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.

1036 lines
40KB

  1. /*
  2. * AAC coefficients encoder
  3. * Copyright (C) 2008-2009 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 libavcodec/aaccoder.c
  23. * AAC coefficients encoder
  24. */
  25. /***********************************
  26. * TODOs:
  27. * speedup quantizer selection
  28. * add sane pulse detection
  29. ***********************************/
  30. #include "avcodec.h"
  31. #include "put_bits.h"
  32. #include "aac.h"
  33. #include "aacenc.h"
  34. #include "aactab.h"
  35. /** bits needed to code codebook run value for long windows */
  36. static const uint8_t run_value_bits_long[64] = {
  37. 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
  38. 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 10,
  39. 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
  40. 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 15
  41. };
  42. /** bits needed to code codebook run value for short windows */
  43. static const uint8_t run_value_bits_short[16] = {
  44. 3, 3, 3, 3, 3, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 9
  45. };
  46. static const uint8_t *run_value_bits[2] = {
  47. run_value_bits_long, run_value_bits_short
  48. };
  49. /**
  50. * Quantize one coefficient.
  51. * @return absolute value of the quantized coefficient
  52. * @see 3GPP TS26.403 5.6.2 "Scalefactor determination"
  53. */
  54. static av_always_inline int quant(float coef, const float Q)
  55. {
  56. return pow(coef * Q, 0.75) + 0.4054;
  57. }
  58. static void quantize_bands(int (*out)[2], const float *in, const float *scaled,
  59. int size, float Q34, int is_signed, int maxval)
  60. {
  61. int i;
  62. double qc;
  63. for (i = 0; i < size; i++) {
  64. qc = scaled[i] * Q34;
  65. out[i][0] = (int)FFMIN((int)qc, maxval);
  66. out[i][1] = (int)FFMIN((int)(qc + 0.4054), maxval);
  67. if (is_signed && in[i] < 0.0f) {
  68. out[i][0] = -out[i][0];
  69. out[i][1] = -out[i][1];
  70. }
  71. }
  72. }
  73. static void abs_pow34_v(float *out, const float *in, const int size)
  74. {
  75. #ifndef USE_REALLY_FULL_SEARCH
  76. int i;
  77. for (i = 0; i < size; i++)
  78. out[i] = pow(fabsf(in[i]), 0.75);
  79. #endif /* USE_REALLY_FULL_SEARCH */
  80. }
  81. static av_always_inline int quant2(float coef, const float Q)
  82. {
  83. return pow(coef * Q, 0.75);
  84. }
  85. static const uint8_t aac_cb_range [12] = {0, 3, 3, 3, 3, 9, 9, 8, 8, 13, 13, 17};
  86. static const uint8_t aac_cb_maxval[12] = {0, 1, 1, 2, 2, 4, 4, 7, 7, 12, 12, 16};
  87. /**
  88. * Calculate rate distortion cost for quantizing with given codebook
  89. *
  90. * @return quantization distortion
  91. */
  92. static float quantize_band_cost(struct AACEncContext *s, const float *in,
  93. const float *scaled, int size, int scale_idx,
  94. int cb, const float lambda, const float uplim,
  95. int *bits)
  96. {
  97. const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
  98. const float Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
  99. const float CLIPPED_ESCAPE = 165140.0f*IQ;
  100. int i, j, k;
  101. float cost = 0;
  102. const int dim = cb < FIRST_PAIR_BT ? 4 : 2;
  103. int resbits = 0;
  104. #ifndef USE_REALLY_FULL_SEARCH
  105. const float Q34 = pow(Q, 0.75);
  106. const int range = aac_cb_range[cb];
  107. const int maxval = aac_cb_maxval[cb];
  108. int offs[4];
  109. #endif /* USE_REALLY_FULL_SEARCH */
  110. if (!cb) {
  111. for (i = 0; i < size; i++)
