cosmetics: Remove unnecessary {} around if/for blocks;
[libav.git] / libavcodec / aaccoder.c
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 /**
23 * @file libavcodec/aaccoder.c
24 * AAC coefficients encoder
25 */
26
27 /***********************************
28 * TODOs:
29 * speedup quantizer selection
30 * add sane pulse detection
31 ***********************************/
32
33 #include "avcodec.h"
34 #include "put_bits.h"
35 #include "aac.h"
36 #include "aacenc.h"
37 #include "aactab.h"
38
39 /** bits needed to code codebook run value for long windows */
40 static const uint8_t run_value_bits_long[64] = {
41 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
42 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 10,
43 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
44 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 15
45 };
46
47 /** bits needed to code codebook run value for short windows */
48 static const uint8_t run_value_bits_short[16] = {
49 3, 3, 3, 3, 3, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 9
50 };
51
52 static const uint8_t *run_value_bits[2] = {
53 run_value_bits_long, run_value_bits_short
54 };
55
56
57 /**
58 * Quantize one coefficient.
59 * @return absolute value of the quantized coefficient
60 * @see 3GPP TS26.403 5.6.2 "Scalefactor determination"
61 */
62 static av_always_inline int quant(float coef, const float Q)
63 {
64 return pow(coef * Q, 0.75) + 0.4054;
65 }
66
67 static void quantize_bands(int (*out)[2], const float *in, const float *scaled,
68 int size, float Q34, int is_signed, int maxval)
69 {
70 int i;
71 double qc;
72 for (i = 0; i < size; i++) {
73 qc = scaled[i] * Q34;
74 out[i][0] = (int)FFMIN((int)qc, maxval);
75 out[i][1] = (int)FFMIN((int)(qc + 0.4054), maxval);
76 if (is_signed && in[i] < 0.0f) {
77 out[i][0] = -out[i][0];
78 out[i][1] = -out[i][1];
79 }
80 }
81 }
82
83 static void abs_pow34_v(float *out, const float *in, const int size)
84 {
85 #ifndef USE_REALLY_FULL_SEARCH
86 int i;
87 for (i = 0; i < size; i++)
88 out[i] = pow(fabsf(in[i]), 0.75);
89 #endif /* USE_REALLY_FULL_SEARCH */
90 }
91
92 static av_always_inline int quant2(float coef, const float Q)
93 {
94 return pow(coef * Q, 0.75);
95 }
96
97 static const uint8_t aac_cb_range [12] = {0, 3, 3, 3, 3, 9, 9, 8, 8, 13, 13, 17};
98 static const uint8_t aac_cb_maxval[12] = {0, 1, 1, 2, 2, 4, 4, 7, 7, 12, 12, 16};
99
100 /**
101 * Calculate rate distortion cost for quantizing with given codebook
102 *
103 * @return quantization distortion
104 */
105 static float quantize_band_cost(struct AACEncContext *s, const float *in,
106 const float *scaled, int size, int scale_idx,
107 int cb, const float lambda, const float uplim,
108 int *bits)
109 {
110 const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
111 const float Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
112 const float CLIPPED_ESCAPE = 165140.0f*IQ;
113 int i, j, k;
114 float cost = 0;
115 const int dim = cb < FIRST_PAIR_BT ? 4 : 2;
116 int resbits = 0;
117 #ifndef USE_REALLY_FULL_SEARCH
118 const float Q34 = pow(Q, 0.75);
119 const int range = aac_cb_range[cb];
120 const int maxval = aac_cb_maxval[cb];
121 int offs[4];
122 #endif /* USE_REALLY_FULL_SEARCH */
123
