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.

1009 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. #define PSY_3GPP_PE_TO_BITS(bits) ((bits) / 1.18f)
  79. /* LAME psy model constants */
  80. #define PSY_LAME_FIR_LEN 21 ///< LAME psy model FIR order
  81. #define AAC_BLOCK_SIZE_LONG 1024 ///< long block size
  82. #define AAC_BLOCK_SIZE_SHORT 128 ///< short block size
  83. #define AAC_NUM_BLOCKS_SHORT 8 ///< number of blocks in a short sequence
  84. #define PSY_LAME_NUM_SUBBLOCKS 3 ///< Number of sub-blocks in each short block
  85. /**
  86. * @}
  87. */
  88. /**
  89. * information for single band used by 3GPP TS26.403-inspired psychoacoustic model
  90. */
  91. typedef struct AacPsyBand{
  92. float energy; ///< band energy
  93. float thr; ///< energy threshold
  94. float thr_quiet; ///< threshold in quiet
  95. float nz_lines; ///< number of non-zero spectral lines
  96. float active_lines; ///< number of active spectral lines
  97. float pe; ///< perceptual entropy
  98. float pe_const; ///< constant part of the PE calculation
  99. float norm_fac; ///< normalization factor for linearization
  100. int avoid_holes; ///< hole avoidance flag
  101. }AacPsyBand;
  102. /**
  103. * single/pair channel context for psychoacoustic model
  104. */
  105. typedef struct AacPsyChannel{
  106. AacPsyBand band[128]; ///< bands information
  107. AacPsyBand prev_band[128]; ///< bands information from the previous frame
  108. float win_energy; ///< sliding average of channel energy
  109. float iir_state[2]; ///< hi-pass IIR filter state
  110. uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence)
  111. enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame
  112. /* LAME psy model specific members */
  113. float attack_threshold; ///< attack threshold for this channel
  114. float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS];
  115. int prev_attack; ///< attack value for the last short block in the previous sequence
  116. }AacPsyChannel;
  117. /**
  118. * psychoacoustic model frame type-dependent coefficients
  119. */
  120. typedef struct AacPsyCoeffs{
  121. float ath; ///< absolute threshold of hearing per bands
  122. float barks; ///< Bark value for each spectral band in long frame
  123. float spread_low[2]; ///< spreading factor for low-to-high threshold spreading in long frame
  124. float spread_hi [2]; ///< spreading factor for high-to-low threshold spreading in long frame
  125. float min_snr; ///< minimal SNR
  126. }AacPsyCoeffs;
  127. /**
  128. * 3GPP TS26.403-inspired psychoacoustic model specific data
  129. */
  130. typedef struct AacPsyContext{
  131. int chan_bitrate; ///< bitrate per channel
  132. int frame_bits; ///< average bits per frame
  133. int fill_level; ///< bit reservoir fill level
  134. struct {
  135. float min; ///< minimum allowed PE for bit factor calculation
  136. float max; ///< maximum allowed PE for bit factor calculation
  137. float previous; ///< allowed PE of the previous frame
  138. float correction; ///< PE correction factor
  139. } pe;
  140. AacPsyCoeffs psy_coef[2][64];
  141. AacPsyChannel *ch;
  142. }AacPsyContext;
  143. /**
  144. * LAME psy model preset struct
  145. */
  146. typedef struct PsyLamePreset {
  147. int quality; ///< Quality to map the rest of the vaules to.
  148. /* This is overloaded to be both kbps per channel in ABR mode, and
  149. * requested quality in constant quality mode.
  150. */
  151. float st_lrm; ///< short threshold for L, R, and M channels
  152. } PsyLamePreset;
  153. /**
  154. * LAME psy model preset table for ABR
  155. */
  156. static const PsyLamePreset psy_abr_map[] = {
  157. /* TODO: Tuning. These were taken from LAME. */
  158. /* kbps/ch st_lrm */
  159. { 8, 6.60},
  160. { 16, 6.60},
  161. { 24, 6.60},
  162. { 32, 6.60},
  163. { 40, 6.60},
  164. { 48, 6.60},
  165. { 56, 6.60},
  166. { 64, 6.40},
  167. { 80, 6.00},
  168. { 96, 5.60},
  169. {112, 5.20},
  170. {128, 5.20},
  171. {160, 5.20}
  172. };
  173. /**
  174. * LAME psy model preset table for constant quality
  175. */
  176. static const PsyLamePreset psy_vbr_map[] = {
  177. /* vbr_q st_lrm */
  178. { 0, 4.20},
  179. { 1, 4.20},
  180. { 2, 4.20},
  181. { 3, 4.20},
  182. { 4, 4.20},
  183. { 5, 4.20},
  184. { 6, 4.20},
  185. { 7, 4.20},
  186. { 8, 4.20},
  187. { 9, 4.20},
  188. {10, 4.20}
  189. };
  190. /**
  191. * LAME psy model FIR coefficient table
  192. */
  193. static const float psy_fir_coeffs[] = {
  194. -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2,
  195. -3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2,
  196. -5.52212e-17 * 2, -0.313819 * 2
  197. };
  198. #if ARCH_MIPS
  199. # include "mips/aacpsy_mips.h"
  200. #endif /* ARCH_MIPS */
  201. /**
  202. * Calculate the ABR attack threshold from the above LAME psymodel table.