  112. cost += in[i]*in[i]*lambda;
  113. return cost;
  114. }
  115. #ifndef USE_REALLY_FULL_SEARCH
  116. offs[0] = 1;
  117. for (i = 1; i < dim; i++)
  118. offs[i] = offs[i-1]*range;
  119. quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
  120. #endif /* USE_REALLY_FULL_SEARCH */
  121. for (i = 0; i < size; i += dim) {
  122. float mincost;
  123. int minidx = 0;
  124. int minbits = 0;
  125. const float *vec;
  126. #ifndef USE_REALLY_FULL_SEARCH
  127. int (*quants)[2] = &s->qcoefs[i];
  128. mincost = 0.0f;
  129. for (j = 0; j < dim; j++)
  130. mincost += in[i+j]*in[i+j]*lambda;
  131. minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
  132. minbits = ff_aac_spectral_bits[cb-1][minidx];
  133. mincost += minbits;
  134. for (j = 0; j < (1<<dim); j++) {
  135. float rd = 0.0f;
  136. int curbits;
  137. int curidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
  138. int same = 0;
  139. for (k = 0; k < dim; k++) {
  140. if ((j & (1 << k)) && quants[k][0] == quants[k][1]) {
  141. same = 1;
  142. break;
  143. }
  144. }
  145. if (same)
  146. continue;
  147. for (k = 0; k < dim; k++)
  148. curidx += quants[k][!!(j & (1 << k))] * offs[dim - 1 - k];
  149. curbits = ff_aac_spectral_bits[cb-1][curidx];
  150. vec = &ff_aac_codebook_vectors[cb-1][curidx*dim];
  151. #else
  152. mincost = INFINITY;
  153. vec = ff_aac_codebook_vectors[cb-1];
  154. for (j = 0; j < ff_aac_spectral_sizes[cb-1]; j++, vec += dim) {
  155. float rd = 0.0f;
  156. int curbits = ff_aac_spectral_bits[cb-1][j];
  157. #endif /* USE_REALLY_FULL_SEARCH */
  158. if (IS_CODEBOOK_UNSIGNED(cb)) {
  159. for (k = 0; k < dim; k++) {
  160. float t = fabsf(in[i+k]);
  161. float di;
  162. //do not code with escape sequence small values
  163. if (vec[k] == 64.0f && t < 39.0f*IQ) {
  164. rd = INFINITY;
  165. break;
  166. }
  167. if (vec[k] == 64.0f) { //FIXME: slow
  168. if (t >= CLIPPED_ESCAPE) {
  169. di = t - CLIPPED_ESCAPE;
  170. curbits += 21;
  171. } else {
  172. int c = av_clip(quant(t, Q), 0, 8191);
  173. di = t - c*cbrt(c)*IQ;
  174. curbits += av_log2(c)*2 - 4 + 1;
  175. }
  176. } else {
  177. di = t - vec[k]*IQ;
  178. }
  179. if (vec[k] != 0.0f)
  180. curbits++;
  181. rd += di*di*lambda;
  182. }
  183. } else {
  184. for (k = 0; k < dim; k++) {
  185. float di = in[i+k] - vec[k]*IQ;
  186. rd += di*di*lambda;
  187. }
  188. }
  189. rd += curbits;
  190. if (rd < mincost) {
  191. mincost = rd;
  192. minidx = j;
  193. minbits = curbits;
  194. }
  195. }
  196. cost += mincost;
  197. resbits += minbits;
  198. if (cost >= uplim)
  199. return uplim;
  200. }
  201. if (bits)
  202. *bits = resbits;
  203. return cost;
  204. }
  205. static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
  206. const float *in, int size, int scale_idx,
  207. int cb, const float lambda)
  208. {
  209. const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
  210. const float Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
  211. const float CLIPPED_ESCAPE = 165140.0f*IQ;
  212. const int dim = (cb < FIRST_PAIR_BT) ? 4 : 2;
  213. int i, j, k;
  214. #ifndef USE_REALLY_FULL_SEARCH
  215. const float Q34 = pow(Q, 0.75);
  216. const int range = aac_cb_range[cb];
  217. const int maxval = aac_cb_maxval[cb];
  218. int offs[4];
  219. float *scaled = s->scoefs;
  220. #endif /* USE_REALLY_FULL_SEARCH */
  221. //START_TIMER
  222. if (!cb)
  223. return;
  224. #ifndef USE_REALLY_FULL_SEARCH
  225. offs[0] = 1;
  226. for (i = 1; i < dim; i++)
