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.

1055 lines
40KB

  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/internal.h"
  27. #include "libavutil/libm.h"
  28. #include "avcodec.h"
  29. #include "aactab.h"
  30. #include "psymodel.h"
  31. /***********************************
  32. * TODOs:
  33. * try other bitrate controlling mechanism (maybe use ratecontrol.c?)
  34. * control quality for quality-based output
  35. **********************************/
  36. /**
  37. * constants for 3GPP AAC psychoacoustic model
  38. * @{
  39. */
  40. #define PSY_3GPP_THR_SPREAD_HI 1.5f // spreading factor for low-to-hi threshold spreading (15 dB/Bark)
  41. #define PSY_3GPP_THR_SPREAD_LOW 3.0f // spreading factor for hi-to-low threshold spreading (30 dB/Bark)
  42. /* spreading factor for low-to-hi energy spreading, long block, > 22kbps/channel (20dB/Bark) */
  43. #define PSY_3GPP_EN_SPREAD_HI_L1 2.0f
  44. /* spreading factor for low-to-hi energy spreading, long block, <= 22kbps/channel (15dB/Bark) */
  45. #define PSY_3GPP_EN_SPREAD_HI_L2 1.5f
  46. /* spreading factor for low-to-hi energy spreading, short block (15 dB/Bark) */
  47. #define PSY_3GPP_EN_SPREAD_HI_S 1.5f
  48. /* spreading factor for hi-to-low energy spreading, long block (30dB/Bark) */
  49. #define PSY_3GPP_EN_SPREAD_LOW_L 3.0f
  50. /* spreading factor for hi-to-low energy spreading, short block (20dB/Bark) */
  51. #define PSY_3GPP_EN_SPREAD_LOW_S 2.0f
  52. #define PSY_3GPP_RPEMIN 0.01f
  53. #define PSY_3GPP_RPELEV 2.0f
  54. #define PSY_3GPP_C1 3.0f /* log2(8) */
  55. #define PSY_3GPP_C2 1.3219281f /* log2(2.5) */
  56. #define PSY_3GPP_C3 0.55935729f /* 1 - C2 / C1 */
  57. #define PSY_SNR_1DB 7.9432821e-1f /* -1dB */
  58. #define PSY_SNR_25DB 3.1622776e-3f /* -25dB */
  59. #define PSY_3GPP_SAVE_SLOPE_L -0.46666667f
  60. #define PSY_3GPP_SAVE_SLOPE_S -0.36363637f
  61. #define PSY_3GPP_SAVE_ADD_L -0.84285712f
  62. #define PSY_3GPP_SAVE_ADD_S -0.75f
  63. #define PSY_3GPP_SPEND_SLOPE_L 0.66666669f
  64. #define PSY_3GPP_SPEND_SLOPE_S 0.81818181f
  65. #define PSY_3GPP_SPEND_ADD_L -0.35f
  66. #define PSY_3GPP_SPEND_ADD_S -0.26111111f
  67. #define PSY_3GPP_CLIP_LO_L 0.2f
  68. #define PSY_3GPP_CLIP_LO_S 0.2f
  69. #define PSY_3GPP_CLIP_HI_L 0.95f
  70. #define PSY_3GPP_CLIP_HI_S 0.75f
  71. #define PSY_3GPP_AH_THR_LONG 0.5f
  72. #define PSY_3GPP_AH_THR_SHORT 0.63f
  73. #define PSY_PE_FORGET_SLOPE 511
  74. enum {
  75. PSY_3GPP_AH_NONE,
  76. PSY_3GPP_AH_INACTIVE,
  77. PSY_3GPP_AH_ACTIVE
  78. };
  79. #define PSY_3GPP_BITS_TO_PE(bits) ((bits) * 1.18f)
  80. #define PSY_3GPP_PE_TO_BITS(bits) ((bits) / 1.18f)
  81. /* LAME psy model constants */
  82. #define PSY_LAME_FIR_LEN 21 ///< LAME psy model FIR order
  83. #define AAC_BLOCK_SIZE_LONG 1024 ///< long block size
  84. #define AAC_BLOCK_SIZE_SHORT 128 ///< short block size
  85. #define AAC_NUM_BLOCKS_SHORT 8 ///< number of blocks in a short sequence
  86. #define PSY_LAME_NUM_SUBBLOCKS 3 ///< Number of sub-blocks in each short block
  87. /**
  88. * @}
  89. */
  90. /**
  91. * information for single band used by 3GPP TS26.403-inspired psychoacoustic model
  92. */
  93. typedef struct AacPsyBand{
  94. float energy; ///< band energy
  95. float thr; ///< energy threshold
  96. float thr_quiet; ///< threshold in quiet
  97. float nz_lines; ///< number of non-zero spectral lines
  98. float active_lines; ///< number of active spectral lines
  99. float pe; ///< perceptual entropy
  100. float pe_const; ///< constant part of the PE calculation
  101. float norm_fac; ///< normalization factor for linearization
  102. int avoid_holes; ///< hole avoidance flag
  103. }AacPsyBand;
  104. /**
  105. * single/pair channel context for psychoacoustic model
  106. */
  107. typedef struct AacPsyChannel{
  108. AacPsyBand band[128]; ///< bands information
  109. AacPsyBand prev_band[128]; ///< bands information from the previous frame
  110. float win_energy; ///< sliding average of channel energy
  111. float iir_state[2]; ///< hi-pass IIR filter state
  112. uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence)
  113. enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame
  114. /* LAME psy model specific members */
  115. float attack_threshold; ///< attack threshold for this channel
  116. float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS];
  117. int prev_attack; ///< attack value for the last short block in the previous sequence
  118. }AacPsyChannel;
  119. /**
  120. * psychoacoustic model frame type-dependent coefficients
  121. */
  122. typedef struct AacPsyCoeffs{
  123. float ath; ///< absolute threshold of hearing per bands
  124. float barks; ///< Bark value for each spectral band in long frame
  125. float spread_low[2]; ///< spreading factor for low-to-high threshold spreading in long frame
  126. float spread_hi [2]; ///< spreading factor for high-to-low threshold spreading in long frame
  127. float min_snr; ///< minimal SNR
  128. }AacPsyCoeffs;
  129. /**
  130. * 3GPP TS26.403-inspired psychoacoustic model specific data
  131. */
  132. typedef struct AacPsyContext{
  133. int chan_bitrate; ///< bitrate per channel
  134. int frame_bits; ///< average bits per frame
  135. int fill_level; ///< bit reservoir fill level
  136. struct {
  137. float min; ///< minimum allowed PE for bit factor calculation
  138. float max; ///< maximum allowed PE for bit factor calculation
  139. float previous; ///< allowed PE of the previous frame
  140. float correction; ///< PE correction factor
  141. } pe;
  142. AacPsyCoeffs psy_coef[2][64];
  143. AacPsyChannel *ch;
  144. float global_quality; ///< normalized global quality taken from avctx
  145. }AacPsyContext;
  146. /**
  147. * LAME psy model preset struct
  148. */
  149. typedef struct PsyLamePreset {
  150. int quality; ///< Quality to map the rest of the vaules to.