124 if (!cb) {
125 for (i = 0; i < size; i++)
126 cost += in[i]*in[i]*lambda;
127 return cost;
128 }
129 #ifndef USE_REALLY_FULL_SEARCH
130 offs[0] = 1;
131 for (i = 1; i < dim; i++)
132 offs[i] = offs[i-1]*range;
133 quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
134 #endif /* USE_REALLY_FULL_SEARCH */
135 for (i = 0; i < size; i += dim) {
136 float mincost;
137 int minidx = 0;
138 int minbits = 0;
139 const float *vec;
140 #ifndef USE_REALLY_FULL_SEARCH
141 int (*quants)[2] = &s->qcoefs[i];
142 mincost = 0.0f;
143 for (j = 0; j < dim; j++)
144 mincost += in[i+j]*in[i+j]*lambda;
145 minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
146 minbits = ff_aac_spectral_bits[cb-1][minidx];
147 mincost += minbits;
148 for (j = 0; j < (1<<dim); j++) {
149 float rd = 0.0f;
150 int curbits;
151 int curidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
152 int same = 0;
153 for (k = 0; k < dim; k++) {
154 if ((j & (1 << k)) && quants[k][0] == quants[k][1]) {
155 same = 1;
156 break;
157 }
158 }
159 if (same)
160 continue;
161 for (k = 0; k < dim; k++)
162 curidx += quants[k][!!(j & (1 << k))] * offs[dim - 1 - k];
163 curbits = ff_aac_spectral_bits[cb-1][curidx];
164 vec = &ff_aac_codebook_vectors[cb-1][curidx*dim];
165 #else
166 mincost = INFINITY;
167 vec = ff_aac_codebook_vectors[cb-1];
168 for (j = 0; j < ff_aac_spectral_sizes[cb-1]; j++, vec += dim) {
169 float rd = 0.0f;
170 int curbits = ff_aac_spectral_bits[cb-1][j];
171 #endif /* USE_REALLY_FULL_SEARCH */
172 if (IS_CODEBOOK_UNSIGNED(cb)) {
173 for (k = 0; k < dim; k++) {
174 float t = fabsf(in[i+k]);
175 float di;
176 //do not code with escape sequence small values
177 if (vec[k] == 64.0f && t < 39.0f*IQ) {
178 rd = INFINITY;
179 break;
180 }
181 if (vec[k] == 64.0f) { //FIXME: slow
182 if (t >= CLIPPED_ESCAPE) {
183 di = t - CLIPPED_ESCAPE;
184 curbits += 21;
185 } else {
186 int c = av_clip(quant(t, Q), 0, 8191);
187 di = t - c*cbrt(c)*IQ;
188 curbits += av_log2(c)*2 - 4 + 1;
189 }
190 } else {
191 di = t - vec[k]*IQ;
192 }
193 if (vec[k] != 0.0f)
194 curbits++;
195 rd += di*di*lambda;
196 }
197 } else {
198 for (k = 0; k < dim; k++) {
199 float di = in[i+k] - vec[k]*IQ;
200 rd += di*di*lambda;
201 }
202 }
203 rd += curbits;
204 if (rd < mincost) {
205 mincost = rd;
206 minidx = j;
207 minbits = curbits;
208 }
209 }
210 cost += mincost;
211 resbits += minbits;
212 if (cost >= uplim)
213 return uplim;
214 }
215
216 if (bits)
217 *bits = resbits;
218 return cost;
219 }
220
221 static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
222 const float *in, int size, int scale_idx,
223 int cb, const float lambda)
224 {
225 const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
226 const float Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
227 const float CLIPPED_ESCAPE = 165140.0f*IQ;
228 const int dim = (cb < FIRST_PAIR_BT) ? 4 : 2;
229 int i, j, k;
230 #ifndef USE_REALLY_FULL_SEARCH
231 const float Q34 = pow(Q, 0.75);
232 const int range = aac_cb_range[cb];
233 const int maxval = aac_cb_maxval[cb];
234 int offs[4];
235 float *scaled = s->scoefs;
236 #endif /* USE_REALLY_FULL_SEARCH */
237
238 //START_TIMER
239 if (!cb)
240 return;
241
242 #ifndef USE_REALLY_FULL_SEARCH
243 offs[0] = 1;