  203. */
  204. static float lame_calc_attack_threshold(int bitrate)
  205. {
  206. /* Assume max bitrate to start with */
  207. int lower_range = 12, upper_range = 12;
  208. int lower_range_kbps = psy_abr_map[12].quality;
  209. int upper_range_kbps = psy_abr_map[12].quality;
  210. int i;
  211. /* Determine which bitrates the value specified falls between.
  212. * If the loop ends without breaking our above assumption of 320kbps was correct.
  213. */
  214. for (i = 1; i < 13; i++) {
  215. if (FFMAX(bitrate, psy_abr_map[i].quality) != bitrate) {
  216. upper_range = i;
  217. upper_range_kbps = psy_abr_map[i ].quality;
  218. lower_range = i - 1;
  219. lower_range_kbps = psy_abr_map[i - 1].quality;
  220. break; /* Upper range found */
  221. }
  222. }
  223. /* Determine which range the value specified is closer to */
  224. if ((upper_range_kbps - bitrate) > (bitrate - lower_range_kbps))
  225. return psy_abr_map[lower_range].st_lrm;
  226. return psy_abr_map[upper_range].st_lrm;
  227. }
  228. /**
  229. * LAME psy model specific initialization
  230. */
  231. static av_cold void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx)
  232. {
  233. int i, j;
  234. for (i = 0; i < avctx->channels; i++) {
  235. AacPsyChannel *pch = &ctx->ch[i];
  236. if (avctx->flags & AV_CODEC_FLAG_QSCALE)
  237. pch->attack_threshold = psy_vbr_map[avctx->global_quality / FF_QP2LAMBDA].st_lrm;
  238. else
  239. pch->attack_threshold = lame_calc_attack_threshold(avctx->bit_rate / avctx->channels / 1000);
  240. for (j = 0; j < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; j++)
  241. pch->prev_energy_subshort[j] = 10.0f;
  242. }
  243. }
  244. /**
  245. * Calculate Bark value for given line.
  246. */
  247. static av_cold float calc_bark(float f)
  248. {
  249. return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f));
  250. }
  251. #define ATH_ADD 4
  252. /**
  253. * Calculate ATH value for given frequency.
  254. * Borrowed from Lame.
  255. */
  256. static av_cold float ath(float f, float add)
  257. {
  258. f /= 1000.0f;
  259. return 3.64 * pow(f, -0.8)
  260. - 6.8 * exp(-0.6 * (f - 3.4) * (f - 3.4))
  261. + 6.0 * exp(-0.15 * (f - 8.7) * (f - 8.7))
  262. + (0.6 + 0.04 * add) * 0.001 * f * f * f * f;
  263. }
  264. static av_cold int psy_3gpp_init(FFPsyContext *ctx) {
  265. AacPsyContext *pctx;
  266. float bark;
  267. int i, j, g, start;
  268. float prev, minscale, minath, minsnr, pe_min;
  269. const int chan_bitrate = ctx->avctx->bit_rate / ctx->avctx->channels;
  270. const int bandwidth = ctx->avctx->cutoff ? ctx->avctx->cutoff : AAC_CUTOFF(ctx->avctx);
  271. const float num_bark = calc_bark((float)bandwidth);
  272. ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext));
  273. if (!ctx->model_priv_data)
  274. return AVERROR(ENOMEM);
  275. pctx = (AacPsyContext*) ctx->model_priv_data;
  276. pctx->chan_bitrate = chan_bitrate;
  277. pctx->frame_bits = chan_bitrate * AAC_BLOCK_SIZE_LONG / ctx->avctx->sample_rate;
  278. pctx->pe.min = 8.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
  279. pctx->pe.max = 12.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
  280. ctx->bitres.size = 6144 - pctx->frame_bits;
  281. ctx->bitres.size -= ctx->bitres.size % 8;
  282. pctx->fill_level = ctx->bitres.size;
  283. minath = ath(3410 - 0.733 * ATH_ADD, ATH_ADD);
  284. for (j = 0; j < 2; j++) {
  285. AacPsyCoeffs *coeffs = pctx->psy_coef[j];
  286. const uint8_t *band_sizes = ctx->bands[j];
  287. float line_to_frequency = ctx->avctx->sample_rate / (j ? 256.f : 2048.0f);
  288. float avg_chan_bits = chan_bitrate * (j ? 128.0f : 1024.0f) / ctx->avctx->sample_rate;
  289. /* reference encoder uses 2.4% here instead of 60% like the spec says */
  290. float bark_pe = 0.024f * PSY_3GPP_BITS_TO_PE(avg_chan_bits) / num_bark;
  291. float en_spread_low = j ? PSY_3GPP_EN_SPREAD_LOW_S : PSY_3GPP_EN_SPREAD_LOW_L;
  292. /* High energy spreading for long blocks <= 22kbps/channel and short blocks are the same. */
  293. float en_spread_hi = (j || (chan_bitrate <= 22.0f)) ? PSY_3GPP_EN_SPREAD_HI_S : PSY_3GPP_EN_SPREAD_HI_L1;