  151. /* This is overloaded to be both kbps per channel in ABR mode, and
  152. * requested quality in constant quality mode.
  153. */
  154. float st_lrm; ///< short threshold for L, R, and M channels
  155. } PsyLamePreset;
  156. /**
  157. * LAME psy model preset table for ABR
  158. */
  159. static const PsyLamePreset psy_abr_map[] = {
  160. /* TODO: Tuning. These were taken from LAME. */
  161. /* kbps/ch st_lrm */
  162. { 8, 6.60},
  163. { 16, 6.60},
  164. { 24, 6.60},
  165. { 32, 6.60},
  166. { 40, 6.60},
  167. { 48, 6.60},
  168. { 56, 6.60},
  169. { 64, 6.40},
  170. { 80, 6.00},
  171. { 96, 5.60},
  172. {112, 5.20},
  173. {128, 5.20},
  174. {160, 5.20}
  175. };
  176. /**
  177. * LAME psy model preset table for constant quality
  178. */
  179. static const PsyLamePreset psy_vbr_map[] = {
  180. /* vbr_q st_lrm */
  181. { 0, 4.20},
  182. { 1, 4.20},
  183. { 2, 4.20},
  184. { 3, 4.20},
  185. { 4, 4.20},
  186. { 5, 4.20},
  187. { 6, 4.20},
  188. { 7, 4.20},
  189. { 8, 4.20},
  190. { 9, 4.20},
  191. {10, 4.20}
  192. };
  193. /**
  194. * LAME psy model FIR coefficient table
  195. */
  196. static const float psy_fir_coeffs[] = {
  197. -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2,
  198. -3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2,
  199. -5.52212e-17 * 2, -0.313819 * 2
  200. };
  201. #if ARCH_MIPS
  202. # include "mips/aacpsy_mips.h"
  203. #endif /* ARCH_MIPS */
  204. /**
  205. * Calculate the ABR attack threshold from the above LAME psymodel table.
  206. */
  207. static float lame_calc_attack_threshold(int bitrate)
  208. {
  209. /* Assume max bitrate to start with */
  210. int lower_range = 12, upper_range = 12;
  211. int lower_range_kbps = psy_abr_map[12].quality;
  212. int upper_range_kbps = psy_abr_map[12].quality;
  213. int i;
  214. /* Determine which bitrates the value specified falls between.
  215. * If the loop ends without breaking our above assumption of 320kbps was correct.
  216. */
  217. for (i = 1; i < 13; i++) {
  218. if (FFMAX(bitrate, psy_abr_map[i].quality) != bitrate) {
  219. upper_range = i;
  220. upper_range_kbps = psy_abr_map[i ].quality;
  221. lower_range = i - 1;
  222. lower_range_kbps = psy_abr_map[i - 1].quality;
  223. break; /* Upper range found */
  224. }
  225. }
  226. /* Determine which range the value specified is closer to */
  227. if ((upper_range_kbps - bitrate) > (bitrate - lower_range_kbps))
  228. return psy_abr_map[lower_range].st_lrm;
  229. return psy_abr_map[upper_range].st_lrm;
  230. }
  231. /**
  232. * LAME psy model specific initialization
  233. */
  234. static av_cold void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx)
  235. {
  236. int i, j;
  237. for (i = 0; i < avctx->channels; i++) {
  238. AacPsyChannel *pch = &ctx->ch[i];
  239. if (avctx->flags & AV_CODEC_FLAG_QSCALE)
  240. pch->attack_threshold = psy_vbr_map[avctx->global_quality / FF_QP2LAMBDA].st_lrm;
  241. else
  242. pch->attack_threshold = lame_calc_attack_threshold(avctx->bit_rate / avctx->channels / 1000);
  243. for (j = 0; j < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; j++)
  244. pch->prev_energy_subshort[j] = 10.0f;
  245. }
  246. }
  247. /**
  248. * Calculate Bark value for given line.
  249. */
  250. static av_cold float calc_bark(float f)
  251. {
  252. return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f));
  253. }
  254. #define ATH_ADD 4
  255. /**
  256. * Calculate ATH value for given frequency.
  257. * Borrowed from Lame.
