abr_map[12].quality;
  208. int upper_range_kbps = psy_abr_map[12].quality;
  209. int i;
  210. /* Determine which bitrates the value specified falls between.
  211. * If the loop ends without breaking our above assumption of 320kbps was correct.
  212. */
  213. for (i = 1; i < 13; i++) {
  214. if (FFMAX(bitrate, psy_abr_map[i].quality) != bitrate) {
  215. upper_range = i;
  216. upper_range_kbps = psy_abr_map[i ].quality;
  217. lower_range = i - 1;
  218. lower_range_kbps = psy_abr_map[i - 1].quality;
  219. break; /* Upper range found */
  220. }
  221. }
  222. /* Determine which range the value specified is closer to */
  223. if ((upper_range_kbps - bitrate) > (bitrate - lower_range_kbps))
  224. return psy_abr_map[lower_range].st_lrm;
  225. return psy_abr_map[upper_range].st_lrm;
  226. }
  227. /**
  228. * LAME psy model specific initialization
  229. */
  230. static av_cold void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx)
  231. {
  232. int i, j;
  233. for (i = 0; i < avctx->channels; i++) {
  234. AacPsyChannel *pch = &ctx->ch[i];
  235. if (avctx->flags & AV_CODEC_FLAG_QSCALE)
  236. pch->attack_threshold = psy_vbr_map[avctx->global_quality / FF_QP2LAMBDA].st_lrm;
  237. else
  238. pch->attack_threshold = lame_calc_attack_threshold(avctx->bit_rate / avctx->channels / 1000);
  239. for (j = 0; j < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; j++)
  240. pch->prev_energy_subshort[j] = 10.0f;
  241. }
  242. }
  243. /**
  244. * Calculate Bark value for given line.
  245. */
  246. static av_cold float calc_bark(float f)
  247. {
  248. return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f));
  249. }
  250. #define ATH_ADD 4
  251. /**
  252. * Calculate ATH value for given frequency.
  253. * Borrowed from Lame.
  254. */
  255. static av_cold float ath(float f, float add)
  256. {
  257. f /= 1000.0f;
  258. return 3.64 * pow(f, -0.8)
  259. - 6.8 * exp(-0.6 * (f - 3.4) * (f - 3.4))
  260. + 6.0 * exp(-0.15 * (f - 8.7) * (f - 8.7))
  261. + (0.6 + 0.04 * add) * 0.001 * f * f * f * f;
  262. }
  263. static av_cold int psy_3gpp_init(FFPsyContext *ctx) {
  264. AacPsyContext *pctx;
  265. float bark;
  266. int i, j, g, start;
  267. float prev, minscale, minath, minsnr, pe_min;
  268. const int chan_bitrate = ctx->avctx->bit_rate / ctx->avctx->channels;
  269. const int bandwidth = ctx->avctx->cutoff ? ctx->avctx->cutoff : AAC_CUTOFF(ctx->avctx);
  270. const float num_bark = calc_bark((float)bandwidth);
  271. ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext));
  272. if (!ctx->model_priv_data)
  273. return AVERROR(ENOMEM);
  274. pctx = (AacPsyContext*) ctx->model_priv_data;
  275. pctx->chan_bitrate = chan_bitrate;
  276. pctx->frame_bits = chan_bitrate * AAC_BLOCK_SIZE_LONG / ctx->avctx->sample_rate;
  277. pctx->pe.min = 8.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
  278. pctx->pe.max = 12.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
  279. ctx->bitres.size = 6144 - pctx->frame_bits;
  280. ctx->bitres.size -= ctx->bitres.size % 8;
  281. pctx->fill_level = ctx->bitres.size;
  282. minath = ath(3410 - 0.733 * ATH_ADD, ATH_ADD);
  283. for (j = 0; j < 2; j++) {
  284. AacPsyCoeffs *coeffs = pctx->psy_coef[j];
  285. const uint8_t *band_sizes = ctx->bands[j];
  286. float line_to_frequency = ctx->avctx->sample_rate / (j ? 256.f : 2048.0f);
  287. float avg_chan_bits = chan_bitrate * (j ? 128.0f : 1024.0f) / ctx->avctx->sample_rate;