  294. i = 0;
  295. prev = 0.0;
  296. for (g = 0; g < ctx->num_bands[j]; g++) {
  297. i += band_sizes[g];
  298. bark = calc_bark((i-1) * line_to_frequency);
  299. coeffs[g].barks = (bark + prev) / 2.0;
  300. prev = bark;
  301. }
  302. for (g = 0; g < ctx->num_bands[j] - 1; g++) {
  303. AacPsyCoeffs *coeff = &coeffs[g];
  304. float bark_width = coeffs[g+1].barks - coeffs->barks;
  305. coeff->spread_low[0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_LOW);
  306. coeff->spread_hi [0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_HI);
  307. coeff->spread_low[1] = pow(10.0, -bark_width * en_spread_low);
  308. coeff->spread_hi [1] = pow(10.0, -bark_width * en_spread_hi);
  309. pe_min = bark_pe * bark_width;
  310. minsnr = exp2(pe_min / band_sizes[g]) - 1.5f;
  311. coeff->min_snr = av_clipf(1.0f / minsnr, PSY_SNR_25DB, PSY_SNR_1DB);
  312. }
  313. start = 0;
  314. for (g = 0; g < ctx->num_bands[j]; g++) {
  315. minscale = ath(start * line_to_frequency, ATH_ADD);
  316. for (i = 1; i < band_sizes[g]; i++)
  317. minscale = FFMIN(minscale, ath((start + i) * line_to_frequency, ATH_ADD));
  318. coeffs[g].ath = minscale - minath;
  319. start += band_sizes[g];
  320. }
  321. }
  322. pctx->ch = av_mallocz_array(ctx->avctx->channels, sizeof(AacPsyChannel));
  323. if (!pctx->ch) {
  324. av_freep(&ctx->model_priv_data);
  325. return AVERROR(ENOMEM);
  326. }
  327. lame_window_init(pctx, ctx->avctx);
  328. return 0;
  329. }
  330. /**
  331. * IIR filter used in block switching decision
  332. */
  333. static float iir_filter(int in, float state[2])
  334. {
  335. float ret;
  336. ret = 0.7548f * (in - state[0]) + 0.5095f * state[1];
  337. state[0] = in;
  338. state[1] = ret;
  339. return ret;
  340. }
  341. /**
  342. * window grouping information stored as bits (0 - new group, 1 - group continues)
  343. */
  344. static const uint8_t window_grouping[9] = {
  345. 0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36
  346. };
  347. /**
  348. * Tell encoder which window types to use.
  349. * @see 3GPP TS26.403 5.4.1 "Blockswitching"
  350. */
  351. static av_unused FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx,
  352. const int16_t *audio,
  353. const int16_t *la,
  354. int channel, int prev_type)
  355. {
  356. int i, j;
  357. int br = ctx->avctx->bit_rate / ctx->avctx->channels;
  358. int attack_ratio = br <= 16000 ? 18 : 10;
  359. AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
  360. AacPsyChannel *pch = &pctx->ch[channel];
  361. uint8_t grouping = 0;
  362. int next_type = pch->next_window_seq;
  363. FFPsyWindowInfo wi = { { 0 } };
  364. if (la) {
  365. float s[8], v;
  366. int switch_to_eight = 0;
  367. float sum = 0.0, sum2 = 0.0;
  368. int attack_n = 0;
  369. int stay_short = 0;
  370. for (i = 0; i < 8; i++) {
  371. for (j = 0; j < 128; j++) {
  372. v = iir_filter(la[i*128+j], pch->iir_state);
  373. sum += v*v;
  374. }
  375. s[i] = sum;
  376. sum2 += sum;
  377. }
  378. for (i = 0; i < 8; i++) {
  379. if (s[i] > pch->win_energy * attack_ratio) {
  380. attack_n = i + 1;
  381. switch_to_eight = 1;
  382. break;
  383. }
  384. }
  385. pch->win_energy = pch->win_energy*7/8 + sum2/64;
  386. wi.window_type[1] = prev_type;
  387. switch (prev_type) {
  388. case ONLY_LONG_SEQUENCE:
  389. wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
  390. next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
  391. break;
  392. case LONG_START_SEQUENCE:
  393. wi.window_type[0] = EIGHT_SHORT_SEQUENCE;
  394. grouping = pch->next_grouping;
  395. next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
  396. break;
  397. case LONG_STOP_SEQUENCE:
  398. wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
  399. next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
  400. break;
  401. case EIGHT_SHORT_SEQUENCE:
  402. stay_short = next_type == EIGHT_SHORT_SEQUENCE || switch_to_eight;
  403. wi.window_type[0] = stay_short ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