  258. */
  259. static av_cold float ath(float f, float add)
  260. {
  261. f /= 1000.0f;
  262. return 3.64 * pow(f, -0.8)
  263. - 6.8 * exp(-0.6 * (f - 3.4) * (f - 3.4))
  264. + 6.0 * exp(-0.15 * (f - 8.7) * (f - 8.7))
  265. + (0.6 + 0.04 * add) * 0.001 * f * f * f * f;
  266. }
  267. static av_cold int psy_3gpp_init(FFPsyContext *ctx) {
  268. AacPsyContext *pctx;
  269. float bark;
  270. int i, j, g, start;
  271. float prev, minscale, minath, minsnr, pe_min;
  272. int chan_bitrate = ctx->avctx->bit_rate / ((ctx->avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : ctx->avctx->channels);
  273. const int bandwidth = ctx->cutoff ? ctx->cutoff : AAC_CUTOFF(ctx->avctx);
  274. const float num_bark = calc_bark((float)bandwidth);
  275. ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext));
  276. if (!ctx->model_priv_data)
  277. return AVERROR(ENOMEM);
  278. pctx = (AacPsyContext*) ctx->model_priv_data;
  279. pctx->global_quality = (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120) * 0.01f;
  280. if (ctx->avctx->flags & CODEC_FLAG_QSCALE) {
  281. /* Use the target average bitrate to compute spread parameters */
  282. chan_bitrate = (int)(chan_bitrate / 120.0 * (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120));
  283. }
  284. pctx->chan_bitrate = chan_bitrate;
  285. pctx->frame_bits = FFMIN(2560, chan_bitrate * AAC_BLOCK_SIZE_LONG / ctx->avctx->sample_rate);
  286. pctx->pe.min = 8.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
  287. pctx->pe.max = 12.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
  288. ctx->bitres.size = 6144 - pctx->frame_bits;
  289. ctx->bitres.size -= ctx->bitres.size % 8;
  290. pctx->fill_level = ctx->bitres.size;
  291. minath = ath(3410 - 0.733 * ATH_ADD, ATH_ADD);
  292. for (j = 0; j < 2; j++) {
  293. AacPsyCoeffs *coeffs = pctx->psy_coef[j];
  294. const uint8_t *band_sizes = ctx->bands[j];
  295. float line_to_frequency = ctx->avctx->sample_rate / (j ? 256.f : 2048.0f);
  296. float avg_chan_bits = chan_bitrate * (j ? 128.0f : 1024.0f) / ctx->avctx->sample_rate;
  297. /* reference encoder uses 2.4% here instead of 60% like the spec says */
  298. float bark_pe = 0.024f * PSY_3GPP_BITS_TO_PE(avg_chan_bits) / num_bark;
  299. float en_spread_low = j ? PSY_3GPP_EN_SPREAD_LOW_S : PSY_3GPP_EN_SPREAD_LOW_L;
  300. /* High energy spreading for long blocks <= 22kbps/channel and short blocks are the same. */
  301. float en_spread_hi = (j || (chan_bitrate <= 22.0f)) ? PSY_3GPP_EN_SPREAD_HI_S : PSY_3GPP_EN_SPREAD_HI_L1;
  302. i = 0;
  303. prev = 0.0;
  304. for (g = 0; g < ctx->num_bands[j]; g++) {
  305. i += band_sizes[g];
  306. bark = calc_bark((i-1) * line_to_frequency);
  307. coeffs[g].barks = (bark + prev) / 2.0;
  308. prev = bark;
  309. }
  310. for (g = 0; g < ctx->num_bands[j] - 1; g++) {
  311. AacPsyCoeffs *coeff = &coeffs[g];
  312. float bark_width = coeffs[g+1].barks - coeffs->barks;
  313. coeff->spread_low[0] = ff_exp10(-bark_width * PSY_3GPP_THR_SPREAD_LOW);
  314. coeff->spread_hi [0] = ff_exp10(-bark_width * PSY_3GPP_THR_SPREAD_HI);
  315. coeff->spread_low[1] = ff_exp10(-bark_width * en_spread_low);
  316. coeff->spread_hi [1] = ff_exp10(-bark_width * en_spread_hi);
  317. pe_min = bark_pe * bark_width;
  318. minsnr = exp2(pe_min / band_sizes[g]) - 1.5f;
  319. coeff->min_snr = av_clipf(1.0f / minsnr, PSY_SNR_25DB, PSY_SNR_1DB);
  320. }
  321. start = 0;
  322. for (g = 0; g < ctx->num_bands[j]; g++) {
  323. minscale = ath(start * line_to_frequency, ATH_ADD);
  324. for (i = 1; i < band_sizes[g]; i++)
  325. minscale = FFMIN(minscale, ath((start + i) * line_to_frequency, ATH_ADD));
  326. coeffs[g].ath = minscale - minath;
  327. start += band_sizes[g];
  328. }
  329. }
  330. pctx->ch = av_mallocz_array(ctx->avctx->channels, sizeof(AacPsyChannel));
  331. if (!pctx->ch) {
  332. av_freep(&ctx->model_priv_data);
  333. return AVERROR(ENOMEM);
  334. }
  335. lame_window_init(pctx, ctx->avctx);
  336. return 0;
  337. }
  338. /**
  339. * IIR filter used in block switching decision
  340. */
  341. static float iir_filter(int in, float state[2])
  342. {
  343. float ret;
  344. ret = 0.7548f * (in - state[0]) + 0.5095f * state[1];
  345. state[0] = in;
  346. state[1] = ret;
  347. return ret;
  348. }
  349. /**
  350. * window grouping information stored as bits (0 - new group, 1 - group continues)
  351. */
  352. static const uint8_t window_grouping[9] = {
  353. 0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36
  354. };
  355. /**
  356. * Tell encoder which window types to use.