  227. offs[i] = offs[i-1]*range;
  228. abs_pow34_v(scaled, in, size);
  229. quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
  230. #endif /* USE_REALLY_FULL_SEARCH */
  231. for (i = 0; i < size; i += dim) {
  232. float mincost;
  233. int minidx = 0;
  234. int minbits = 0;
  235. const float *vec;
  236. #ifndef USE_REALLY_FULL_SEARCH
  237. int (*quants)[2] = &s->qcoefs[i];
  238. mincost = 0.0f;
  239. for (j = 0; j < dim; j++)
  240. mincost += in[i+j]*in[i+j]*lambda;
  241. minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
  242. minbits = ff_aac_spectral_bits[cb-1][minidx];
  243. mincost += minbits;
  244. for (j = 0; j < (1<<dim); j++) {
  245. float rd = 0.0f;
  246. int curbits;
  247. int curidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
  248. int same = 0;
  249. for (k = 0; k < dim; k++) {
  250. if ((j & (1 << k)) && quants[k][0] == quants[k][1]) {
  251. same = 1;
  252. break;
  253. }
  254. }
  255. if (same)
  256. continue;
  257. for (k = 0; k < dim; k++)
  258. curidx += quants[k][!!(j & (1 << k))] * offs[dim - 1 - k];
  259. curbits = ff_aac_spectral_bits[cb-1][curidx];
  260. vec = &ff_aac_codebook_vectors[cb-1][curidx*dim];
  261. #else
  262. vec = ff_aac_codebook_vectors[cb-1];
  263. mincost = INFINITY;
  264. for (j = 0; j < ff_aac_spectral_sizes[cb-1]; j++, vec += dim) {
  265. float rd = 0.0f;
  266. int curbits = ff_aac_spectral_bits[cb-1][j];
  267. int curidx = j;
  268. #endif /* USE_REALLY_FULL_SEARCH */
  269. if (IS_CODEBOOK_UNSIGNED(cb)) {
  270. for (k = 0; k < dim; k++) {
  271. float t = fabsf(in[i+k]);
  272. float di;
  273. //do not code with escape sequence small values
  274. if (vec[k] == 64.0f && t < 39.0f*IQ) {
  275. rd = INFINITY;
  276. break;
  277. }
  278. if (vec[k] == 64.0f) { //FIXME: slow
  279. if (t >= CLIPPED_ESCAPE) {
  280. di = t - CLIPPED_ESCAPE;
  281. curbits += 21;
  282. } else {
  283. int c = av_clip(quant(t, Q), 0, 8191);
  284. di = t - c*cbrt(c)*IQ;
  285. curbits += av_log2(c)*2 - 4 + 1;
  286. }
  287. } else {
  288. di = t - vec[k]*IQ;
  289. }
  290. if (vec[k] != 0.0f)
  291. curbits++;
  292. rd += di*di*lambda;
  293. }
  294. } else {
  295. for (k = 0; k < dim; k++) {
  296. float di = in[i+k] - vec[k]*IQ;
  297. rd += di*di*lambda;
  298. }
  299. }
  300. rd += curbits;
  301. if (rd < mincost) {
  302. mincost = rd;
  303. minidx = curidx;
  304. minbits = curbits;
  305. }
  306. }
  307. put_bits(pb, ff_aac_spectral_bits[cb-1][minidx], ff_aac_spectral_codes[cb-1][minidx]);
  308. if (IS_CODEBOOK_UNSIGNED(cb))
  309. for (j = 0; j < dim; j++)
  310. if (ff_aac_codebook_vectors[cb-1][minidx*dim+j] != 0.0f)
  311. put_bits(pb, 1, in[i+j] < 0.0f);
  312. if (cb == ESC_BT) {
  313. for (j = 0; j < 2; j++) {
  314. if (ff_aac_codebook_vectors[cb-1][minidx*2+j] == 64.0f) {
  315. int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
  316. int len = av_log2(coef);
  317. put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
  318. put_bits(pb, len, coef & ((1 << len) - 1));
  319. }
  320. }
  321. }
  322. }
  323. //STOP_TIMER("quantize_and_encode")
  324. }
  325. /**
  326. * structure used in optimal codebook search
  327. */
  328. typedef struct BandCodingPath {
  329. int prev_idx; ///< pointer to the previous path point
  330. int codebook; ///< codebook for coding band run
  331. float cost; ///< path cost
  332. int run;
  333. } BandCodingPath;
  334. /**
  335. * Encode band info for single window group bands.