  404. grouping = next_type == EIGHT_SHORT_SEQUENCE ? pch->next_grouping : 0;
  405. next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
  406. break;
  407. }
  408. pch->next_grouping = window_grouping[attack_n];
  409. pch->next_window_seq = next_type;
  410. } else {
  411. for (i = 0; i < 3; i++)
  412. wi.window_type[i] = prev_type;
  413. grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0;
  414. }
  415. wi.window_shape = 1;
  416. if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
  417. wi.num_windows = 1;
  418. wi.grouping[0] = 1;
  419. } else {
  420. int lastgrp = 0;
  421. wi.num_windows = 8;
  422. for (i = 0; i < 8; i++) {
  423. if (!((grouping >> i) & 1))
  424. lastgrp = i;
  425. wi.grouping[lastgrp]++;
  426. }
  427. }
  428. return wi;
  429. }
  430. /* 5.6.1.2 "Calculation of Bit Demand" */
  431. static int calc_bit_demand(AacPsyContext *ctx, float pe, int bits, int size,
  432. int short_window)
  433. {
  434. const float bitsave_slope = short_window ? PSY_3GPP_SAVE_SLOPE_S : PSY_3GPP_SAVE_SLOPE_L;
  435. const float bitsave_add = short_window ? PSY_3GPP_SAVE_ADD_S : PSY_3GPP_SAVE_ADD_L;
  436. const float bitspend_slope = short_window ? PSY_3GPP_SPEND_SLOPE_S : PSY_3GPP_SPEND_SLOPE_L;
  437. const float bitspend_add = short_window ? PSY_3GPP_SPEND_ADD_S : PSY_3GPP_SPEND_ADD_L;
  438. const float clip_low = short_window ? PSY_3GPP_CLIP_LO_S : PSY_3GPP_CLIP_LO_L;
  439. const float clip_high = short_window ? PSY_3GPP_CLIP_HI_S : PSY_3GPP_CLIP_HI_L;
  440. float clipped_pe, bit_save, bit_spend, bit_factor, fill_level;
  441. ctx->fill_level += ctx->frame_bits - bits;
  442. ctx->fill_level = av_clip(ctx->fill_level, 0, size);
  443. fill_level = av_clipf((float)ctx->fill_level / size, clip_low, clip_high);
  444. clipped_pe = av_clipf(pe, ctx->pe.min, ctx->pe.max);
  445. bit_save = (fill_level + bitsave_add) * bitsave_slope;
  446. assert(bit_save <= 0.3f && bit_save >= -0.05000001f);
  447. bit_spend = (fill_level + bitspend_add) * bitspend_slope;
  448. assert(bit_spend <= 0.5f && bit_spend >= -0.1f);
  449. /* The bit factor graph in the spec is obviously incorrect.
  450. * bit_spend + ((bit_spend - bit_spend))...
  451. * The reference encoder subtracts everything from 1, but also seems incorrect.
  452. * 1 - bit_save + ((bit_spend + bit_save))...
  453. * Hopefully below is correct.
  454. */
  455. bit_factor = 1.0f - bit_save + ((bit_spend - bit_save) / (ctx->pe.max - ctx->pe.min)) * (clipped_pe - ctx->pe.min);
  456. /* NOTE: The reference encoder attempts to center pe max/min around the current pe. */
  457. ctx->pe.max = FFMAX(pe, ctx->pe.max);
  458. ctx->pe.min = FFMIN(pe, ctx->pe.min);
  459. return FFMIN(ctx->frame_bits * bit_factor, ctx->frame_bits + size - bits);
  460. }
  461. static float calc_pe_3gpp(AacPsyBand *band)
  462. {
  463. float pe, a;
  464. band->pe = 0.0f;
  465. band->pe_const = 0.0f;
  466. band->active_lines = 0.0f;
  467. if (band->energy > band->thr) {
  468. a = log2f(band->energy);
  469. pe = a - log2f(band->thr);
  470. band->active_lines = band->nz_lines;
  471. if (pe < PSY_3GPP_C1) {
  472. pe = pe * PSY_3GPP_C3 + PSY_3GPP_C2;
  473. a = a * PSY_3GPP_C3 + PSY_3GPP_C2;
  474. band->active_lines *= PSY_3GPP_C3;
  475. }
  476. band->pe = pe * band->nz_lines;
  477. band->pe_const = a * band->nz_lines;
  478. }
  479. return band->pe;
  480. }
  481. static float calc_reduction_3gpp(float a, float desired_pe, float pe,
  482. float active_lines)
  483. {
  484. float thr_avg, reduction;
  485. if(active_lines == 0.0)
  486. return 0;
  487. thr_avg = exp2f((a - pe) / (4.0f * active_lines));
  488. reduction = exp2f((a - desired_pe) / (4.0f * active_lines)) - thr_avg;
  489. return FFMAX(reduction, 0.0f);
  490. }
  491. static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr,
  492. float reduction)
  493. {
  494. float thr = band->thr;
  495. if (band->energy > thr) {
  496. thr = sqrtf(thr);
  497. thr = sqrtf(thr) + reduction;
  498. thr *= thr;
  499. thr *= thr;
  500. /* This deviates from the 3GPP spec to match the reference encoder.