  357. * @see 3GPP TS26.403 5.4.1 "Blockswitching"
  358. */
  359. static av_unused FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx,
  360. const int16_t *audio,
  361. const int16_t *la,
  362. int channel, int prev_type)
  363. {
  364. int i, j;
  365. int br = ((AacPsyContext*)ctx->model_priv_data)->chan_bitrate;
  366. int attack_ratio = br <= 16000 ? 18 : 10;
  367. AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
  368. AacPsyChannel *pch = &pctx->ch[channel];
  369. uint8_t grouping = 0;
  370. int next_type = pch->next_window_seq;
  371. FFPsyWindowInfo wi = { { 0 } };
  372. if (la) {
  373. float s[8], v;
  374. int switch_to_eight = 0;
  375. float sum = 0.0, sum2 = 0.0;
  376. int attack_n = 0;
  377. int stay_short = 0;
  378. for (i = 0; i < 8; i++) {
  379. for (j = 0; j < 128; j++) {
  380. v = iir_filter(la[i*128+j], pch->iir_state);
  381. sum += v*v;
  382. }
  383. s[i] = sum;
  384. sum2 += sum;
  385. }
  386. for (i = 0; i < 8; i++) {
  387. if (s[i] > pch->win_energy * attack_ratio) {
  388. attack_n = i + 1;
  389. switch_to_eight = 1;
  390. break;
  391. }
  392. }
  393. pch->win_energy = pch->win_energy*7/8 + sum2/64;
  394. wi.window_type[1] = prev_type;
  395. switch (prev_type) {
  396. case ONLY_LONG_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 LONG_START_SEQUENCE:
  401. wi.window_type[0] = EIGHT_SHORT_SEQUENCE;
  402. grouping = pch->next_grouping;
  403. next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
  404. break;
  405. case LONG_STOP_SEQUENCE:
  406. wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
  407. next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
  408. break;
  409. case EIGHT_SHORT_SEQUENCE:
  410. stay_short = next_type == EIGHT_SHORT_SEQUENCE || switch_to_eight;
  411. wi.window_type[0] = stay_short ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
  412. grouping = next_type == EIGHT_SHORT_SEQUENCE ? pch->next_grouping : 0;
  413. next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
  414. break;
  415. }
  416. pch->next_grouping = window_grouping[attack_n];
  417. pch->next_window_seq = next_type;
  418. } else {
  419. for (i = 0; i < 3; i++)
  420. wi.window_type[i] = prev_type;
  421. grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0;
  422. }
  423. wi.window_shape = 1;
  424. if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
  425. wi.num_windows = 1;
  426. wi.grouping[0] = 1;
  427. } else {
  428. int lastgrp = 0;
  429. wi.num_windows = 8;
  430. for (i = 0; i < 8; i++) {
  431. if (!((grouping >> i) & 1))
  432. lastgrp = i;
  433. wi.grouping[lastgrp]++;
  434. }
  435. }
  436. return wi;
  437. }
  438. /* 5.6.1.2 "Calculation of Bit Demand" */
  439. static int calc_bit_demand(AacPsyContext *ctx, float pe, int bits, int size,
  440. int short_window)
  441. {
  442. const float bitsave_slope = short_window ? PSY_3GPP_SAVE_SLOPE_S : PSY_3GPP_SAVE_SLOPE_L;
  443. const float bitsave_add = short_window ? PSY_3GPP_SAVE_ADD_S : PSY_3GPP_SAVE_ADD_L;
  444. const float bitspend_slope = short_window ? PSY_3GPP_SPEND_SLOPE_S : PSY_3GPP_SPEND_SLOPE_L;
  445. const float bitspend_add = short_window ? PSY_3GPP_SPEND_ADD_S : PSY_3GPP_SPEND_ADD_L;
  446. const float clip_low = short_window ? PSY_3GPP_CLIP_LO_S : PSY_3GPP_CLIP_LO_L;
  447. const float clip_high = short_window ? PSY_3GPP_CLIP_HI_S : PSY_3GPP_CLIP_HI_L;
  448. float clipped_pe, bit_save, bit_spend, bit_factor, fill_level, forgetful_min_pe;
  449. ctx->fill_level += ctx->frame_bits - bits;
  450. ctx->fill_level = av_clip(ctx->fill_level, 0, size);
  451. fill_level = av_clipf((float)ctx->fill_level / size, clip_low, clip_high);
  452. clipped_pe = av_clipf(pe, ctx->pe.min, ctx->pe.max);
  453. bit_save = (fill_level + bitsave_add) * bitsave_slope;
  454. assert(bit_save <= 0.3f && bit_save >= -0.05000001f);
  455. bit_spend = (fill_level + bitspend_add) * bitspend_slope;
  456. assert(bit_spend <= 0.5f && bit_spend >= -0.1f);
  457. /* The bit factor graph in the spec is obviously incorrect.
  458. * bit_spend + ((bit_spend - bit_spend))...
  459. * The reference encoder subtracts everything from 1, but also seems incorrect.
  460. * 1 - bit_save + ((bit_spend + bit_save))...
  461. * Hopefully below is correct.
  462. */
  463. bit_factor = 1.0f - bit_save + ((bit_spend - bit_save) / (ctx->pe.max - ctx->pe.min)) * (clipped_pe - ctx->pe.min);
  464. /* NOTE: The reference encoder attempts to center pe max/min around the current pe.