  288. /* reference encoder uses 2.4% here instead of 60% like the spec says */
  289. float bark_pe = 0.024f * PSY_3GPP_BITS_TO_PE(avg_chan_bits) / num_bark;
  290. float en_spread_low = j ? PSY_3GPP_EN_SPREAD_LOW_S : PSY_3GPP_EN_SPREAD_LOW_L;
  291. /* High energy spreading for long blocks <= 22kbps/channel and short blocks are the same. */
  292. float en_spread_hi = (j || (chan_bitrate <= 22.0f)) ? PSY_3GPP_EN_SPREAD_HI_S : PSY_3GPP_EN_SPREAD_HI_L1;
  293. i = 0;
  294. prev = 0.0;
  295. for (g = 0; g < ctx->num_bands[j]; g++) {
  296. i += band_sizes[g];
  297. bark = calc_bark((i-1) * line_to_frequency);
  298. coeffs[g].barks = (bark + prev) / 2.0;
  299. prev = bark;
  300. }
  301. for (g = 0; g < ctx->num_bands[j] - 1; g++) {
  302. AacPsyCoeffs *coeff = &coeffs[g];
  303. float bark_width = coeffs[g+1].barks - coeffs->barks;
  304. coeff->spread_low[0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_LOW);
  305. coeff->spread_hi [0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_HI);
  306. coeff->spread_low[1] = pow(10.0, -bark_width * en_spread_low);
  307. coeff->spread_hi [1] = pow(10.0, -bark_width * en_spread_hi);
  308. pe_min = bark_pe * bark_width;
  309. minsnr = exp2(pe_min / band_sizes[g]) - 1.5f;
  310. coeff->min_snr = av_clipf(1.0f / minsnr, PSY_SNR_25DB, PSY_SNR_1DB);
  311. }
  312. start = 0;
  313. for (g = 0; g < ctx->num_bands[j]; g++) {
  314. minscale = ath(start * line_to_frequency, ATH_ADD);
  315. for (i = 1; i < band_sizes[g]; i++)
  316. minscale = FFMIN(minscale, ath((start + i) * line_to_frequency, ATH_ADD));
  317. coeffs[g].ath = minscale - minath;
  318. start += band_sizes[g];
  319. }
  320. }
  321. pctx->ch = av_mallocz_array(ctx->avctx->channels, sizeof(AacPsyChannel));
  322. if (!pctx->ch) {
  323. av_freep(&ctx->model_priv_data);
  324. return AVERROR(ENOMEM);
  325. }
  326. lame_window_init(pctx, ctx->avctx);
  327. return 0;
  328. }
  329. /**
  330. * IIR filter used in block switching decision
  331. */
  332. static float iir_filter(int in, float state[2])
  333. {
  334. float ret;
  335. ret = 0.7548f * (in - state[0]) + 0.5095f * state[1];
  336. state[0] = in;
  337. state[1] = ret;
  338. return ret;
  339. }
  340. /**
  341. * window grouping information stored as bits (0 - new group, 1 - group continues)
  342. */
  343. static const uint8_t window_grouping[9] = {
  344. 0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36
  345. };
  346. /**
  347. * Tell encoder which window types to use.
  348. * @see 3GPP TS26.403 5.4.1 "Blockswitching"
  349. */
  350. static av_unused FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx,
  351. const int16_t *audio,
  352. const int16_t *la,
  353. int channel, int prev_type)
  354. {
  355. int i, j;
  356. int br = ctx->avctx->bit_rate / ctx->avctx->channels;
  357. int attack_ratio = br <= 16000 ? 18 : 10;
  358. AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
  359. AacPsyChannel *pch = &pctx->ch[channel];
  360. uint8_t grouping = 0;
  361. int next_type = pch->next_window_seq;
  362. FFPsyWindowInfo wi = { { 0 } };
  363. if (la) {
  364. float s[8], v;
  365. int switch_to_eight = 0;
  366. float sum = 0.0, sum2 = 0.0;
  367. int attack_n = 0;
  368. int stay_short = 0;
  369. for (i = 0; i < 8; i++) {
  370. for (j = 0; j < 128; j++) {
  371. v = iir_filter(la[i*128+j], pch->iir_state);
  372. sum += v*v;
  373. }
  374. s[i] = sum;