  501. * It performs min(thr_reduced, max(thr, energy/min_snr)) only for bands
  502. * that have hole avoidance on (active or inactive). It always reduces the
  503. * threshold of bands with hole avoidance off.
  504. */
  505. if (thr > band->energy * min_snr && band->avoid_holes != PSY_3GPP_AH_NONE) {
  506. thr = FFMAX(band->thr, band->energy * min_snr);
  507. band->avoid_holes = PSY_3GPP_AH_ACTIVE;
  508. }
  509. }
  510. return thr;
  511. }
  512. #ifndef calc_thr_3gpp
  513. static void calc_thr_3gpp(const FFPsyWindowInfo *wi, const int num_bands, AacPsyChannel *pch,
  514. const uint8_t *band_sizes, const float *coefs)
  515. {
  516. int i, w, g;
  517. int start = 0;
  518. for (w = 0; w < wi->num_windows*16; w += 16) {
  519. for (g = 0; g < num_bands; g++) {
  520. AacPsyBand *band = &pch->band[w+g];
  521. float form_factor = 0.0f;
  522. float Temp;
  523. band->energy = 0.0f;
  524. for (i = 0; i < band_sizes[g]; i++) {
  525. band->energy += coefs[start+i] * coefs[start+i];
  526. form_factor += sqrtf(fabs(coefs[start+i]));
  527. }
  528. Temp = band->energy > 0 ? sqrtf((float)band_sizes[g] / band->energy) : 0;
  529. band->thr = band->energy * 0.001258925f;
  530. band->nz_lines = form_factor * sqrtf(Temp);
  531. start += band_sizes[g];
  532. }
  533. }
  534. }
  535. #endif /* calc_thr_3gpp */
  536. #ifndef psy_hp_filter
  537. static void psy_hp_filter(const float *firbuf, float *hpfsmpl, const float *psy_fir_coeffs)
  538. {
  539. int i, j;
  540. for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) {
  541. float sum1, sum2;
  542. sum1 = firbuf[i + (PSY_LAME_FIR_LEN - 1) / 2];
  543. sum2 = 0.0;
  544. for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) {
  545. sum1 += psy_fir_coeffs[j] * (firbuf[i + j] + firbuf[i + PSY_LAME_FIR_LEN - j]);
  546. sum2 += psy_fir_coeffs[j + 1] * (firbuf[i + j + 1] + firbuf[i + PSY_LAME_FIR_LEN - j - 1]);
  547. }
  548. /* NOTE: The LAME psymodel expects it's input in the range -32768 to 32768.
  549. * Tuning this for normalized floats would be difficult. */
  550. hpfsmpl[i] = (sum1 + sum2) * 32768.0f;
  551. }
  552. }
  553. #endif /* psy_hp_filter */
  554. /**
  555. * Calculate band thresholds as suggested in 3GPP TS26.403
  556. */
  557. static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel,
  558. const float *coefs, const FFPsyWindowInfo *wi)
  559. {
  560. AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
  561. AacPsyChannel *pch = &pctx->ch[channel];
  562. int i, w, g;
  563. float desired_bits, desired_pe, delta_pe, reduction= NAN, spread_en[128] = {0};
  564. float a = 0.0f, active_lines = 0.0f, norm_fac = 0.0f;
  565. float pe = pctx->chan_bitrate > 32000 ? 0.0f : FFMAX(50.0f, 100.0f - pctx->chan_bitrate * 100.0f / 32000.0f);
  566. const int num_bands = ctx->num_bands[wi->num_windows == 8];
  567. const uint8_t *band_sizes = ctx->bands[wi->num_windows == 8];
  568. AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8];
  569. const float avoid_hole_thr = wi->num_windows == 8 ? PSY_3GPP_AH_THR_SHORT : PSY_3GPP_AH_THR_LONG;
  570. //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation"
  571. calc_thr_3gpp(wi, num_bands, pch, band_sizes, coefs);
  572. //modify thresholds and energies - spread, threshold in quiet, pre-echo control
  573. for (w = 0; w < wi->num_windows*16; w += 16) {
  574. AacPsyBand *bands = &pch->band[w];
  575. /* 5.4.2.3 "Spreading" & 5.4.3 "Spread Energy Calculation" */
  576. spread_en[0] = bands[0].energy;
  577. for (g = 1; g < num_bands; g++) {
  578. bands[g].thr = FFMAX(bands[g].thr, bands[g-1].thr * coeffs[g].spread_hi[0]);
  579. spread_en[w+g] = FFMAX(bands[g].energy, spread_en[w+g-1] * coeffs[g].spread_hi[1]);
  580. }
  581. for (g = num_bands - 2; g >= 0; g--) {
  582. bands[g].thr = FFMAX(bands[g].thr, bands[g+1].thr * coeffs[g].spread_low[0]);
  583. spread_en[w+g] = FFMAX(spread_en[w+g], spread_en[w+g+1] * coeffs[g].spread_low[1]);
  584. }
  585. //5.4.2.4 "Threshold in quiet"
  586. for (g = 0; g < num_bands; g++) {
  587. AacPsyBand *band = &bands[g];
  588. band->thr_quiet = band->thr = FFMAX(band->thr, coeffs[g].ath);
  589. //5.4.2.5 "Pre-echo control"