244 for (i = 1; i < dim; i++)
245 offs[i] = offs[i-1]*range;
246 abs_pow34_v(scaled, in, size);
247 quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
248 #endif /* USE_REALLY_FULL_SEARCH */
249 for (i = 0; i < size; i += dim) {
250 float mincost;
251 int minidx = 0;
252 int minbits = 0;
253 const float *vec;
254 #ifndef USE_REALLY_FULL_SEARCH
255 int (*quants)[2] = &s->qcoefs[i];
256 mincost = 0.0f;
257 for (j = 0; j < dim; j++)
258 mincost += in[i+j]*in[i+j]*lambda;
259 minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
260 minbits = ff_aac_spectral_bits[cb-1][minidx];
261 mincost += minbits;
262 for (j = 0; j < (1<<dim); j++) {
263 float rd = 0.0f;
264 int curbits;
265 int curidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
266 int same = 0;
267 for (k = 0; k < dim; k++) {
268 if ((j & (1 << k)) && quants[k][0] == quants[k][1]) {
269 same = 1;
270 break;
271 }
272 }
273 if (same)
274 continue;
275 for (k = 0; k < dim; k++)
276 curidx += quants[k][!!(j & (1 << k))] * offs[dim - 1 - k];
277 curbits = ff_aac_spectral_bits[cb-1][curidx];
278 vec = &ff_aac_codebook_vectors[cb-1][curidx*dim];
279 #else
280 vec = ff_aac_codebook_vectors[cb-1];
281 mincost = INFINITY;
282 for (j = 0; j < ff_aac_spectral_sizes[cb-1]; j++, vec += dim) {
283 float rd = 0.0f;
284 int curbits = ff_aac_spectral_bits[cb-1][j];
285 int curidx = j;
286 #endif /* USE_REALLY_FULL_SEARCH */
287 if (IS_CODEBOOK_UNSIGNED(cb)) {
288 for (k = 0; k < dim; k++) {
289 float t = fabsf(in[i+k]);
290 float di;
291 //do not code with escape sequence small values
292 if (vec[k] == 64.0f && t < 39.0f*IQ) {
293 rd = INFINITY;
294 break;
295 }
296 if (vec[k] == 64.0f) { //FIXME: slow
297 if (t >= CLIPPED_ESCAPE) {
298 di = t - CLIPPED_ESCAPE;
299 curbits += 21;
300 } else {
301 int c = av_clip(quant(t, Q), 0, 8191);
302 di = t - c*cbrt(c)*IQ;
303 curbits += av_log2(c)*2 - 4 + 1;
304 }
305 } else {
306 di = t - vec[k]*IQ;
307 }
308 if (vec[k] != 0.0f)
309 curbits++;
310 rd += di*di*lambda;
311 }
312 } else {
313 for (k = 0; k < dim; k++) {
314 float di = in[i+k] - vec[k]*IQ;
315 rd += di*di*lambda;
316 }
317 }
318 rd += curbits;
319 if (rd < mincost) {
320 mincost = rd;
321 minidx = curidx;
322 minbits = curbits;
323 }
324 }
325 put_bits(pb, ff_aac_spectral_bits[cb-1][minidx], ff_aac_spectral_codes[cb-1][minidx]);
326 if (IS_CODEBOOK_UNSIGNED(cb))
327 for (j = 0; j < dim; j++)
328 if (ff_aac_codebook_vectors[cb-1][minidx*dim+j] != 0.0f)
329 put_bits(pb, 1, in[i+j] < 0.0f);
330 if (cb == ESC_BT) {
331 for (j = 0; j < 2; j++) {
332 if (ff_aac_codebook_vectors[cb-1][minidx*2+j] == 64.0f) {
333 int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
334 int len = av_log2(coef);
335
336 put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
337 put_bits(pb, len, coef & ((1 << len) - 1));
338 }
339 }
340 }
341 }
342 //STOP_TIMER("quantize_and_encode")
343 }
344
345 /**
346 * structure used in optimal codebook search
347 */
348 typedef struct BandCodingPath {
349 int prev_idx; ///< pointer to the previous path point
350 int codebook; ///< codebook for coding band run
351 float cost; ///< path cost
352 int run;
353 } BandCodingPath;
354
355 /**
356 * Encode band info for single window group bands.