  375. sum2 += sum;
  376. }
  377. for (i = 0; i < 8; i++) {
  378. if (s[i] > pch->win_energy * attack_ratio) {
  379. attack_n = i + 1;
  380. switch_to_eight = 1;
  381. break;
  382. }
  383. }
  384. pch->win_energy = pch->win_energy*7/8 + sum2/64;
  385. wi.window_type[1] = prev_type;
  386. switch (prev_type) {
  387. case ONLY_LONG_SEQUENCE:
  388. wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
  389. next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
  390. break;
  391. case LONG_START_SEQUENCE:
  392. wi.window_type[0] = EIGHT_SHORT_SEQUENCE;
  393. grouping = pch->next_grouping;
  394. next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
  395. break;
  396. case LONG_STOP_SEQUENCE:
  397. wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
  398. next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
  399. break;
  400. case EIGHT_SHORT_SEQUENCE:
  401. stay_short = next_type == EIGHT_SHORT_SEQUENCE || switch_to_eight;
  402. wi.window_type[0] = stay_short ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
  403. grouping = next_type == EIGHT_SHORT_SEQUENCE ? pch->next_grouping : 0;
  404. next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
  405. break;
  406. }
  407. pch->next_grouping = window_grouping[attack_n];
  408. pch->next_window_seq = next_type;
  409. } else {
  410. for (i = 0; i < 3; i++)
  411. wi.window_type[i] = prev_type;
  412. grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0;
  413. }
  414. wi.window_shape = 1;
  415. if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
  416. wi.num_windows = 1;
  417. wi.grouping[0] = 1;
  418. } else {
  419. int lastgrp = 0;
  420. wi.num_windows = 8;
  421. for (i = 0; i < 8; i++) {
  422. if (!((grouping >> i) & 1))
  423. lastgrp = i;
  424. wi.grouping[lastgrp]++;
  425. }
  426. }
  427. return wi;
  428. }
  429. /* 5.6.1.2 "Calculation of Bit Demand" */
  430. static int calc_bit_demand(AacPsyContext *ctx, float pe, int bits, int size,
  431. int short_window)
  432. {
  433. const float bitsave_slope = short_window ? PSY_3GPP_SAVE_SLOPE_S : PSY_3GPP_SAVE_SLOPE_L;
  434. const float bitsave_add = short_window ? PSY_3GPP_SAVE_ADD_S : PSY_3GPP_SAVE_ADD_L;
  435. const float bitspend_slope = short_window ? PSY_3GPP_SPEND_SLOPE_S : PSY_3GPP_SPEND_SLOPE_L;
  436. const float bitspend_add = short_window ? PSY_3GPP_SPEND_ADD_S : PSY_3GPP_SPEND_ADD_L;
  437. const float clip_low = short_window ? PSY_3GPP_CLIP_LO_S : PSY_3GPP_CLIP_LO_L;
  438. const float clip_high = short_window ? PSY_3GPP_CLIP_HI_S : PSY_3GPP_CLIP_HI_L;
  439. float clipped_pe, bit_save, bit_spend, bit_factor, fill_level;
  440. ctx->fill_level += ctx->frame_bits - bits;
  441. ctx->fill_level = av_clip(ctx->fill_level, 0, size);
  442. fill_level = av_clipf((float)ctx->fill_level / size, clip_low, clip_high);
  443. clipped_pe = av_clipf(pe, ctx->pe.min, ctx->pe.max);
  444. bit_save = (fill_level + bitsave_add) * bitsave_slope;
  445. assert(bit_save <= 0.3f && bit_save >= -0.05000001f);
  446. bit_spend = (fill_level + bitspend_add) * bitspend_slope;
  447. assert(bit_spend <= 0.5f && bit_spend >= -0.1f);
  448. /* The bit factor graph in the spec is obviously incorrect.
  449. * bit_spend + ((bit_spend - bit_spend))...
  450. * The reference encoder subtracts everything from 1, but also seems incorrect.
  451. * 1 - bit_save + ((bit_spend + bit_save))...
  452. * Hopefully below is correct.