  336. */
  337. static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
  338. int win, int group_len, const float lambda)
  339. {
  340. BandCodingPath path[120][12];
  341. int w, swb, cb, start, start2, size;
  342. int i, j;
  343. const int max_sfb = sce->ics.max_sfb;
  344. const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
  345. const int run_esc = (1 << run_bits) - 1;
  346. int idx, ppos, count;
  347. int stackrun[120], stackcb[120], stack_len;
  348. float next_minrd = INFINITY;
  349. int next_mincb = 0;
  350. abs_pow34_v(s->scoefs, sce->coeffs, 1024);
  351. start = win*128;
  352. for (cb = 0; cb < 12; cb++) {
  353. path[0][cb].cost = 0.0f;
  354. path[0][cb].prev_idx = -1;
  355. path[0][cb].run = 0;
  356. }
  357. for (swb = 0; swb < max_sfb; swb++) {
  358. start2 = start;
  359. size = sce->ics.swb_sizes[swb];
  360. if (sce->zeroes[win*16 + swb]) {
  361. for (cb = 0; cb < 12; cb++) {
  362. path[swb+1][cb].prev_idx = cb;
  363. path[swb+1][cb].cost = path[swb][cb].cost;
  364. path[swb+1][cb].run = path[swb][cb].run + 1;
  365. }
  366. } else {
  367. float minrd = next_minrd;
  368. int mincb = next_mincb;
  369. next_minrd = INFINITY;
  370. next_mincb = 0;
  371. for (cb = 0; cb < 12; cb++) {
  372. float cost_stay_here, cost_get_here;
  373. float rd = 0.0f;
  374. for (w = 0; w < group_len; w++) {
  375. FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(win+w)*16+swb];
  376. rd += quantize_band_cost(s, sce->coeffs + start + w*128,
  377. s->scoefs + start + w*128, size,
  378. sce->sf_idx[(win+w)*16+swb], cb,
  379. lambda / band->threshold, INFINITY, NULL);
  380. }
  381. cost_stay_here = path[swb][cb].cost + rd;
  382. cost_get_here = minrd + rd + run_bits + 4;
  383. if ( run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
  384. != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
  385. cost_stay_here += run_bits;
  386. if (cost_get_here < cost_stay_here) {
  387. path[swb+1][cb].prev_idx = mincb;
  388. path[swb+1][cb].cost = cost_get_here;
  389. path[swb+1][cb].run = 1;
  390. } else {
  391. path[swb+1][cb].prev_idx = cb;
  392. path[swb+1][cb].cost = cost_stay_here;
  393. path[swb+1][cb].run = path[swb][cb].run + 1;
  394. }
  395. if (path[swb+1][cb].cost < next_minrd) {
  396. next_minrd = path[swb+1][cb].cost;
  397. next_mincb = cb;
  398. }
  399. }
  400. }
  401. start += sce->ics.swb_sizes[swb];
  402. }
  403. //convert resulting path from backward-linked list
  404. stack_len = 0;
  405. idx = 0;
  406. for (cb = 1; cb < 12; cb++)
  407. if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
  408. idx = cb;
  409. ppos = max_sfb;
  410. while (ppos > 0) {
  411. cb = idx;
  412. stackrun[stack_len] = path[ppos][cb].run;
  413. stackcb [stack_len] = cb;
  414. idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
  415. ppos -= path[ppos][cb].run;
  416. stack_len++;
  417. }
  418. //perform actual band info encoding
  419. start = 0;
  420. for (i = stack_len - 1; i >= 0; i--) {
  421. put_bits(&s->pb, 4, stackcb[i]);
  422. count = stackrun[i];
  423. memset(sce->zeroes + win*16 + start, !stackcb[i], count);
  424. //XXX: memset when band_type is also uint8_t
  425. for (j = 0; j < count; j++) {
  426. sce->band_type[win*16 + start] = stackcb[i];
  427. start++;
  428. }
  429. while (count >= run_esc) {
  430. put_bits(&s->pb, run_bits, run_esc);
  431. count -= run_esc;
  432. }
  433. put_bits(&s->pb, run_bits, count);
  434. }
  435. }
  436. typedef struct TrellisPath {
  437. float cost;
  438. int prev;
  439. int min_val;
  440. int max_val;
  441. } TrellisPath;
  442. static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
  443. SingleChannelElement *sce,
  444. const float lambda)
  445. {
  446. int q, w, w2, g, start = 0;
  447. int i;
  448. int idx;
  449. TrellisPath paths[256*121];
  450. int bandaddr[121];
  451. int minq;
  452. float mincost;
  453. for (i = 0; i < 256; i++) {
  454. paths[i].cost = 0.0f;
  455. paths[i].prev = -1;
  456. paths[i].min_val = i;
  457. paths[i].max_val = i;
  458. }
  459. for (i = 256; i < 256*121; i++) {
  460. paths[i].cost = INFINITY;
  461. paths[i].prev = -2;
  462. paths[i].min_val = INT_MAX;
  463. paths[i].max_val = 0;
  464. }
  465. idx = 256;
  466. abs_pow34_v(s->scoefs, sce->coeffs, 1024);
  467. for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
  468. start = w*128;
  469. for (g = 0; g < sce->ics.num_swb; g++) {
  470. const float *coefs = sce->coeffs + start;