  590. if (!(wi->window_type[0] == LONG_STOP_SEQUENCE || (wi->window_type[1] == LONG_START_SEQUENCE && !w)))
  591. band->thr = FFMAX(PSY_3GPP_RPEMIN*band->thr, FFMIN(band->thr,
  592. PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet));
  593. /* 5.6.1.3.1 "Preparatory steps of the perceptual entropy calculation" */
  594. pe += calc_pe_3gpp(band);
  595. a += band->pe_const;
  596. active_lines += band->active_lines;
  597. /* 5.6.1.3.3 "Selection of the bands for avoidance of holes" */
  598. if (spread_en[w+g] * avoid_hole_thr > band->energy || coeffs[g].min_snr > 1.0f)
  599. band->avoid_holes = PSY_3GPP_AH_NONE;
  600. else
  601. band->avoid_holes = PSY_3GPP_AH_INACTIVE;
  602. }
  603. }
  604. /* 5.6.1.3.2 "Calculation of the desired perceptual entropy" */
  605. ctx->ch[channel].entropy = pe;
  606. desired_bits = calc_bit_demand(pctx, pe, ctx->bitres.bits, ctx->bitres.size, wi->num_windows == 8);
  607. desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits);
  608. /* NOTE: PE correction is kept simple. During initial testing it had very
  609. * little effect on the final bitrate. Probably a good idea to come
  610. * back and do more testing later.
  611. */
  612. if (ctx->bitres.bits > 0)
  613. desired_pe *= av_clipf(pctx->pe.previous / PSY_3GPP_BITS_TO_PE(ctx->bitres.bits),
  614. 0.85f, 1.15f);
  615. pctx->pe.previous = PSY_3GPP_BITS_TO_PE(desired_bits);
  616. ctx->bitres.alloc = desired_bits;
  617. if (desired_pe < pe) {
  618. /* 5.6.1.3.4 "First Estimation of the reduction value" */
  619. for (w = 0; w < wi->num_windows*16; w += 16) {
  620. reduction = calc_reduction_3gpp(a, desired_pe, pe, active_lines);
  621. pe = 0.0f;
  622. a = 0.0f;
  623. active_lines = 0.0f;
  624. for (g = 0; g < num_bands; g++) {
  625. AacPsyBand *band = &pch->band[w+g];
  626. band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
  627. /* recalculate PE */
  628. pe += calc_pe_3gpp(band);
  629. a += band->pe_const;
  630. active_lines += band->active_lines;
  631. }
  632. }
  633. /* 5.6.1.3.5 "Second Estimation of the reduction value" */
  634. for (i = 0; i < 2; i++) {
  635. float pe_no_ah = 0.0f, desired_pe_no_ah;
  636. active_lines = a = 0.0f;
  637. for (w = 0; w < wi->num_windows*16; w += 16) {
  638. for (g = 0; g < num_bands; g++) {
  639. AacPsyBand *band = &pch->band[w+g];
  640. if (band->avoid_holes != PSY_3GPP_AH_ACTIVE) {
  641. pe_no_ah += band->pe;
  642. a += band->pe_const;
  643. active_lines += band->active_lines;
  644. }
  645. }
  646. }
  647. desired_pe_no_ah = FFMAX(desired_pe - (pe - pe_no_ah), 0.0f);
  648. if (active_lines > 0.0f)
  649. reduction = calc_reduction_3gpp(a, desired_pe_no_ah, pe_no_ah, active_lines);
  650. pe = 0.0f;
  651. for (w = 0; w < wi->num_windows*16; w += 16) {
  652. for (g = 0; g < num_bands; g++) {
  653. AacPsyBand *band = &pch->band[w+g];
  654. if (active_lines > 0.0f)
  655. band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
  656. pe += calc_pe_3gpp(band);
  657. if (band->thr > 0.0f)
  658. band->norm_fac = band->active_lines / band->thr;
  659. else
  660. band->norm_fac = 0.0f;
  661. norm_fac += band->norm_fac;
  662. }
  663. }
  664. delta_pe = desired_pe - pe;
  665. if (fabs(delta_pe) > 0.05f * desired_pe)
  666. break;
  667. }
  668. if (pe < 1.15f * desired_pe) {
  669. /* 6.6.1.3.6 "Final threshold modification by linearization" */
  670. norm_fac = 1.0f / norm_fac;
  671. for (w = 0; w < wi->num_windows*16; w += 16) {
  672. for (g = 0; g < num_bands; g++) {
  673. AacPsyBand *band = &pch->band[w+g];
  674. if (band->active_lines > 0.5f) {
  675. float delta_sfb_pe = band->norm_fac * norm_fac * delta_pe;
  676. float thr = band->thr;
  677. thr *= exp2f(delta_sfb_pe / band->active_lines);
  678. if (thr > coeffs[g].min_snr * band->energy && band->avoid_holes == PSY_3GPP_AH_INACTIVE)
  679. thr = FFMAX(band->thr, coeffs[g].min_snr * band->energy);
  680. band->thr = thr;
  681. }
  682. }
  683. }
  684. } else {
  685. /* 5.6.1.3.7 "Further perceptual entropy reduction" */
  686. g = num_bands;
  687. while (pe > desired_pe && g--) {
  688. for (w = 0; w < wi->num_windows*16; w+= 16) {