  453. */
  454. bit_factor = 1.0f - bit_save + ((bit_spend - bit_save) / (ctx->pe.max - ctx->pe.min)) * (clipped_pe - ctx->pe.min);
  455. /* NOTE: The reference encoder attempts to center pe max/min around the current pe. */
  456. ctx->pe.max = FFMAX(pe, ctx->pe.max);
  457. ctx->pe.min = FFMIN(pe, ctx->pe.min);
  458. return FFMIN(ctx->frame_bits * bit_factor, ctx->frame_bits + size - bits);
  459. }
  460. static float calc_pe_3gpp(AacPsyBand *band)
  461. {
  462. float pe, a;
  463. band->pe = 0.0f;
  464. band->pe_const = 0.0f;
  465. band->active_lines = 0.0f;
  466. if (band->energy > band->thr) {
  467. a = log2f(band->energy);
  468. pe = a - log2f(band->thr);
  469. band->active_lines = band->nz_lines;
  470. if (pe < PSY_3GPP_C1) {
  471. pe = pe * PSY_3GPP_C3 + PSY_3GPP_C2;
  472. a = a * PSY_3GPP_C3 + PSY_3GPP_C2;
  473. band->active_lines *= PSY_3GPP_C3;
  474. }
  475. band->pe = pe * band->nz_lines;
  476. band->pe_const = a * band->nz_lines;
  477. }
  478. return band->pe;
  479. }
  480. static float calc_reduction_3gpp(float a, float desired_pe, float pe,
  481. float active_lines)
  482. {
  483. float thr_avg, reduction;
  484. if(active_lines == 0.0)
  485. return 0;
  486. thr_avg = exp2f((a - pe) / (4.0f * active_lines));
  487. reduction = exp2f((a - desired_pe) / (4.0f * active_lines)) - thr_avg;
  488. return FFMAX(reduction, 0.0f);
  489. }
  490. static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr,
  491. float reduction)
  492. {
  493. float thr = band->thr;
  494. if (band->energy > thr) {
  495. thr = sqrtf(thr);
  496. thr = sqrtf(thr) + reduction;
  497. thr *= thr;
  498. thr *= thr;
  499. /* This deviates from the 3GPP spec to match the reference encoder.
  500. * It performs min(thr_reduced, max(thr, energy/min_snr)) only for bands
  501. * that have hole avoidance on (active or inactive). It always reduces the
  502. * threshold of bands with hole avoidance off.
  503. */
  504. if (thr > band->energy * min_snr && band->avoid_holes != PSY_3GPP_AH_NONE) {
  505. thr = FFMAX(band->thr, band->energy * min_snr);
  506. band->avoid_holes = PSY_3GPP_AH_ACTIVE;
  507. }
  508. }
  509. return thr;
  510. }
  511. #ifndef calc_thr_3gpp
  512. static void calc_thr_3gpp(const FFPsyWindowInfo *wi, const int num_bands, AacPsyChannel *pch,
  513. const uint8_t *band_sizes, const float *coefs)
  514. {
  515. int i, w, g;
  516. int start = 0;
  517. for (w = 0; w < wi->num_windows*16; w += 16) {
  518. for (g = 0; g < num_bands; g++) {
  519. AacPsyBand *band = &pch->band[w+g];
  520. float form_factor = 0.0f;
  521. float Temp;
  522. band->energy = 0.0f;
  523. for (i = 0; i < band_sizes[g]; i++) {
  524. band->energy += coefs[start+i] * coefs[start+i];
  525. form_factor += sqrtf(fabs(coefs[start+i]));
  526. }
  527. Temp = band->energy > 0 ? sqrtf((float)band_sizes[g] / band->energy) : 0;
  528. band->thr = band->energy * 0.001258925f;
  529. band->nz_lines = form_factor * sqrtf(Temp);
  530. start += band_sizes[g];
  531. }
  532. }
  533. }
  534. #endif /* calc_thr_3gpp */
  535. #ifndef psy_hp_filter
  536. static void psy_hp_filter(const float *firbuf, float *hpfsmpl, const float *psy_fir_coeffs)
  537. {
  538. int i, j;
  539. for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) {
  540. float sum1, sum2;
  541. sum1 = firbuf[i + (PSY_LAME_FIR_LEN - 1) / 2];
  542. sum2 = 0.0;
  543. for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) {
  544. sum1 += psy_fir_coeffs[j] * (firbuf[i + j] + firbuf[i + PSY_LAME_FIR_LEN - j]);
  545. sum2 += psy_fir_coeffs[j + 1] * (firbuf[i + j + 1] + firbuf[i + PSY_LAME_FIR_LEN - j - 1]);
  546. }
  547. /* NOTE: The LAME psymodel expects it's input in the range -32768 to 32768.