  471. float qmin, qmax;
  472. int nz = 0;
  473. bandaddr[idx >> 8] = w * 16 + g;
  474. qmin = INT_MAX;
  475. qmax = 0.0f;
  476. for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
  477. FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
  478. if (band->energy <= band->threshold || band->threshold == 0.0f) {
  479. sce->zeroes[(w+w2)*16+g] = 1;
  480. continue;
  481. }
  482. sce->zeroes[(w+w2)*16+g] = 0;
  483. nz = 1;
  484. for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
  485. float t = fabsf(coefs[w2*128+i]);
  486. if (t > 0.0f)
  487. qmin = FFMIN(qmin, t);
  488. qmax = FFMAX(qmax, t);
  489. }
  490. }
  491. if (nz) {
  492. int minscale, maxscale;
  493. float minrd = INFINITY;
  494. //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
  495. minscale = av_clip_uint8(log2(qmin)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
  496. //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
  497. maxscale = av_clip_uint8(log2(qmax)*4 + 6 + SCALE_ONE_POS - SCALE_DIV_512);
  498. for (q = minscale; q < maxscale; q++) {
  499. float dists[12], dist;
  500. memset(dists, 0, sizeof(dists));
  501. for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
  502. FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
  503. int cb;
  504. for (cb = 0; cb <= ESC_BT; cb++)
  505. dists[cb] += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
  506. q, cb, lambda / band->threshold, INFINITY, NULL);
  507. }
  508. dist = dists[0];
  509. for (i = 1; i <= ESC_BT; i++)
  510. dist = FFMIN(dist, dists[i]);
  511. minrd = FFMIN(minrd, dist);
  512. for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, 256); i++) {
  513. float cost;
  514. int minv, maxv;
  515. if (isinf(paths[idx - 256 + i].cost))
  516. continue;
  517. cost = paths[idx - 256 + i].cost + dist
  518. + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
  519. minv = FFMIN(paths[idx - 256 + i].min_val, q);
  520. maxv = FFMAX(paths[idx - 256 + i].max_val, q);
  521. if (cost < paths[idx + q].cost && maxv-minv < SCALE_MAX_DIFF) {
  522. paths[idx + q].cost = cost;
  523. paths[idx + q].prev = idx - 256 + i;
  524. paths[idx + q].min_val = minv;
  525. paths[idx + q].max_val = maxv;
  526. }
  527. }
  528. }
  529. } else {
  530. for (q = 0; q < 256; q++) {
  531. if (!isinf(paths[idx - 256 + q].cost)) {
  532. paths[idx + q].cost = paths[idx - 256 + q].cost + 1;
  533. paths[idx + q].prev = idx - 256 + q;
  534. paths[idx + q].min_val = FFMIN(paths[idx - 256 + q].min_val, q);
  535. paths[idx + q].max_val = FFMAX(paths[idx - 256 + q].max_val, q);
  536. continue;
  537. }
  538. for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, 256); i++) {
  539. float cost;
  540. int minv, maxv;
  541. if (isinf(paths[idx - 256 + i].cost))
  542. continue;
  543. cost = paths[idx - 256 + i].cost + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
  544. minv = FFMIN(paths[idx - 256 + i].min_val, q);
  545. maxv = FFMAX(paths[idx - 256 + i].max_val, q);
  546. if (cost < paths[idx + q].cost && maxv-minv < SCALE_MAX_DIFF) {
  547. paths[idx + q].cost = cost;
  548. paths[idx + q].prev = idx - 256 + i;
  549. paths[idx + q].min_val = minv;
  550. paths[idx + q].max_val = maxv;
  551. }
  552. }
  553. }
  554. }
  555. sce->zeroes[w*16+g] = !nz;
  556. start += sce->ics.swb_sizes[g];
  557. idx += 256;
  558. }
  559. }
  560. idx -= 256;
  561. mincost = paths[idx].cost;
  562. minq = idx;
  563. for (i = 1; i < 256; i++) {
  564. if (paths[idx + i].cost < mincost) {
  565. mincost = paths[idx + i].cost;
  566. minq = idx + i;
  567. }
  568. }
  569. while (minq >= 256) {
  570. sce->sf_idx[bandaddr[minq>>8]] = minq & 0xFF;
  571. minq = paths[minq].prev;
  572. }
  573. //set the same quantizers inside window groups
  574. for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
  575. for (g = 0; g < sce->ics.num_swb; g++)
  576. for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
  577. sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
  578. }
  579. /**
  580. * two-loop quantizers search taken from ISO 13818-7 Appendix C
  581. */
  582. static void search_for_quantizers_twoloop(AVCodecContext *avctx,
  583. AACEncContext *s,
  584. SingleChannelElement *sce,
  585. const float lambda)
  586. {
  587. int start = 0, i, w, w2, g;
  588. int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels;
  589. float dists[128], uplims[128];
  590. int fflag, minscaler;
  591. int its = 0;