  465. * Here we do that by slowly forgetting pe.min when pe stays in a range that makes
  466. * it unlikely (ie: above the mean)
  467. */
  468. ctx->pe.max = FFMAX(pe, ctx->pe.max);
  469. forgetful_min_pe = ((ctx->pe.min * PSY_PE_FORGET_SLOPE)
  470. + FFMAX(ctx->pe.min, pe * (pe / ctx->pe.max))) / (PSY_PE_FORGET_SLOPE + 1);
  471. ctx->pe.min = FFMIN(pe, forgetful_min_pe);
  472. /* NOTE: allocate a minimum of 1/8th average frame bits, to avoid
  473. * reservoir starvation from producing zero-bit frames
  474. */
  475. return FFMIN(
  476. ctx->frame_bits * bit_factor,
  477. FFMAX(ctx->frame_bits + size - bits, ctx->frame_bits / 8));
  478. }
  479. static float calc_pe_3gpp(AacPsyBand *band)
  480. {
  481. float pe, a;
  482. band->pe = 0.0f;
  483. band->pe_const = 0.0f;
  484. band->active_lines = 0.0f;
  485. if (band->energy > band->thr) {
  486. a = log2f(band->energy);
  487. pe = a - log2f(band->thr);
  488. band->active_lines = band->nz_lines;
  489. if (pe < PSY_3GPP_C1) {
  490. pe = pe * PSY_3GPP_C3 + PSY_3GPP_C2;
  491. a = a * PSY_3GPP_C3 + PSY_3GPP_C2;
  492. band->active_lines *= PSY_3GPP_C3;
  493. }
  494. band->pe = pe * band->nz_lines;
  495. band->pe_const = a * band->nz_lines;
  496. }
  497. return band->pe;
  498. }
  499. static float calc_reduction_3gpp(float a, float desired_pe, float pe,
  500. float active_lines)
  501. {
  502. float thr_avg, reduction;
  503. if(active_lines == 0.0)
  504. return 0;
  505. thr_avg = exp2f((a - pe) / (4.0f * active_lines));
  506. reduction = exp2f((a - desired_pe) / (4.0f * active_lines)) - thr_avg;
  507. return FFMAX(reduction, 0.0f);
  508. }
  509. static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr,
  510. float reduction)
  511. {
  512. float thr = band->thr;
  513. if (band->energy > thr) {
  514. thr = sqrtf(thr);
  515. thr = sqrtf(thr) + reduction;
  516. thr *= thr;
  517. thr *= thr;
  518. /* This deviates from the 3GPP spec to match the reference encoder.
  519. * It performs min(thr_reduced, max(thr, energy/min_snr)) only for bands
  520. * that have hole avoidance on (active or inactive). It always reduces the
  521. * threshold of bands with hole avoidance off.
  522. */
  523. if (thr > band->energy * min_snr && band->avoid_holes != PSY_3GPP_AH_NONE) {
  524. thr = FFMAX(band->thr, band->energy * min_snr);
  525. band->avoid_holes = PSY_3GPP_AH_ACTIVE;
  526. }
  527. }
  528. return thr;
  529. }
  530. #ifndef calc_thr_3gpp
  531. static void calc_thr_3gpp(const FFPsyWindowInfo *wi, const int num_bands, AacPsyChannel *pch,
  532. const uint8_t *band_sizes, const float *coefs, const int cutoff)
  533. {
  534. int i, w, g;
  535. int start = 0, wstart = 0;
  536. for (w = 0; w < wi->num_windows*16; w += 16) {
  537. wstart = 0;
  538. for (g = 0; g < num_bands; g++) {
  539. AacPsyBand *band = &pch->band[w+g];
  540. float form_factor = 0.0f;
  541. float Temp;
  542. band->energy = 0.0f;
  543. if (wstart < cutoff) {
  544. for (i = 0; i < band_sizes[g]; i++) {
  545. band->energy += coefs[start+i] * coefs[start+i];
  546. form_factor += sqrtf(fabs(coefs[start+i]));
  547. }
  548. }
  549. Temp = band->energy > 0 ? sqrtf((float)band_sizes[g] / band->energy) : 0;
  550. band->thr = band->energy * 0.001258925f;
  551. band->nz_lines = form_factor * sqrtf(Temp);
  552. start += band_sizes[g];
  553. wstart += band_sizes[g];
  554. }
  555. }
  556. }
  557. #endif /* calc_thr_3gpp */
  558. #ifndef psy_hp_filter
  559. static void psy_hp_filter(const float *firbuf, float *hpfsmpl, const float *psy_fir_coeffs)
  560. {
  561. int i, j;
  562. for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) {
  563. float sum1, sum2;
  564. sum1 = firbuf[i + (PSY_LAME_FIR_LEN - 1) / 2];
  565. sum2 = 0.0;
  566. for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) {
  567. sum1 += psy_fir_coeffs[j] * (firbuf[i + j] + firbuf[i + PSY_LAME_FIR_LEN - j]);
  568. sum2 += psy_fir_coeffs[j + 1] * (firbuf[i + j + 1] + firbuf[i + PSY_LAME_FIR_LEN - j - 1]);
  569. }
  570. /* NOTE: The LAME psymodel expects it's input in the range -32768 to 32768.