  548. * Tuning this for normalized floats would be difficult. */
  549. hpfsmpl[i] = (sum1 + sum2) * 32768.0f;
  550. }
  551. }
  552. #endif /* psy_hp_filter */
  553. /**
  554. * Calculate band thresholds as suggested in 3GPP TS26.403
  555. */
  556. static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel,
  557. const float *coefs, const FFPsyWindowInfo *wi)
  558. {
  559. AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
  560. AacPsyChannel *pch = &pctx->ch[channel];
  561. int i, w, g;
  562. float desired_bits, desired_pe, delta_pe, reduction= NAN, spread_en[128] = {0};
  563. float a = 0.0f, active_lines = 0.0f, norm_fac = 0.0f;
  564. float pe = pctx->chan_bitrate > 32000 ? 0.0f : FFMAX(50.0f, 100.0f - pctx->chan_bitrate * 100.0f / 32000.0f);
  565. const int num_bands = ctx->num_bands[wi->num_windows == 8];
  566. const uint8_t *band_sizes = ctx->bands[wi->num_windows == 8];
  567. AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8];
  568. const float avoid_hole_thr = wi->num_windows == 8 ? PSY_3GPP_AH_THR_SHORT : PSY_3GPP_AH_THR_LONG;
  569. //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation"
  570. calc_thr_3gpp(wi, num_bands, pch, band_sizes, coefs);
  571. //modify thresholds and energies - spread, threshold in quiet, pre-echo control
  572. for (w = 0; w < wi->num_windows*16; w += 16) {
  573. AacPsyBand *bands = &pch->band[w];
  574. /* 5.4.2.3 "Spreading" & 5.4.3 "Spread Energy Calculation" */
  575. spread_en[0] = bands[0].energy;
  576. for (g = 1; g < num_bands; g++) {
  577. bands[g].thr = FFMAX(bands[g].thr, bands[g-1].thr * coeffs[g].spread_hi[0]);
  578. spread_en[w+g] = FFMAX(bands[g].energy, spread_en[w+g-1] * coeffs[g].spread_hi[1]);
  579. }
  580. for (g = num_bands - 2; g >= 0; g--) {
  581. bands[g].thr = FFMAX(bands[g].thr, bands[g+1].thr * coeffs[g].spread_low[0]);
  582. spread_en[w+g] = FFMAX(spread_en[w+g], spread_en[w+g+1] * coeffs[g].spread_low[1]);
  583. }
  584. //5.4.2.4 "Threshold in quiet"
  585. for (g = 0; g < num_bands; g++) {
  586. AacPsyBand *band = &bands[g];
  587. band->thr_quiet = band->thr = FFMAX(band->thr, coeffs[g].ath);
  588. //5.4.2.5 "Pre-echo control"
  589. if (!(wi->window_type[0] == LONG_STOP_SEQUENCE || (wi->window_type[1] == LONG_START_SEQUENCE && !w)))
  590. band->thr = FFMAX(PSY_3GPP_RPEMIN*band->thr, FFMIN(band->thr,
  591. PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet));
  592. /* 5.6.1.3.1 "Preparatory steps of the perceptual entropy calculation" */
  593. pe += calc_pe_3gpp(band);
  594. a += band->pe_const;
  595. active_lines += band->active_lines;
  596. /* 5.6.1.3.3 "Selection of the bands for avoidance of holes" */
  597. if (spread_en[w+g] * avoid_hole_thr > band->energy || coeffs[g].min_snr > 1.0f)
  598. band->avoid_holes = PSY_3GPP_AH_NONE;
  599. else
  600. band->avoid_holes = PSY_3GPP_AH_INACTIVE;
  601. }
  602. }
  603. /* 5.6.1.3.2 "Calculation of the desired perceptual entropy" */
  604. ctx->ch[channel].entropy = pe;
  605. desired_bits = calc_bit_demand(pctx, pe, ctx->bitres.bits, ctx->bitres.size, wi->num_windows == 8);
  606. desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits);
  607. /* NOTE: PE correction is kept simple. During initial testing it had very
  608. * little effect on the final bitrate. Probably a good idea to come
  609. * back and do more testing later.