  592. int allz = 0;
  593. float minthr = INFINITY;
  594. //XXX: some heuristic to determine initial quantizers will reduce search time
  595. memset(dists, 0, sizeof(dists));
  596. //determine zero bands and upper limits
  597. for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
  598. for (g = 0; g < sce->ics.num_swb; g++) {
  599. int nz = 0;
  600. float uplim = 0.0f;
  601. for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
  602. FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
  603. uplim += band->threshold;
  604. if (band->energy <= band->threshold || band->threshold == 0.0f) {
  605. sce->zeroes[(w+w2)*16+g] = 1;
  606. continue;
  607. }
  608. nz = 1;
  609. }
  610. uplims[w*16+g] = uplim *512;
  611. sce->zeroes[w*16+g] = !nz;
  612. if (nz)
  613. minthr = FFMIN(minthr, uplim);
  614. allz = FFMAX(allz, nz);
  615. }
  616. }
  617. for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
  618. for (g = 0; g < sce->ics.num_swb; g++) {
  619. if (sce->zeroes[w*16+g]) {
  620. sce->sf_idx[w*16+g] = SCALE_ONE_POS;
  621. continue;
  622. }
  623. sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2(uplims[w*16+g]/minthr)*4,59);
  624. }
  625. }
  626. if (!allz)
  627. return;
  628. abs_pow34_v(s->scoefs, sce->coeffs, 1024);
  629. //perform two-loop search
  630. //outer loop - improve quality
  631. do {
  632. int tbits, qstep;
  633. minscaler = sce->sf_idx[0];
  634. //inner loop - quantize spectrum to fit into given number of bits
  635. qstep = its ? 1 : 32;
  636. do {
  637. int prev = -1;
  638. tbits = 0;
  639. fflag = 0;
  640. for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
  641. start = w*128;
  642. for (g = 0; g < sce->ics.num_swb; g++) {
  643. const float *coefs = sce->coeffs + start;
  644. const float *scaled = s->scoefs + start;
  645. int bits = 0;
  646. int cb;
  647. float mindist = INFINITY;
  648. int minbits = 0;
  649. if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218)
  650. continue;
  651. minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
  652. for (cb = 0; cb <= ESC_BT; cb++) {
  653. float dist = 0.0f;
  654. int bb = 0;
  655. for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
  656. int b;
  657. dist += quantize_band_cost(s, coefs + w2*128,
  658. scaled + w2*128,
  659. sce->ics.swb_sizes[g],
  660. sce->sf_idx[w*16+g],
  661. ESC_BT,
  662. lambda,
  663. INFINITY,
  664. &b);
  665. bb += b;
  666. }
  667. if (dist < mindist) {
  668. mindist = dist;
  669. minbits = bb;
  670. }
  671. }
  672. dists[w*16+g] = (mindist - minbits) / lambda;
  673. bits = minbits;
  674. if (prev != -1) {
  675. bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
  676. }
  677. tbits += bits;
  678. start += sce->ics.swb_sizes[g];
  679. prev = sce->sf_idx[w*16+g];
  680. }
  681. }
  682. if (tbits > destbits) {
  683. for (i = 0; i < 128; i++)
  684. if (sce->sf_idx[i] < 218 - qstep)
  685. sce->sf_idx[i] += qstep;
  686. } else {
  687. for (i = 0; i < 128; i++)
  688. if (sce->sf_idx[i] > 60 - qstep)
  689. sce->sf_idx[i] -= qstep;
  690. }
  691. qstep >>= 1;
  692. if (!qstep && tbits > destbits*1.02)
  693. qstep = 1;
  694. if (sce->sf_idx[0] >= 217)
  695. break;
  696. } while (qstep);
  697. fflag = 0;
  698. minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
  699. for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
  700. start = w*128;
  701. for (g = 0; g < sce->ics.num_swb; g++) {
  702. int prevsc = sce->sf_idx[w*16+g];
  703. if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60)
  704. sce->sf_idx[w*16+g]--;
  705. sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
  706. sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
  707. if (sce->sf_idx[w*16+g] != prevsc)
  708. fflag = 1;
  709. }
  710. }
  711. its++;
  712. } while (fflag && its < 10);
  713. }
  714. static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
  715. SingleChannelElement *sce,
  716. const float lambda)
  717. {
  718. int start = 0, i, w, w2, g;
  719. float uplim[128], maxq[128];
  720. int minq, maxsf;
  721. float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
  722. int last = 0, lastband = 0, curband = 0;
  723. float avg_energy = 0.0;
  724. if (sce->ics.num_windows == 1) {
  725. start = 0;
  726. for (i = 0; i < 1024; i++) {
  727. if (i - start >= sce->ics.swb_sizes[curband]) {
  728. start += sce->ics.swb_sizes[curband];
  729. curband++;
  730. }
  731. if (sce->coeffs[i]) {