  571. * Tuning this for normalized floats would be difficult. */
  572. hpfsmpl[i] = (sum1 + sum2) * 32768.0f;
  573. }
  574. }
  575. #endif /* psy_hp_filter */
  576. /**
  577. * Calculate band thresholds as suggested in 3GPP TS26.403
  578. */
  579. static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel,
  580. const float *coefs, const FFPsyWindowInfo *wi)
  581. {
  582. AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
  583. AacPsyChannel *pch = &pctx->ch[channel];
  584. int i, w, g;
  585. float desired_bits, desired_pe, delta_pe, reduction= NAN, spread_en[128] = {0};
  586. float a = 0.0f, active_lines = 0.0f, norm_fac = 0.0f;
  587. float pe = pctx->chan_bitrate > 32000 ? 0.0f : FFMAX(50.0f, 100.0f - pctx->chan_bitrate * 100.0f / 32000.0f);
  588. const int num_bands = ctx->num_bands[wi->num_windows == 8];
  589. const uint8_t *band_sizes = ctx->bands[wi->num_windows == 8];
  590. AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8];
  591. const float avoid_hole_thr = wi->num_windows == 8 ? PSY_3GPP_AH_THR_SHORT : PSY_3GPP_AH_THR_LONG;
  592. const int bandwidth = ctx->cutoff ? ctx->cutoff : AAC_CUTOFF(ctx->avctx);
  593. const int cutoff = bandwidth * 2048 / wi->num_windows / ctx->avctx->sample_rate;
  594. //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation"
  595. calc_thr_3gpp(wi, num_bands, pch, band_sizes, coefs, cutoff);
  596. //modify thresholds and energies - spread, threshold in quiet, pre-echo control
  597. for (w = 0; w < wi->num_windows*16; w += 16) {
  598. AacPsyBand *bands = &pch->band[w];
  599. /* 5.4.2.3 "Spreading" & 5.4.3 "Spread Energy Calculation" */
  600. spread_en[0] = bands[0].energy;
  601. for (g = 1; g < num_bands; g++) {
  602. bands[g].thr = FFMAX(bands[g].thr, bands[g-1].thr * coeffs[g].spread_hi[0]);
  603. spread_en[w+g] = FFMAX(bands[g].energy, spread_en[w+g-1] * coeffs[g].spread_hi[1]);
  604. }
  605. for (g = num_bands - 2; g >= 0; g--) {
  606. bands[g].thr = FFMAX(bands[g].thr, bands[g+1].thr * coeffs[g].spread_low[0]);
  607. spread_en[w+g] = FFMAX(spread_en[w+g], spread_en[w+g+1] * coeffs[g].spread_low[1]);
  608. }
  609. //5.4.2.4 "Threshold in quiet"
  610. for (g = 0; g < num_bands; g++) {
  611. AacPsyBand *band = &bands[g];
  612. band->thr_quiet = band->thr = FFMAX(band->thr, coeffs[g].ath);
  613. //5.4.2.5 "Pre-echo control"
  614. if (!(wi->window_type[0] == LONG_STOP_SEQUENCE || (wi->window_type[1] == LONG_START_SEQUENCE && !w)))
  615. band->thr = FFMAX(PSY_3GPP_RPEMIN*band->thr, FFMIN(band->thr,
  616. PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet));
  617. /* 5.6.1.3.1 "Preparatory steps of the perceptual entropy calculation" */
  618. pe += calc_pe_3gpp(band);
  619. a += band->pe_const;
  620. active_lines += band->active_lines;
  621. /* 5.6.1.3.3 "Selection of the bands for avoidance of holes" */
  622. if (spread_en[w+g] * avoid_hole_thr > band->energy || coeffs[g].min_snr > 1.0f)
  623. band->avoid_holes = PSY_3GPP_AH_NONE;
  624. else
  625. band->avoid_holes = PSY_3GPP_AH_INACTIVE;
  626. }
  627. }
  628. /* 5.6.1.3.2 "Calculation of the desired perceptual entropy" */
  629. ctx->ch[channel].entropy = pe;
  630. if (ctx->avctx->flags & CODEC_FLAG_QSCALE) {
  631. /* (2.5 * 120) achieves almost transparent rate, and we want to give
  632. * ample room downwards, so we make that equivalent to QSCALE=2.4
  633. */
  634. desired_pe = pe * (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120) / (2 * 2.5f * 120.0f);
  635. desired_bits = FFMIN(2560, PSY_3GPP_PE_TO_BITS(desired_pe));
  636. desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); // reflect clipping
  637. /* PE slope smoothing */
  638. if (ctx->bitres.bits > 0) {
  639. desired_bits = FFMIN(2560, PSY_3GPP_PE_TO_BITS(desired_pe));
  640. desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); // reflect clipping
  641. }
  642. pctx->pe.max = FFMAX(pe, pctx->pe.max);
  643. pctx->pe.min = FFMIN(pe, pctx->pe.min);
  644. } else {
  645. desired_bits = calc_bit_demand(pctx, pe, ctx->bitres.bits, ctx->bitres.size, wi->num_windows == 8);
  646. desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits);
  647. /* NOTE: PE correction is kept simple. During initial testing it had very
  648. * little effect on the final bitrate. Probably a good idea to come
  649. * back and do more testing later.
  650. */
  651. if (ctx->bitres.bits > 0)
  652. desired_pe *= av_clipf(pctx->pe.previous / PSY_3GPP_BITS_TO_PE(ctx->bitres.bits),
  653. 0.85f, 1.15f);
  654. }
  655. pctx->pe.previous = PSY_3GPP_BITS_TO_PE(desired_bits);
  656. ctx->bitres.alloc = desired_bits;
  657. if (desired_pe < pe) {
  658. /* 5.6.1.3.4 "First Estimation of the reduction value" */
  659. for (w = 0; w < wi->num_windows*16; w += 16) {
  660. reduction = calc_reduction_3gpp(a, desired_pe, pe, active_lines);
  661. pe = 0.0f;
  662. a = 0.0f;
  663. active_lines = 0.0f;
  664. for (g = 0; g < num_bands; g++) {
  665. AacPsyBand *band = &pch->band[w+g];
  666. band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
  667. /* recalculate PE */
  668. pe += calc_pe_3gpp(band);
  669. a += band->pe_const;
  670. active_lines += band->active_lines;
  671. }
  672. }
  673. /* 5.6.1.3.5 "Second Estimation of the reduction value" */
  674. for (i = 0; i < 2; i++) {
  675. float pe_no_ah = 0.0f, desired_pe_no_ah;
  676. active_lines = a = 0.0f;
  677. for (w = 0; w < wi->num_windows*16; w += 16) {
  678. for (g = 0; g < num_bands; g++) {
  679. AacPsyBand *band = &pch->band[w+g];
  680. if (band->avoid_holes != PSY_3GPP_AH_ACTIVE) {
  681. pe_no_ah += band->pe;
  682. a += band->pe_const;