357 */
358 static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
359 int win, int group_len, const float lambda)
360 {
361 BandCodingPath path[120][12];
362 int w, swb, cb, start, start2, size;
363 int i, j;
364 const int max_sfb = sce->ics.max_sfb;
365 const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
366 const int run_esc = (1 << run_bits) - 1;
367 int idx, ppos, count;
368 int stackrun[120], stackcb[120], stack_len;
369 float next_minrd = INFINITY;
370 int next_mincb = 0;
371
372 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
373 start = win*128;
374 for (cb = 0; cb < 12; cb++) {
375 path[0][cb].cost = 0.0f;
376 path[0][cb].prev_idx = -1;
377 path[0][cb].run = 0;
378 }
379 for (swb = 0; swb < max_sfb; swb++) {
380 start2 = start;
381 size = sce->ics.swb_sizes[swb];
382 if (sce->zeroes[win*16 + swb]) {
383 for (cb = 0; cb < 12; cb++) {
384 path[swb+1][cb].prev_idx = cb;
385 path[swb+1][cb].cost = path[swb][cb].cost;
386 path[swb+1][cb].run = path[swb][cb].run + 1;
387 }
388 } else {
389 float minrd = next_minrd;
390 int mincb = next_mincb;
391 next_minrd = INFINITY;
392 next_mincb = 0;
393 for (cb = 0; cb < 12; cb++) {
394 float cost_stay_here, cost_get_here;
395 float rd = 0.0f;
396 for (w = 0; w < group_len; w++) {
397 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(win+w)*16+swb];
398 rd += quantize_band_cost(s, sce->coeffs + start + w*128,
399 s->scoefs + start + w*128, size,
400 sce->sf_idx[(win+w)*16+swb], cb,
401 lambda / band->threshold, INFINITY, NULL);
402 }
403 cost_stay_here = path[swb][cb].cost + rd;
404 cost_get_here = minrd + rd + run_bits + 4;
405 if ( run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
406 != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
407 cost_stay_here += run_bits;
408 if (cost_get_here < cost_stay_here) {
409 path[swb+1][cb].prev_idx = mincb;
410 path[swb+1][cb].cost = cost_get_here;
411 path[swb+1][cb].run = 1;
412 } else {
413 path[swb+1][cb].prev_idx = cb;
414 path[swb+1][cb].cost = cost_stay_here;
415 path[swb+1][cb].run = path[swb][cb].run + 1;
416 }
417 if (path[swb+1][cb].cost < next_minrd) {
418 next_minrd = path[swb+1][cb].cost;
419 next_mincb = cb;
420 }
421 }
422 }
423 start += sce->ics.swb_sizes[swb];
424 }
425
426 //convert resulting path from backward-linked list
427 stack_len = 0;
428 idx = 0;
429 for (cb = 1; cb < 12; cb++)
430 if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
431 idx = cb;
432 ppos = max_sfb;
433 while (ppos > 0) {
434 cb = idx;
435 stackrun[stack_len] = path[ppos][cb].run;
436 stackcb [stack_len] = cb;
437 idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
438 ppos -= path[ppos][cb].run;
439 stack_len++;
440 }
441 //perform actual band info encoding
442 start = 0;
443 for (i = stack_len - 1; i >= 0; i--) {
444 put_bits(&s->pb, 4, stackcb[i]);
445 count = stackrun[i];
446 memset(sce->zeroes + win*16 + start, !stackcb[i], count);
447 //XXX: memset when band_type is also uint8_t
448 for (j = 0; j < count; j++) {
449 sce->band_type[win*16 + start] = stackcb[i];
450 start++;
451 }
452 while (count >= run_esc) {
453 put_bits(&s->pb, run_bits, run_esc);
454 count -= run_esc;
455 }
456 put_bits(&s->pb, run_bits, count);
457 }
458 }
459
460 static void encode_window_bands_info_fixed(AACEncContext *s,
461 SingleChannelElement *sce,
462 int win, int group_len,
463 const float lambda)
464 {
465 encode_window_bands_info(s, sce, win, group_len, 1.0f);
466 }
467
468
469 typedef struct TrellisPath {
470 float cost;
471 int prev;
472 int min_val;
473 int max_val;
474 } TrellisPath;
475
476 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
477 SingleChannelElement *sce,
478 const float lambda)
479 {
480 int q, w, w2, g, start = 0;
481 int i;
482 int idx;
483 TrellisPath paths[256*121];
484 int bandaddr[121];
485 int minq;
486 float mincost;
487
488 for (i = 0; i < 256; i++) {
489 paths[i].cost = 0.0f;
490 paths[i].prev = -1;
491 paths[i].min_val = i;
492 paths[i].max_val = i;
493 }
494 for (i = 256; i < 256*121; i++) {
495 paths[i].cost = INFINITY;
496 paths[i].prev = -2;
497 paths[i].min_val = INT_MAX;
498 paths[i].max_val = 0;
499 }
500 idx = 256;