  610. */
  611. if (ctx->bitres.bits > 0)
  612. desired_pe *= av_clipf(pctx->pe.previous / PSY_3GPP_BITS_TO_PE(ctx->bitres.bits),
  613. 0.85f, 1.15f);
  614. pctx->pe.previous = PSY_3GPP_BITS_TO_PE(desired_bits);
  615. if (desired_pe < pe) {
  616. /* 5.6.1.3.4 "First Estimation of the reduction value" */
  617. for (w = 0; w < wi->num_windows*16; w += 16) {
  618. reduction = calc_reduction_3gpp(a, desired_pe, pe, active_lines);
  619. pe = 0.0f;
  620. a = 0.0f;
  621. active_lines = 0.0f;
  622. for (g = 0; g < num_bands; g++) {
  623. AacPsyBand *band = &pch->band[w+g];
  624. band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
  625. /* recalculate PE */
  626. pe += calc_pe_3gpp(band);
  627. a += band->pe_const;
  628. active_lines += band->active_lines;
  629. }
  630. }
  631. /* 5.6.1.3.5 "Second Estimation of the reduction value" */
  632. for (i = 0; i < 2; i++) {
  633. float pe_no_ah = 0.0f, desired_pe_no_ah;
  634. active_lines = a = 0.0f;
  635. for (w = 0; w < wi->num_windows*16; w += 16) {
  636. for (g = 0; g < num_bands; g++) {
  637. AacPsyBand *band = &pch->band[w+g];
  638. if (band->avoid_holes != PSY_3GPP_AH_ACTIVE) {
  639. pe_no_ah += band->pe;
  640. a += band->pe_const;
  641. active_lines += band->active_lines;
  642. }
  643. }
  644. }
  645. desired_pe_no_ah = FFMAX(desired_pe - (pe - pe_no_ah), 0.0f);
  646. if (active_lines > 0.0f)
  647. reduction = calc_reduction_3gpp(a, desired_pe_no_ah, pe_no_ah, active_lines);
  648. pe = 0.0f;
  649. for (w = 0; w < wi->num_windows*16; w += 16) {
  650. for (g = 0; g < num_bands; g++) {
  651. AacPsyBand *band = &pch->band[w+g];
  652. if (active_lines > 0.0f)
  653. band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
  654. pe += calc_pe_3gpp(band);
  655. if (band->thr > 0.0f)
  656. band->norm_fac = band->active_lines / band->thr;
  657. else
  658. band->norm_fac = 0.0f;
  659. norm_fac += band->norm_fac;
  660. }
  661. }
  662. delta_pe = desired_pe - pe;
  663. if (fabs(delta_pe) > 0.05f * desired_pe)
  664. break;
  665. }
  666. if (pe < 1.15f * desired_pe) {
  667. /* 6.6.1.3.6 "Final threshold modification by linearization" */
  668. norm_fac = 1.0f / norm_fac;
  669. for (w = 0; w < wi->num_windows*16; w += 16) {
  670. for (g = 0; g < num_bands; g++) {
  671. AacPsyBand *band = &pch->band[w+g];
  672. if (band->active_lines > 0.5f) {
  673. float delta_sfb_pe = band->norm_fac * norm_fac * delta_pe;
  674. float thr = band->thr;
  675. thr *= exp2f(delta_sfb_pe / band->active_lines);
  676. if (thr > coeffs[g].min_snr * band->energy && band->avoid_holes == PSY_3GPP_AH_INACTIVE)
  677. thr = FFMAX(band->thr, coeffs[g].min_snr * band->energy);
  678. band->thr = thr;
  679. }
  680. }
  681. }
  682. } else {
  683. /* 5.6.1.3.7 "Further perceptual entropy reduction" */
  684. g = num_bands;
  685. while (pe > desired_pe && g--) {
  686. for (w = 0; w < wi->num_windows*16; w+= 16) {
  687. AacPsyBand *band = &pch->band[w+g];
  688. if (band->avoid_holes != PSY_3GPP_AH_NONE && coeffs[g].min_snr < PSY_SNR_1DB) {
  689. coeffs[g].min_snr = PSY_SNR_1DB;
  690. band->thr = band->energy * PSY_SNR_1DB;
  691. pe += band->active_lines * 1.5f - band->pe;
  692. }
  693. }
  694. }
  695. /* TODO: allow more holes (unused without mid/side) */
  696. }
  697. }
  698. for (w = 0; w < wi->num_windows*16; w += 16) {
  699. for (g = 0; g < num_bands; g++) {
  700. AacPsyBand *band = &pch->band[w+g];
  701. FFPsyBand *psy_band = &ctx->ch[channel].psy_bands[w+g];
  702. psy_band->threshold = band->thr;
  703. psy_band->energy = band->energy;