  683. active_lines += band->active_lines;
  684. }
  685. }
  686. }
  687. desired_pe_no_ah = FFMAX(desired_pe - (pe - pe_no_ah), 0.0f);
  688. if (active_lines > 0.0f)
  689. reduction = calc_reduction_3gpp(a, desired_pe_no_ah, pe_no_ah, active_lines);
  690. pe = 0.0f;
  691. for (w = 0; w < wi->num_windows*16; w += 16) {
  692. for (g = 0; g < num_bands; g++) {
  693. AacPsyBand *band = &pch->band[w+g];
  694. if (active_lines > 0.0f)
  695. band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
  696. pe += calc_pe_3gpp(band);
  697. if (band->thr > 0.0f)
  698. band->norm_fac = band->active_lines / band->thr;
  699. else
  700. band->norm_fac = 0.0f;
  701. norm_fac += band->norm_fac;
  702. }
  703. }
  704. delta_pe = desired_pe - pe;
  705. if (fabs(delta_pe) > 0.05f * desired_pe)
  706. break;
  707. }
  708. if (pe < 1.15f * desired_pe) {
  709. /* 6.6.1.3.6 "Final threshold modification by linearization" */
  710. norm_fac = 1.0f / norm_fac;
  711. for (w = 0; w < wi->num_windows*16; w += 16) {
  712. for (g = 0; g < num_bands; g++) {
  713. AacPsyBand *band = &pch->band[w+g];
  714. if (band->active_lines > 0.5f) {
  715. float delta_sfb_pe = band->norm_fac * norm_fac * delta_pe;
  716. float thr = band->thr;
  717. thr *= exp2f(delta_sfb_pe / band->active_lines);
  718. if (thr > coeffs[g].min_snr * band->energy && band->avoid_holes == PSY_3GPP_AH_INACTIVE)
  719. thr = FFMAX(band->thr, coeffs[g].min_snr * band->energy);
  720. band->thr = thr;
  721. }
  722. }
  723. }
  724. } else {
  725. /* 5.6.1.3.7 "Further perceptual entropy reduction" */
  726. g = num_bands;
  727. while (pe > desired_pe && g--) {
  728. for (w = 0; w < wi->num_windows*16; w+= 16) {
  729. AacPsyBand *band = &pch->band[w+g];
  730. if (band->avoid_holes != PSY_3GPP_AH_NONE && coeffs[g].min_snr < PSY_SNR_1DB) {
  731. coeffs[g].min_snr = PSY_SNR_1DB;
  732. band->thr = band->energy * PSY_SNR_1DB;
  733. pe += band->active_lines * 1.5f - band->pe;
  734. }
  735. }
  736. }
  737. /* TODO: allow more holes (unused without mid/side) */
  738. }
  739. }
  740. for (w = 0; w < wi->num_windows*16; w += 16) {
  741. for (g = 0; g < num_bands; g++) {
  742. AacPsyBand *band = &pch->band[w+g];
  743. FFPsyBand *psy_band = &ctx->ch[channel].psy_bands[w+g];
  744. psy_band->threshold = band->thr;
  745. psy_band->energy = band->energy;
  746. psy_band->spread = band->active_lines * 2.0f / band_sizes[g];
  747. psy_band->bits = PSY_3GPP_PE_TO_BITS(band->pe);
  748. }
  749. }
  750. memcpy(pch->prev_band, pch->band, sizeof(pch->band));
  751. }
  752. static void psy_3gpp_analyze(FFPsyContext *ctx, int channel,
  753. const float **coeffs, const FFPsyWindowInfo *wi)
  754. {
  755. int ch;
  756. FFPsyChannelGroup *group = ff_psy_find_group(ctx, channel);
  757. for (ch = 0; ch < group->num_ch; ch++)
  758. psy_3gpp_analyze_channel(ctx, channel + ch, coeffs[ch], &wi[ch]);
  759. }
  760. static av_cold void psy_3gpp_end(FFPsyContext *apc)
  761. {
  762. AacPsyContext *pctx = (AacPsyContext*) apc->model_priv_data;
  763. av_freep(&pctx->ch);
  764. av_freep(&apc->model_priv_data);
  765. }
  766. static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock)
  767. {
  768. int blocktype = ONLY_LONG_SEQUENCE;
  769. if (uselongblock) {
  770. if (ctx->next_window_seq == EIGHT_SHORT_SEQUENCE)
  771. blocktype = LONG_STOP_SEQUENCE;
  772. } else {
  773. blocktype = EIGHT_SHORT_SEQUENCE;
  774. if (ctx->next_window_seq == ONLY_LONG_SEQUENCE)
  775. ctx->next_window_seq = LONG_START_SEQUENCE;
  776. if (ctx->next_window_seq == LONG_STOP_SEQUENCE)
  777. ctx->next_window_seq = EIGHT_SHORT_SEQUENCE;
  778. }
  779. wi->window_type[0] = ctx->next_window_seq;
  780. ctx->next_window_seq = blocktype;
  781. }
  782. static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx, const float *audio,
  783. const float *la, int channel, int prev_type)
  784. {
  785. AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
  786. AacPsyChannel *pch = &pctx->ch[channel];
  787. int grouping = 0;
  788. int uselongblock = 1;
  789. int attacks[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
  790. float clippings[AAC_NUM_BLOCKS_SHORT];
  791. int i;
  792. FFPsyWindowInfo wi = { { 0 } };
  793. if (la) {
  794. float hpfsmpl[AAC_BLOCK_SIZE_LONG];
  795. float const *pf = hpfsmpl;
  796. float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
  797. float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
  798. float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
  799. const float *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN);
  800. int att_sum = 0;
  801. /* LAME comment: apply high pass filter of fs/4 */
  802. psy_hp_filter(firbuf, hpfsmpl, psy_fir_coeffs);
  803. /* Calculate the energies of each sub-shortblock */
  804. for (i = 0; i < PSY_LAME_NUM_SUBBLOCKS; i++) {
  805. energy_subshort[i] = pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS)];
  806. assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)] > 0);
  807. attack_intensity[i] = energy_subshort[i] / pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)];
  808. energy_short[0] += energy_subshort[i];
  809. }
  810. for (i = 0; i < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; i++) {
  811. float const *const pfe = pf + AAC_BLOCK_SIZE_LONG / (AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS);
  812. float p = 1.0f;
  813. for (; pf < pfe; pf++)
  814. p = FFMAX(p, fabsf(*pf));
  815. pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p;
  816. energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p;
  817. /* NOTE: The indexes below are [i + 3 - 2] in the LAME source.