501 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
502 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
503 start = w*128;
504 for (g = 0; g < sce->ics.num_swb; g++) {
505 const float *coefs = sce->coeffs + start;
506 float qmin, qmax;
507 int nz = 0;
508
509 bandaddr[idx >> 8] = w * 16 + g;
510 qmin = INT_MAX;
511 qmax = 0.0f;
512 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
513 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
514 if (band->energy <= band->threshold || band->threshold == 0.0f) {
515 sce->zeroes[(w+w2)*16+g] = 1;
516 continue;
517 }
518 sce->zeroes[(w+w2)*16+g] = 0;
519 nz = 1;
520 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
521 float t = fabsf(coefs[w2*128+i]);
522 if (t > 0.0f)
523 qmin = fminf(qmin, t);
524 qmax = fmaxf(qmax, t);
525 }
526 }
527 if (nz) {
528 int minscale, maxscale;
529 float minrd = INFINITY;
530 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
531 minscale = av_clip_uint8(log2(qmin)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
532 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
533 maxscale = av_clip_uint8(log2(qmax)*4 + 6 + SCALE_ONE_POS - SCALE_DIV_512);
534 for (q = minscale; q < maxscale; q++) {
535 float dists[12], dist;
536 memset(dists, 0, sizeof(dists));
537 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
538 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
539 int cb;
540 for (cb = 0; cb <= ESC_BT; cb++)
541 dists[cb] += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
542 q, cb, lambda / band->threshold, INFINITY, NULL);
543 }
544 dist = dists[0];
545 for (i = 1; i <= ESC_BT; i++)
546 dist = fminf(dist, dists[i]);
547 minrd = fminf(minrd, dist);
548
549 for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, 256); i++) {
550 float cost;
551 int minv, maxv;
552 if (isinf(paths[idx - 256 + i].cost))
553 continue;
554 cost = paths[idx - 256 + i].cost + dist
555 + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
556 minv = FFMIN(paths[idx - 256 + i].min_val, q);
557 maxv = FFMAX(paths[idx - 256 + i].max_val, q);
558 if (cost < paths[idx + q].cost && maxv-minv < SCALE_MAX_DIFF) {
559 paths[idx + q].cost = cost;
560 paths[idx + q].prev = idx - 256 + i;
561 paths[idx + q].min_val = minv;
562 paths[idx + q].max_val = maxv;
563 }
564 }
565 }
566 } else {
567 for (q = 0; q < 256; q++) {
568 if (!isinf(paths[idx - 256 + q].cost)) {
569 paths[idx + q].cost = paths[idx - 256 + q].cost + 1;
570 paths[idx + q].prev = idx - 256 + q;
571 paths[idx + q].min_val = FFMIN(paths[idx - 256 + q].min_val, q);
572 paths[idx + q].max_val = FFMAX(paths[idx - 256 + q].max_val, q);
573 continue;
574 }
575 for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, 256); i++) {
576 float cost;
577 int minv, maxv;
578 if (isinf(paths[idx - 256 + i].cost))
579 continue;
580 cost = paths[idx - 256 + i].cost + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
581 minv = FFMIN(paths[idx - 256 + i].min_val, q);
582 maxv = FFMAX(paths[idx - 256 + i].max_val, q);
583 if (cost < paths[idx + q].cost && maxv-minv < SCALE_MAX_DIFF) {
584 paths[idx + q].cost = cost;
585 paths[idx + q].prev = idx - 256 + i;
586 paths[idx + q].min_val = minv;
587 paths[idx + q].max_val = maxv;
588 }
589 }
590 }
591 }
592 sce->zeroes[w*16+g] = !nz;
593 start += sce->ics.swb_sizes[g];
594 idx += 256;
595 }
596 }
597 idx -= 256;
598 mincost = paths[idx].cost;
599 minq = idx;
600 for (i = 1; i < 256; i++) {
601 if (paths[idx + i].cost < mincost) {
602 mincost = paths[idx + i].cost;
603 minq = idx + i;
604 }
605 }
606 while (minq >= 256) {
607 sce->sf_idx[bandaddr[minq>>8]] = minq & 0xFF;
608 minq = paths[minq].prev;
609 }
610 //set the same quantizers inside window groups
611 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
612 for (g = 0; g < sce->ics.num_swb; g++)
613 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
614 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
615 }
616
617 /**
618 * two-loop quantizers search taken from ISO 13818-7 Appendix C
619 */
620 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
621 AACEncContext *s,
622 SingleChannelElement *sce,
623 const float lambda)
624 {
625 int start = 0, i, w, w2, g;