  732. avg_energy += sce->coeffs[i] * sce->coeffs[i];
  733. last = i;
  734. lastband = curband;
  735. }
  736. }
  737. } else {
  738. for (w = 0; w < 8; w++) {
  739. const float *coeffs = sce->coeffs + w*128;
  740. start = 0;
  741. for (i = 0; i < 128; i++) {
  742. if (i - start >= sce->ics.swb_sizes[curband]) {
  743. start += sce->ics.swb_sizes[curband];
  744. curband++;
  745. }
  746. if (coeffs[i]) {
  747. avg_energy += coeffs[i] * coeffs[i];
  748. last = FFMAX(last, i);
  749. lastband = FFMAX(lastband, curband);
  750. }
  751. }
  752. }
  753. }
  754. last++;
  755. avg_energy /= last;
  756. if (avg_energy == 0.0f) {
  757. for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
  758. sce->sf_idx[i] = SCALE_ONE_POS;
  759. return;
  760. }
  761. for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
  762. start = w*128;
  763. for (g = 0; g < sce->ics.num_swb; g++) {
  764. float *coefs = sce->coeffs + start;
  765. const int size = sce->ics.swb_sizes[g];
  766. int start2 = start, end2 = start + size, peakpos = start;
  767. float maxval = -1, thr = 0.0f, t;
  768. maxq[w*16+g] = 0.0f;
  769. if (g > lastband) {
  770. maxq[w*16+g] = 0.0f;
  771. start += size;
  772. for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
  773. memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
  774. continue;
  775. }
  776. for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
  777. for (i = 0; i < size; i++) {
  778. float t = coefs[w2*128+i]*coefs[w2*128+i];
  779. maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
  780. thr += t;
  781. if (sce->ics.num_windows == 1 && maxval < t) {
  782. maxval = t;
  783. peakpos = start+i;
  784. }
  785. }
  786. }
  787. if (sce->ics.num_windows == 1) {
  788. start2 = FFMAX(peakpos - 2, start2);
  789. end2 = FFMIN(peakpos + 3, end2);
  790. } else {
  791. start2 -= start;
  792. end2 -= start;
  793. }
  794. start += size;
  795. thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
  796. t = 1.0 - (1.0 * start2 / last);
  797. uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
  798. }
  799. }
  800. memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
  801. abs_pow34_v(s->scoefs, sce->coeffs, 1024);
  802. for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
  803. start = w*128;
  804. for (g = 0; g < sce->ics.num_swb; g++) {
  805. const float *coefs = sce->coeffs + start;
  806. const float *scaled = s->scoefs + start;
  807. const int size = sce->ics.swb_sizes[g];
  808. int scf, prev_scf, step;
  809. int min_scf = 0, max_scf = 255;
  810. float curdiff;
  811. if (maxq[w*16+g] < 21.544) {
  812. sce->zeroes[w*16+g] = 1;
  813. start += size;
  814. continue;
  815. }
  816. sce->zeroes[w*16+g] = 0;
  817. scf = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2(1/maxq[w*16+g])*16/3, 60, 218);
  818. step = 16;
  819. for (;;) {
  820. float dist = 0.0f;
  821. int quant_max;
  822. for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
  823. int b;
  824. dist += quantize_band_cost(s, coefs + w2*128,
  825. scaled + w2*128,
  826. sce->ics.swb_sizes[g],
  827. scf,
  828. ESC_BT,
  829. lambda,
  830. INFINITY,
  831. &b);
  832. dist -= b;
  833. }
  834. dist *= 1.0f / 512.0f / lambda;
  835. quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[200 - scf + SCALE_ONE_POS - SCALE_DIV_512]);
  836. if (quant_max >= 8191) { // too much, return to the previous quantizer
  837. sce->sf_idx[w*16+g] = prev_scf;
  838. break;
  839. }
  840. prev_scf = scf;
  841. curdiff = fabsf(dist - uplim[w*16+g]);
  842. if (curdiff == 0.0f)
  843. step = 0;
  844. else
  845. step = fabsf(log2(curdiff));
  846. if (dist > uplim[w*16+g])
  847. step = -step;
  848. if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
  849. sce->sf_idx[w*16+g] = scf;
  850. break;
  851. }
  852. scf += step;
  853. if (step > 0)
  854. min_scf = scf;
  855. else
  856. max_scf = scf;
  857. }
  858. start += size;
  859. }
  860. }
  861. minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
  862. for (i = 1; i < 128; i++) {
  863. if (!sce->sf_idx[i])
  864. sce->sf_idx[i] = sce->sf_idx[i-1];
  865. else
  866. minq = FFMIN(minq, sce->sf_idx[i]);
  867. }
  868. if (minq == INT_MAX)
  869. minq = 0;
  870. minq = FFMIN(minq, SCALE_MAX_POS);
  871. maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
  872. for (i = 126; i >= 0; i--) {
  873. if (!sce->sf_idx[i])
  874. sce->sf_idx[i] = sce->sf_idx[i+1];