  818. * Obviously the 3 and 2 have some significance, or this would be just [i + 1]
  819. * (which is what we use here). What the 3 stands for is ambiguous, as it is both
  820. * number of short blocks, and the number of sub-short blocks.
  821. * It seems that LAME is comparing each sub-block to sub-block + 1 in the
  822. * previous block.
  823. */
  824. if (p > energy_subshort[i + 1])
  825. p = p / energy_subshort[i + 1];
  826. else if (energy_subshort[i + 1] > p * 10.0f)
  827. p = energy_subshort[i + 1] / (p * 10.0f);
  828. else
  829. p = 0.0;
  830. attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p;
  831. }
  832. /* compare energy between sub-short blocks */
  833. for (i = 0; i < (AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS; i++)
  834. if (!attacks[i / PSY_LAME_NUM_SUBBLOCKS])
  835. if (attack_intensity[i] > pch->attack_threshold)
  836. attacks[i / PSY_LAME_NUM_SUBBLOCKS] = (i % PSY_LAME_NUM_SUBBLOCKS) + 1;
  837. /* should have energy change between short blocks, in order to avoid periodic signals */
  838. /* Good samples to show the effect are Trumpet test songs */
  839. /* GB: tuned (1) to avoid too many short blocks for test sample TRUMPET */
  840. /* RH: tuned (2) to let enough short blocks through for test sample FSOL and SNAPS */
  841. for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) {
  842. float const u = energy_short[i - 1];
  843. float const v = energy_short[i];
  844. float const m = FFMAX(u, v);
  845. if (m < 40000) { /* (2) */
  846. if (u < 1.7f * v && v < 1.7f * u) { /* (1) */
  847. if (i == 1 && attacks[0] < attacks[i])
  848. attacks[0] = 0;
  849. attacks[i] = 0;
  850. }
  851. }
  852. att_sum += attacks[i];
  853. }
  854. if (attacks[0] <= pch->prev_attack)
  855. attacks[0] = 0;
  856. att_sum += attacks[0];
  857. /* 3 below indicates the previous attack happened in the last sub-block of the previous sequence */
  858. if (pch->prev_attack == 3 || att_sum) {
  859. uselongblock = 0;
  860. for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++)
  861. if (attacks[i] && attacks[i-1])
  862. attacks[i] = 0;
  863. }
  864. } else {
  865. /* We have no lookahead info, so just use same type as the previous sequence. */
  866. uselongblock = !(prev_type == EIGHT_SHORT_SEQUENCE);
  867. }
  868. lame_apply_block_type(pch, &wi, uselongblock);
  869. /* Calculate input sample maximums and evaluate clipping risk */
  870. if (audio) {
  871. for (i = 0; i < AAC_NUM_BLOCKS_SHORT; i++) {
  872. const float *wbuf = audio + i * AAC_BLOCK_SIZE_SHORT;
  873. float max = 0;
  874. int j;
  875. for (j = 0; j < AAC_BLOCK_SIZE_SHORT; j++)
  876. max = FFMAX(max, fabsf(wbuf[j]));
  877. clippings[i] = max;
  878. }
  879. } else {
  880. for (i = 0; i < 8; i++)
  881. clippings[i] = 0;
  882. }
  883. wi.window_type[1] = prev_type;
  884. if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
  885. float clipping = 0.0f;
  886. wi.num_windows = 1;
  887. wi.grouping[0] = 1;
  888. if (wi.window_type[0] == LONG_START_SEQUENCE)
  889. wi.window_shape = 0;
  890. else
  891. wi.window_shape = 1;
  892. for (i = 0; i < 8; i++)
  893. clipping = FFMAX(clipping, clippings[i]);
  894. wi.clipping[0] = clipping;
  895. } else {
  896. int lastgrp = 0;
  897. wi.num_windows = 8;
  898. wi.window_shape = 0;
  899. for (i = 0; i < 8; i++) {
  900. if (!((pch->next_grouping >> i) & 1))
  901. lastgrp = i;
  902. wi.grouping[lastgrp]++;
  903. }
  904. for (i = 0; i < 8; i += wi.grouping[i]) {
  905. int w;
  906. float clipping = 0.0f;
  907. for (w = 0; w < wi.grouping[i] && !clipping; w++)
  908. clipping = FFMAX(clipping, clippings[i+w]);
  909. wi.clipping[i] = clipping;
  910. }
  911. }
  912. /* Determine grouping, based on the location of the first attack, and save for
  913. * the next frame.
  914. * FIXME: Move this to analysis.
  915. * TODO: Tune groupings depending on attack location
  916. * TODO: Handle more than one attack in a group
  917. */
  918. for (i = 0; i < 9; i++) {
  919. if (attacks[i]) {
  920. grouping = i;
  921. break;
  922. }
  923. }
  924. pch->next_grouping = window_grouping[grouping];
  925. pch->prev_attack = attacks[8];
  926. return wi;
  927. }
  928. const FFPsyModel ff_aac_psy_model =
  929. {
  930. .name = "3GPP TS 26.403-inspired model",
  931. .init = psy_3gpp_init,
  932. .window = psy_lame_window,
  933. .analyze = psy_3gpp_analyze,
  934. .end = psy_3gpp_end,
  935. };