626 int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels;
627 float dists[128], uplims[128];
628 int fflag, minscaler;
629 int its = 0;
630 int allz = 0;
631 float minthr = INFINITY;
632
633 //XXX: some heuristic to determine initial quantizers will reduce search time
634 memset(dists, 0, sizeof(dists));
635 //determine zero bands and upper limits
636 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
637 for (g = 0; g < sce->ics.num_swb; g++) {
638 int nz = 0;
639 float uplim = 0.0f;
640 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
641 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
642 uplim += band->threshold;
643 if (band->energy <= band->threshold || band->threshold == 0.0f) {
644 sce->zeroes[(w+w2)*16+g] = 1;
645 continue;
646 }
647 nz = 1;
648 }
649 uplims[w*16+g] = uplim *512;
650 sce->zeroes[w*16+g] = !nz;
651 if (nz)
652 minthr = fminf(minthr, uplim);
653 allz = FFMAX(allz, nz);
654 }
655 }
656 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
657 for (g = 0; g < sce->ics.num_swb; g++) {
658 if (sce->zeroes[w*16+g]) {
659 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
660 continue;
661 }
662 sce->sf_idx[w*16+g] = SCALE_ONE_POS + fminf(log2(uplims[w*16+g]/minthr)*4,59);
663 }
664 }
665
666 if (!allz)
667 return;
668 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
669 //perform two-loop search
670 //outer loop - improve quality
671 do {
672 int tbits, qstep;
673 minscaler = sce->sf_idx[0];
674 //inner loop - quantize spectrum to fit into given number of bits
675 qstep = its ? 1 : 32;
676 do {
677 int prev = -1;
678 tbits = 0;
679 fflag = 0;
680 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
681 start = w*128;
682 for (g = 0; g < sce->ics.num_swb; g++) {
683 const float *coefs = sce->coeffs + start;
684 const float *scaled = s->scoefs + start;
685 int bits = 0;
686 int cb;
687 float mindist = INFINITY;
688 int minbits = 0;
689
690 if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218)
691 continue;
692 minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
693 for (cb = 0; cb <= ESC_BT; cb++) {
694 float dist = 0.0f;
695 int bb = 0;
696 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
697 int b;
698 dist += quantize_band_cost(s, coefs + w2*128,
699 scaled + w2*128,
700 sce->ics.swb_sizes[g],
701 sce->sf_idx[w*16+g],
702 ESC_BT,
703 1.0,
704 INFINITY,
705 &b);
706 bb += b;
707 }
708 if (dist < mindist) {
709 mindist = dist;
710 minbits = bb;
711 }
712 }
713 dists[w*16+g] = mindist - minbits;
714 bits = minbits;
715 if (prev != -1) {
716 bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
717 }
718 tbits += bits;
719 start += sce->ics.swb_sizes[g];
720 prev = sce->sf_idx[w*16+g];
721 }
722 }
723 if (tbits > destbits) {
724 for (i = 0; i < 128; i++)
725 if (sce->sf_idx[i] < 218 - qstep)
726 sce->sf_idx[i] += qstep;
727 } else {
728 for (i = 0; i < 128; i++)
729 if (sce->sf_idx[i] > 60 - qstep)
730 sce->sf_idx[i] -= qstep;
731 }
732 qstep >>= 1;
733 if (!qstep && tbits > destbits*1.02)
734 qstep = 1;
735 if (sce->sf_idx[0] >= 217)
736 break;
737 } while (qstep);
738
739 fflag = 0;
740 minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
741 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
742 start = w*128;
743 for (g = 0; g < sce->ics.num_swb; g++) {
744 int prevsc = sce->sf_idx[w*16+g];
745 if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60)
746 sce->sf_idx[w*16+g]--;
747 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
748 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
749 if (sce->sf_idx[w*16+g] != prevsc)
750 fflag = 1;
751 }
752 }
753 its++;
754 } while (fflag && its < 10);
755 }
756
757 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
758 SingleChannelElement *sce,
759 const float lambda)
760 {
761 int start = 0, i, w, w2, g;
762 float uplim[128], maxq[128];
763 int minq, maxsf;
764 float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
765 int last = 0, lastband = 0, curband = 0;
766 float avg_energy = 0.0;
767 if (sce->ics.num_windows == 1) {
768 start = 0;
769 for (i = 0; i < 1024; i++) {
770 if (i - start >= sce->ics.swb_sizes[curband]) {