  875. sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
  876. }
  877. }
  878. static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
  879. SingleChannelElement *sce,
  880. const float lambda)
  881. {
  882. int start = 0, i, w, w2, g;
  883. int minq = 255;
  884. memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
  885. for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
  886. start = w*128;
  887. for (g = 0; g < sce->ics.num_swb; g++) {
  888. for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
  889. FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
  890. if (band->energy <= band->threshold) {
  891. sce->sf_idx[(w+w2)*16+g] = 218;
  892. sce->zeroes[(w+w2)*16+g] = 1;
  893. } else {
  894. sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2(band->threshold), 80, 218);
  895. sce->zeroes[(w+w2)*16+g] = 0;
  896. }
  897. minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
  898. }
  899. }
  900. }
  901. for (i = 0; i < 128; i++) {
  902. sce->sf_idx[i] = 140;
  903. //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
  904. }
  905. //set the same quantizers inside window groups
  906. for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
  907. for (g = 0; g < sce->ics.num_swb; g++)
  908. for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
  909. sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
  910. }
  911. static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
  912. const float lambda)
  913. {
  914. int start = 0, i, w, w2, g;
  915. float M[128], S[128];
  916. float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
  917. SingleChannelElement *sce0 = &cpe->ch[0];
  918. SingleChannelElement *sce1 = &cpe->ch[1];
  919. if (!cpe->common_window)
  920. return;
  921. for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
  922. for (g = 0; g < sce0->ics.num_swb; g++) {
  923. if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
  924. float dist1 = 0.0f, dist2 = 0.0f;
  925. for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
  926. FFPsyBand *band0 = &s->psy.psy_bands[(s->cur_channel+0)*PSY_MAX_BANDS+(w+w2)*16+g];
  927. FFPsyBand *band1 = &s->psy.psy_bands[(s->cur_channel+1)*PSY_MAX_BANDS+(w+w2)*16+g];
  928. float minthr = FFMIN(band0->threshold, band1->threshold);
  929. float maxthr = FFMAX(band0->threshold, band1->threshold);
  930. for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
  931. M[i] = (sce0->coeffs[start+w2*128+i]
  932. + sce1->coeffs[start+w2*128+i]) * 0.5;
  933. S[i] = sce0->coeffs[start+w2*128+i]
  934. - sce1->coeffs[start+w2*128+i];
  935. }
  936. abs_pow34_v(L34, sce0->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
  937. abs_pow34_v(R34, sce1->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
  938. abs_pow34_v(M34, M, sce0->ics.swb_sizes[g]);
  939. abs_pow34_v(S34, S, sce0->ics.swb_sizes[g]);
  940. dist1 += quantize_band_cost(s, sce0->coeffs + start + w2*128,
  941. L34,
  942. sce0->ics.swb_sizes[g],
  943. sce0->sf_idx[(w+w2)*16+g],
  944. sce0->band_type[(w+w2)*16+g],
  945. lambda / band0->threshold, INFINITY, NULL);
  946. dist1 += quantize_band_cost(s, sce1->coeffs + start + w2*128,
  947. R34,
  948. sce1->ics.swb_sizes[g],
  949. sce1->sf_idx[(w+w2)*16+g],
  950. sce1->band_type[(w+w2)*16+g],
  951. lambda / band1->threshold, INFINITY, NULL);
  952. dist2 += quantize_band_cost(s, M,
  953. M34,
  954. sce0->ics.swb_sizes[g],
  955. sce0->sf_idx[(w+w2)*16+g],
  956. sce0->band_type[(w+w2)*16+g],
  957. lambda / maxthr, INFINITY, NULL);
  958. dist2 += quantize_band_cost(s, S,
  959. S34,
  960. sce1->ics.swb_sizes[g],
  961. sce1->sf_idx[(w+w2)*16+g],
  962. sce1->band_type[(w+w2)*16+g],
  963. lambda / minthr, INFINITY, NULL);
  964. }
  965. cpe->ms_mask[w*16+g] = dist2 < dist1;
  966. }
  967. start += sce0->ics.swb_sizes[g];
  968. }
  969. }
  970. }
  971. AACCoefficientsEncoder ff_aac_coders[] = {
  972. {
  973. search_for_quantizers_faac,
  974. encode_window_bands_info,
  975. quantize_and_encode_band,
  976. // search_for_ms,
  977. },
  978. {
  979. search_for_quantizers_anmr,
  980. encode_window_bands_info,
  981. quantize_and_encode_band,
  982. // search_for_ms,
  983. },
  984. {
  985. search_for_quantizers_twoloop,
  986. encode_window_bands_info,
  987. quantize_and_encode_band,
  988. // search_for_ms,
  989. },
  990. {
  991. search_for_quantizers_fast,
  992. encode_window_bands_info,
  993. quantize_and_encode_band,
  994. // search_for_ms,
  995. },
  996. };