  704. psy_band->spread = band->active_lines * 2.0f / band_sizes[g];
  705. }
  706. }
  707. memcpy(pch->prev_band, pch->band, sizeof(pch->band));
  708. }
  709. static void psy_3gpp_analyze(FFPsyContext *ctx, int channel,
  710. const float **coeffs, const FFPsyWindowInfo *wi)
  711. {
  712. int ch;
  713. FFPsyChannelGroup *group = ff_psy_find_group(ctx, channel);
  714. for (ch = 0; ch < group->num_ch; ch++)
  715. psy_3gpp_analyze_channel(ctx, channel + ch, coeffs[ch], &wi[ch]);
  716. }
  717. static av_cold void psy_3gpp_end(FFPsyContext *apc)
  718. {
  719. AacPsyContext *pctx = (AacPsyContext*) apc->model_priv_data;
  720. av_freep(&pctx->ch);
  721. av_freep(&apc->model_priv_data);
  722. }
  723. static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock)
  724. {
  725. int blocktype = ONLY_LONG_SEQUENCE;
  726. if (uselongblock) {
  727. if (ctx->next_window_seq == EIGHT_SHORT_SEQUENCE)
  728. blocktype = LONG_STOP_SEQUENCE;
  729. } else {
  730. blocktype = EIGHT_SHORT_SEQUENCE;
  731. if (ctx->next_window_seq == ONLY_LONG_SEQUENCE)
  732. ctx->next_window_seq = LONG_START_SEQUENCE;
  733. if (ctx->next_window_seq == LONG_STOP_SEQUENCE)
  734. ctx->next_window_seq = EIGHT_SHORT_SEQUENCE;
  735. }
  736. wi->window_type[0] = ctx->next_window_seq;
  737. ctx->next_window_seq = blocktype;
  738. }
  739. static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx, const float *audio,
  740. const float *la, int channel, int prev_type)
  741. {
  742. AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
  743. AacPsyChannel *pch = &pctx->ch[channel];
  744. int grouping = 0;
  745. int uselongblock = 1;
  746. int attacks[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
  747. float clippings[AAC_NUM_BLOCKS_SHORT];
  748. int i;
  749. FFPsyWindowInfo wi = { { 0 } };
  750. if (la) {
  751. float hpfsmpl[AAC_BLOCK_SIZE_LONG];
  752. float const *pf = hpfsmpl;
  753. float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
  754. float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
  755. float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
  756. const float *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN);
  757. int att_sum = 0;
  758. /* LAME comment: apply high pass filter of fs/4 */
  759. psy_hp_filter(firbuf, hpfsmpl, psy_fir_coeffs);
  760. /* Calculate the energies of each sub-shortblock */
  761. for (i = 0; i < PSY_LAME_NUM_SUBBLOCKS; i++) {
  762. energy_subshort[i] = pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS)];
  763. assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)] > 0);
  764. attack_intensity[i] = energy_subshort[i] / pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)];
  765. energy_short[0] += energy_subshort[i];
  766. }
  767. for (i = 0; i < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; i++) {
  768. float const *const pfe = pf + AAC_BLOCK_SIZE_LONG / (AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS);
  769. float p = 1.0f;
  770. for (; pf < pfe; pf++)
  771. p = FFMAX(p, fabsf(*pf));
  772. pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p;
  773. energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p;
  774. /* NOTE: The indexes below are [i + 3 - 2] in the LAME source.
  775. * Obviously the 3 and 2 have some significance, or this would be just [i + 1]
  776. * (which is what we use here). What the 3 stands for is ambiguous, as it is both
  777. * number of short blocks, and the number of sub-short blocks.
  778. * It seems that LAME is comparing each sub-block to sub-block + 1 in the
  779. * previous block.