771 start += sce->ics.swb_sizes[curband];
772 curband++;
773 }
774 if (sce->coeffs[i]) {
775 avg_energy += sce->coeffs[i] * sce->coeffs[i];
776 last = i;
777 lastband = curband;
778 }
779 }
780 } else {
781 for (w = 0; w < 8; w++) {
782 const float *coeffs = sce->coeffs + w*128;
783 start = 0;
784 for (i = 0; i < 128; i++) {
785 if (i - start >= sce->ics.swb_sizes[curband]) {
786 start += sce->ics.swb_sizes[curband];
787 curband++;
788 }
789 if (coeffs[i]) {
790 avg_energy += coeffs[i] * coeffs[i];
791 last = FFMAX(last, i);
792 lastband = FFMAX(lastband, curband);
793 }
794 }
795 }
796 }
797 last++;
798 avg_energy /= last;
799 if (avg_energy == 0.0f) {
800 for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
801 sce->sf_idx[i] = SCALE_ONE_POS;
802 return;
803 }
804 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
805 start = w*128;
806 for (g = 0; g < sce->ics.num_swb; g++) {
807 float *coefs = sce->coeffs + start;
808 const int size = sce->ics.swb_sizes[g];
809 int start2 = start, end2 = start + size, peakpos = start;
810 float maxval = -1, thr = 0.0f, t;
811 maxq[w*16+g] = 0.0f;
812 if (g > lastband) {
813 maxq[w*16+g] = 0.0f;
814 start += size;
815 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
816 memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
817 continue;
818 }
819 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
820 for (i = 0; i < size; i++) {
821 float t = coefs[w2*128+i]*coefs[w2*128+i];
822 maxq[w*16+g] = fmaxf(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
823 thr += t;
824 if (sce->ics.num_windows == 1 && maxval < t) {
825 maxval = t;
826 peakpos = start+i;
827 }
828 }
829 }
830 if (sce->ics.num_windows == 1) {
831 start2 = FFMAX(peakpos - 2, start2);
832 end2 = FFMIN(peakpos + 3, end2);
833 } else {
834 start2 -= start;
835 end2 -= start;
836 }
837 start += size;
838 thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
839 t = 1.0 - (1.0 * start2 / last);
840 uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
841 }
842 }
843 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
844 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
845 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
846 start = w*128;
847 for (g = 0; g < sce->ics.num_swb; g++) {
848 const float *coefs = sce->coeffs + start;
849 const float *scaled = s->scoefs + start;
850 const int size = sce->ics.swb_sizes[g];
851 int scf, prev_scf, step;
852 int min_scf = 0, max_scf = 255;
853 float curdiff;
854 if (maxq[w*16+g] < 21.544) {
855 sce->zeroes[w*16+g] = 1;
856 start += size;
857 continue;
858 }
859 sce->zeroes[w*16+g] = 0;
860 scf = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2(1/maxq[w*16+g])*16/3, 60, 218);
861 step = 16;
862 for (;;) {
863 float dist = 0.0f;
864 int quant_max;
865
866 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
867 int b;
868 dist += quantize_band_cost(s, coefs + w2*128,
869 scaled + w2*128,
870 sce->ics.swb_sizes[g],
871 scf,
872 ESC_BT,
873 1.0,
874 INFINITY,
875 &b);
876 dist -= b;
877 }
878 dist *= 1.0f/512.0f;
879 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[200 - scf + SCALE_ONE_POS - SCALE_DIV_512]);
880 if (quant_max >= 8191) { // too much, return to the previous quantizer
881 sce->sf_idx[w*16+g] = prev_scf;
882 break;
883 }
884 prev_scf = scf;
885 curdiff = fabsf(dist - uplim[w*16+g]);
886 if (curdiff == 0.0f)
887 step = 0;
888 else
889 step = fabsf(log2(curdiff));
890 if (dist > uplim[w*16+g])
891 step = -step;
892 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
893 sce->sf_idx[w*16+g] = scf;
894 break;
895 }
896 scf += step;
897 if (step > 0)
898 min_scf = scf;
899 else
900 max_scf = scf;
901 }
902 start += size;
903 }
904 }
905 minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
906 for (i = 1; i < 128; i++) {
907 if (!sce->sf_idx[i])
908 sce->sf_idx[i] = sce->sf_idx[i-1];
909 else
910 minq = FFMIN(minq, sce->sf_idx[i]);
911 }
912 if (minq == INT_MAX)
913 minq = 0;
914 minq = FFMIN(minq, SCALE_MAX_POS);
915 maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
916 for (i = 126; i >= 0; i--) {
917 if (!sce->sf_idx[i])
918 sce->sf_idx[i] = sce->sf_idx[i+1];
919 sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