  689. AacPsyBand *band = &pch->band[w+g];
  690. if (band->avoid_holes != PSY_3GPP_AH_NONE && coeffs[g].min_snr < PSY_SNR_1DB) {
  691. coeffs[g].min_snr = PSY_SNR_1DB;
  692. band->thr = band->energy * PSY_SNR_1DB;
  693. pe += band->active_lines * 1.5f - band->pe;
  694. }
  695. }
  696. }
  697. /* TODO: allow more holes (unused without mid/side) */
  698. }
  699. }
  700. for (w = 0; w < wi->num_windows*16; w += 16) {
  701. for (g = 0; g < num_bands; g++) {
  702. AacPsyBand *band = &pch->band[w+g];
  703. FFPsyBand *psy_band = &ctx->ch[channel].psy_bands[w+g];
  704. psy_band->threshold = band->thr;
  705. psy_band->energy = band->energy;
  706. psy_band->spread = band->active_lines * 2.0f / band_sizes[g];
  707. psy_band->bits = PSY_3GPP_PE_TO_BITS(band->pe);
  708. }
  709. }
  710. memcpy(pch->prev_band, pch->band, sizeof(pch->band));
  711. }
  712. static void psy_3gpp_analyze(FFPsyContext *ctx, int channel,
  713. const float **coeffs, const FFPsyWindowInfo *wi)
  714. {
  715. int ch;
  716. FFPsyChannelGroup *group = ff_psy_find_group(ctx, channel);
  717. for (ch = 0; ch < group->num_ch; ch++)
  718. psy_3gpp_analyze_channel(ctx, channel + ch, coeffs[ch], &wi[ch]);
  719. }
  720. static av_cold void psy_3gpp_end(FFPsyContext *apc)
  721. {
  722. AacPsyContext *pctx = (AacPsyContext*) apc->model_priv_data;
  723. av_freep(&pctx->ch);
  724. av_freep(&apc->model_priv_data);
  725. }
  726. static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock)
  727. {
  728. int blocktype = ONLY_LONG_SEQUENCE;
  729. if (uselongblock) {
  730. if (ctx->next_window_seq == EIGHT_SHORT_SEQUENCE)
  731. blocktype = LONG_STOP_SEQUENCE;
  732. } else {
  733. blocktype = EIGHT_SHORT_SEQUENCE;
  734. if (ctx->next_window_seq == ONLY_LONG_SEQUENCE)
  735. ctx->next_window_seq = LONG_START_SEQUENCE;
  736. if (ctx->next_window_seq == LONG_STOP_SEQUENCE)
  737. ctx->next_window_seq = EIGHT_SHORT_SEQUENCE;
  738. }
  739. wi->window_type[0] = ctx->next_window_seq;
  740. ctx->next_window_seq = blocktype;
  741. }
  742. static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx, const float *audio,
  743. const float *la, int channel, int prev_type)
  744. {
  745. AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
  746. AacPsyChannel *pch = &pctx->ch[channel];
  747. int grouping = 0;
  748. int uselongblock = 1;
  749. int attacks[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
  750. float clippings[AAC_NUM_BLOCKS_SHORT];
  751. int i;
  752. FFPsyWindowInfo wi = { { 0 } };
  753. if (la) {
  754. float hpfsmpl[AAC_BLOCK_SIZE_LONG];
  755. float const *pf = hpfsmpl;
  756. float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
  757. float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
  758. float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
  759. const float *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN);
  760. int att_sum = 0;
  761. /* LAME comment: apply high pass filter of fs/4 */
  762. psy_hp_filter(firbuf, hpfsmpl, psy_fir_coeffs);
  763. /* Calculate the energies of each sub-shortblock */
  764. for (i = 0; i < PSY_LAME_NUM_SUBBLOCKS; i++) {
  765. energy_subshort[i] = pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS)];
  766. assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)] > 0);
  767. attack_intensity[i] = energy_subshort[i] / pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)];
  768. energy_short[0] += energy_subshort[i];
  769. }
  770. for (i = 0; i < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; i++) {
  771. float const *const pfe = pf + AAC_BLOCK_SIZE_LONG / (AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS);
  772. float p = 1.0f;
  773. for (; pf < pfe; pf++)
  774. p = FFMAX(p, fabsf(*pf));
  775. pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p;
  776. energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p;
  777. /* NOTE: The indexes below are [i + 3 - 2] in the LAME source.
  778. * Obviously the 3 and 2 have some significance, or this would be just [i + 1]
  779. * (which is what we use here). What the 3 stands for is ambiguous, as it is both
  780. * number of short blocks, and the number of sub-short blocks.