  780. */
  781. if (p > energy_subshort[i + 1])
  782. p = p / energy_subshort[i + 1];
  783. else if (energy_subshort[i + 1] > p * 10.0f)
  784. p = energy_subshort[i + 1] / (p * 10.0f);
  785. else
  786. p = 0.0;
  787. attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p;
  788. }
  789. /* compare energy between sub-short blocks */
  790. for (i = 0; i < (AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS; i++)
  791. if (!attacks[i / PSY_LAME_NUM_SUBBLOCKS])
  792. if (attack_intensity[i] > pch->attack_threshold)
  793. attacks[i / PSY_LAME_NUM_SUBBLOCKS] = (i % PSY_LAME_NUM_SUBBLOCKS) + 1;
  794. /* should have energy change between short blocks, in order to avoid periodic signals */
  795. /* Good samples to show the effect are Trumpet test songs */
  796. /* GB: tuned (1) to avoid too many short blocks for test sample TRUMPET */
  797. /* RH: tuned (2) to let enough short blocks through for test sample FSOL and SNAPS */
  798. for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) {
  799. float const u = energy_short[i - 1];
  800. float const v = energy_short[i];
  801. float const m = FFMAX(u, v);
  802. if (m < 40000) { /* (2) */
  803. if (u < 1.7f * v && v < 1.7f * u) { /* (1) */
  804. if (i == 1 && attacks[0] < attacks[i])
  805. attacks[0] = 0;
  806. attacks[i] = 0;
  807. }
  808. }
  809. att_sum += attacks[i];
  810. }
  811. if (attacks[0] <= pch->prev_attack)
  812. attacks[0] = 0;
  813. att_sum += attacks[0];
  814. /* 3 below indicates the previous attack happened in the last sub-block of the previous sequence */
  815. if (pch->prev_attack == 3 || att_sum) {
  816. uselongblock = 0;
  817. for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++)
  818. if (attacks[i] && attacks[i-1])
  819. attacks[i] = 0;
  820. }
  821. } else {
  822. /* We have no lookahead info, so just use same type as the previous sequence. */
  823. uselongblock = !(prev_type == EIGHT_SHORT_SEQUENCE);
  824. }
  825. lame_apply_block_type(pch, &wi, uselongblock);
  826. /* Calculate input sample maximums and evaluate clipping risk */
  827. if (audio) {
  828. for (i = 0; i < AAC_NUM_BLOCKS_SHORT; i++) {
  829. const float *wbuf = audio + i * AAC_BLOCK_SIZE_SHORT;
  830. float max = 0;
  831. int j;
  832. for (j = 0; j < AAC_BLOCK_SIZE_SHORT; j++)
  833. max = FFMAX(max, fabsf(wbuf[j]));
  834. clippings[i] = max;
  835. }
  836. } else {
  837. for (i = 0; i < 8; i++)
  838. clippings[i] = 0;
  839. }
  840. wi.window_type[1] = prev_type;
  841. if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
  842. float clipping = 0.0f;
  843. wi.num_windows = 1;
  844. wi.grouping[0] = 1;
  845. if (wi.window_type[0] == LONG_START_SEQUENCE)
  846. wi.window_shape = 0;
  847. else
  848. wi.window_shape = 1;
  849. for (i = 0; i < 8; i++)
  850. clipping = FFMAX(clipping, clippings[i]);
  851. wi.clipping[0] = clipping;
  852. } else {
  853. int lastgrp = 0;
  854. wi.num_windows = 8;
  855. wi.window_shape = 0;
  856. for (i = 0; i < 8; i++) {
  857. if (!((pch->next_grouping >> i) & 1))
  858. lastgrp = i;
  859. wi.grouping[lastgrp]++;
  860. }
  861. for (i = 0; i < 8; i += wi.grouping[i]) {
  862. int w;
  863. float clipping = 0.0f;
  864. for (w = 0; w < wi.grouping[i] && !clipping; w++)
  865. clipping = FFMAX(clipping, clippings[i+w]);
  866. wi.clipping[i] = clipping;
  867. }
  868. }
  869. /* Determine grouping, based on the location of the first attack, and save for
  870. * the next frame.
  871. * FIXME: Move this to analysis.
  872. * TODO: Tune groupings depending on attack location
  873. * TODO: Handle more than one attack in a group
  874. */
  875. for (i = 0; i < 9; i++) {
  876. if (attacks[i]) {
  877. grouping = i;
  878. break;
  879. }
  880. }
  881. pch->next_grouping = window_grouping[grouping];
  882. pch->prev_attack = attacks[8];
  883. return wi;
  884. }
  885. const FFPsyModel ff_aac_psy_model =
  886. {
  887. .name = "3GPP TS 26.403-inspired model",
  888. .init = psy_3gpp_init,
  889. .window = psy_lame_window,
  890. .analyze = psy_3gpp_analyze,
  891. .end = psy_3gpp_end,
  892. };