920 }
921 }
922
923 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
924 SingleChannelElement *sce,
925 const float lambda)
926 {
927 int start = 0, i, w, w2, g;
928 int minq = 255;
929
930 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
931 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
932 start = w*128;
933 for (g = 0; g < sce->ics.num_swb; g++) {
934 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
935 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
936 if (band->energy <= band->threshold) {
937 sce->sf_idx[(w+w2)*16+g] = 218;
938 sce->zeroes[(w+w2)*16+g] = 1;
939 } else {
940 sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2(band->threshold), 80, 218);
941 sce->zeroes[(w+w2)*16+g] = 0;
942 }
943 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
944 }
945 }
946 }
947 for (i = 0; i < 128; i++) {
948 sce->sf_idx[i] = 140;
949 //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
950 }
951 //set the same quantizers inside window groups
952 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
953 for (g = 0; g < sce->ics.num_swb; g++)
954 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
955 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
956 }
957
958 static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
959 const float lambda)
960 {
961 int start = 0, i, w, w2, g;
962 float M[128], S[128];
963 float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
964 SingleChannelElement *sce0 = &cpe->ch[0];
965 SingleChannelElement *sce1 = &cpe->ch[1];
966 if (!cpe->common_window)
967 return;
968 for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
969 for (g = 0; g < sce0->ics.num_swb; g++) {
970 if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
971 float dist1 = 0.0f, dist2 = 0.0f;
972 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
973 FFPsyBand *band0 = &s->psy.psy_bands[(s->cur_channel+0)*PSY_MAX_BANDS+(w+w2)*16+g];
974 FFPsyBand *band1 = &s->psy.psy_bands[(s->cur_channel+1)*PSY_MAX_BANDS+(w+w2)*16+g];
975 float minthr = fminf(band0->threshold, band1->threshold);
976 float maxthr = fmaxf(band0->threshold, band1->threshold);
977 for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
978 M[i] = (sce0->coeffs[start+w2*128+i]
979 + sce1->coeffs[start+w2*128+i]) * 0.5;
980 S[i] = sce0->coeffs[start+w2*128+i]
981 - sce1->coeffs[start+w2*128+i];
982 }
983 abs_pow34_v(L34, sce0->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
984 abs_pow34_v(R34, sce1->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
985 abs_pow34_v(M34, M, sce0->ics.swb_sizes[g]);
986 abs_pow34_v(S34, S, sce0->ics.swb_sizes[g]);
987 dist1 += quantize_band_cost(s, sce0->coeffs + start + w2*128,
988 L34,
989 sce0->ics.swb_sizes[g],
990 sce0->sf_idx[(w+w2)*16+g],
991 sce0->band_type[(w+w2)*16+g],
992 lambda / band0->threshold, INFINITY, NULL);
993 dist1 += quantize_band_cost(s, sce1->coeffs + start + w2*128,
994 R34,
995 sce1->ics.swb_sizes[g],
996 sce1->sf_idx[(w+w2)*16+g],
997 sce1->band_type[(w+w2)*16+g],
998 lambda / band1->threshold, INFINITY, NULL);
999 dist2 += quantize_band_cost(s, M,
1000 M34,
1001 sce0->ics.swb_sizes[g],
1002 sce0->sf_idx[(w+w2)*16+g],
1003 sce0->band_type[(w+w2)*16+g],
1004 lambda / maxthr, INFINITY, NULL);
1005 dist2 += quantize_band_cost(s, S,
1006 S34,
1007 sce1->ics.swb_sizes[g],
1008 sce1->sf_idx[(w+w2)*16+g],
1009 sce1->band_type[(w+w2)*16+g],
1010 lambda / minthr, INFINITY, NULL);
1011 }
1012 cpe->ms_mask[w*16+g] = dist2 < dist1;
1013 }
1014 start += sce0->ics.swb_sizes[g];
1015 }
1016 }
1017 }
1018
1019 AACCoefficientsEncoder ff_aac_coders[] = {
1020 {
1021 search_for_quantizers_faac,
1022 encode_window_bands_info_fixed,
1023 quantize_and_encode_band,
1024 // search_for_ms,
1025 },
1026 {
1027 search_for_quantizers_anmr,
1028 encode_window_bands_info,
1029 quantize_and_encode_band,
1030 // search_for_ms,
1031 },
1032 {
1033 search_for_quantizers_twoloop,
1034 encode_window_bands_info,
1035 quantize_and_encode_band,
1036 // search_for_ms,
1037 },
1038 {
1039 search_for_quantizers_fast,
1040 encode_window_bands_info,
1041 quantize_and_encode_band,
1042 // search_for_ms,
1043 },
1044 };