  781. * It seems that LAME is comparing each sub-block to sub-block + 1 in the
  782. * previous block.
  783. */
  784. if (p > energy_subshort[i + 1])
  785. p = p / energy_subshort[i + 1];
  786. else if (energy_subshort[i + 1] > p * 10.0f)
  787. p = energy_subshort[i + 1] / (p * 10.0f);
  788. else
  789. p = 0.0;
  790. attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p;
  791. }
  792. /* compare energy between sub-short blocks */
  793. for (i = 0; i < (AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS; i++)
  794. if (!attacks[i / PSY_LAME_NUM_SUBBLOCKS])
  795. if (attack_intensity[i] > pch->attack_threshold)
  796. attacks[i / PSY_LAME_NUM_SUBBLOCKS] = (i % PSY_LAME_NUM_SUBBLOCKS) + 1;
  797. /* should have energy change between short blocks, in order to avoid periodic signals */
  798. /* Good samples to show the effect are Trumpet test songs */
  799. /* GB: tuned (1) to avoid too many short blocks for test sample TRUMPET */
  800. /* RH: tuned (2) to let enough short blocks through for test sample FSOL and SNAPS */
  801. for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) {
  802. float const u = energy_short[i - 1];
  803. float const v = energy_short[i];
  804. float const m = FFMAX(u, v);
  805. if (m < 40000) { /* (2) */
  806. if (u < 1.7f * v && v < 1.7f * u) { /* (1) */
  807. if (i == 1 && attacks[0] < attacks[i])
  808. attacks[0] = 0;
  809. attacks[i] = 0;
  810. }
  811. }
  812. att_sum += attacks[i];
  813. }
  814. if (attacks[0] <= pch->prev_attack)
  815. attacks[0] = 0;
  816. att_sum += attacks[0];
  817. /* 3 below indicates the previous attack happened in the last sub-block of the previous sequence */
  818. if (pch->prev_attack == 3 || att_sum) {
  819. uselongblock = 0;
  820. for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++)
  821. if (attacks[i] && attacks[i-1])
  822. attacks[i] = 0;
  823. }
  824. } else {
  825. /* We have no lookahead info, so just use same type as the previous sequence. */
  826. uselongblock = !(prev_type == EIGHT_SHORT_SEQUENCE);
  827. }
  828. lame_apply_block_type(pch, &wi, uselongblock);
  829. /* Calculate input sample maximums and evaluate clipping risk */
  830. if (audio) {
  831. for (i = 0; i < AAC_NUM_BLOCKS_SHORT; i++) {
  832. const float *wbuf = audio + i * AAC_BLOCK_SIZE_SHORT;
  833. float max = 0;
  834. int j;
  835. for (j = 0; j < AAC_BLOCK_SIZE_SHORT; j++)
  836. max = FFMAX(max, fabsf(wbuf[j]));
  837. clippings[i] = max;
  838. }
  839. } else {
  840. for (i = 0; i < 8; i++)
  841. clippings[i] = 0;
  842. }
  843. wi.window_type[1] = prev_type;
  844. if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
  845. float clipping = 0.0f;
  846. wi.num_windows = 1;
  847. wi.grouping[0] = 1;
  848. if (wi.window_type[0] == LONG_START_SEQUENCE)
  849. wi.window_shape = 0;
  850. else
  851. wi.window_shape = 1;
  852. for (i = 0; i < 8; i++)
  853. clipping = FFMAX(clipping, clippings[i]);
  854. wi.clipping[0] = clipping;
  855. } else {
  856. int lastgrp = 0;
  857. wi.num_windows = 8;
  858. wi.window_shape = 0;
  859. for (i = 0; i < 8; i++) {
  860. if (!((pch->next_grouping >> i) & 1))
  861. lastgrp = i;
  862. wi.grouping[lastgrp]++;
  863. }
  864. for (i = 0; i < 8; i += wi.grouping[i]) {
  865. int w;
  866. float clipping = 0.0f;
  867. for (w = 0; w < wi.grouping[i] && !clipping; w++)
  868. clipping = FFMAX(clipping, clippings[i+w]);
  869. wi.clipping[i] = clipping;
  870. }
  871. }
  872. /* Determine grouping, based on the location of the first attack, and save for
  873. * the next frame.
  874. * FIXME: Move this to analysis.
  875. * TODO: Tune groupings depending on attack location
  876. * TODO: Handle more than one attack in a group
  877. */
  878. for (i = 0; i < 9; i++) {
  879. if (attacks[i]) {
  880. grouping = i;
  881. break;
  882. }
  883. }
  884. pch->next_grouping = window_grouping[grouping];
  885. pch->prev_attack = attacks[8];
  886. return wi;
  887. }
  888. const FFPsyModel ff_aac_psy_model =
  889. {
  890. .name = "3GPP TS 26.403-inspired model",
  891. .init = psy_3gpp_init,
  892. .window = psy_lame_window,
  893. .analyze = psy_3gpp_analyze,
  894. .end = psy_3gpp_end,
  895. };