Imported Debian version 2.5.3~trusty1
[deb_ffmpeg.git] / ffmpeg / 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
24 * AAC coefficients encoder
25 */
26
27 /***********************************
28 * TODOs:
29 * speedup quantizer selection
30 * add sane pulse detection
31 ***********************************/
32
33 #include "libavutil/libm.h" // brought forward to work around cygwin header breakage
34
35 #include <float.h>
36 #include "libavutil/mathematics.h"
37 #include "avcodec.h"
38 #include "put_bits.h"
39 #include "aac.h"
40 #include "aacenc.h"
41 #include "aactab.h"
42
43 /** bits needed to code codebook run value for long windows */
44 static const uint8_t run_value_bits_long[64] = {
45 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
46 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 10,
47 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
48 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 15
49 };
50
51 /** bits needed to code codebook run value for short windows */
52 static const uint8_t run_value_bits_short[16] = {
53 3, 3, 3, 3, 3, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 9
54 };
55
56 static const uint8_t * const run_value_bits[2] = {
57 run_value_bits_long, run_value_bits_short
58 };
59
60
61 /**
62 * Quantize one coefficient.
63 * @return absolute value of the quantized coefficient
64 * @see 3GPP TS26.403 5.6.2 "Scalefactor determination"
65 */
66 static av_always_inline int quant(float coef, const float Q)
67 {
68 float a = coef * Q;
69 return sqrtf(a * sqrtf(a)) + 0.4054;
70 }
71
72 static void quantize_bands(int *out, const float *in, const float *scaled,
73 int size, float Q34, int is_signed, int maxval)
74 {
75 int i;
76 double qc;
77 for (i = 0; i < size; i++) {
78 qc = scaled[i] * Q34;
79 out[i] = (int)FFMIN(qc + 0.4054, (double)maxval);
80 if (is_signed && in[i] < 0.0f) {
81 out[i] = -out[i];
82 }
83 }
84 }
85
86 static void abs_pow34_v(float *out, const float *in, const int size)
87 {
88 #ifndef USE_REALLY_FULL_SEARCH
89 int i;
90 for (i = 0; i < size; i++) {
91 float a = fabsf(in[i]);
92 out[i] = sqrtf(a * sqrtf(a));
93 }
94 #endif /* USE_REALLY_FULL_SEARCH */
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 av_always_inline float quantize_and_encode_band_cost_template(
106 struct AACEncContext *s,
107 PutBitContext *pb, const float *in,
108 const float *scaled, int size, int scale_idx,
109 int cb, const float lambda, const float uplim,
110 int *bits, int BT_ZERO, int BT_UNSIGNED,
111 int BT_PAIR, int BT_ESC)
112 {
113 const int q_idx = POW_SF2_ZERO - scale_idx + SCALE_ONE_POS - SCALE_DIV_512;
114 const float Q = ff_aac_pow2sf_tab [q_idx];
115 const float Q34 = ff_aac_pow34sf_tab[q_idx];
116 const float IQ = ff_aac_pow2sf_tab [POW_SF2_ZERO + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
117 const float CLIPPED_ESCAPE = 165140.0f*IQ;
118 int i, j;
119 float cost = 0;
120 const int dim = BT_PAIR ? 2 : 4;
121 int resbits = 0;
122 const int range = aac_cb_range[cb];
123 const int maxval = aac_cb_maxval[cb];
124 int off;
125
126 if (BT_ZERO) {
127 for (i = 0; i < size; i++)
128 cost += in[i]*in[i];
129 if (bits)
130 *bits = 0;
131 return cost * lambda;
132 }
133 if (!scaled) {
134 abs_pow34_v(s->scoefs, in, size);
135 scaled = s->scoefs;
136 }
137 quantize_bands(s->qcoefs, in, scaled, size, Q34, !BT_UNSIGNED, maxval);
138 if (BT_UNSIGNED) {
139 off = 0;
140 } else {
141 off = maxval;
142 }
143 for (i = 0; i < size; i += dim) {
144 const float *vec;
145 int *quants = s->qcoefs + i;
146 int curidx = 0;
147 int curbits;
148 float rd = 0.0f;
149 for (j = 0; j < dim; j++) {
150 curidx *= range;
151 curidx += quants[j] + off;
152 }
153 curbits = ff_aac_spectral_bits[cb-1][curidx];
154 vec = &ff_aac_codebook_vectors[cb-1][curidx*dim];
155 if (BT_UNSIGNED) {
156 for (j = 0; j < dim; j++) {
157 float t = fabsf(in[i+j]);
158 float di;
159 if (BT_ESC && vec[j] == 64.0f) { //FIXME: slow
160 if (t >= CLIPPED_ESCAPE) {
161 di = t - CLIPPED_ESCAPE;
162 curbits += 21;
163 } else {
164 int c = av_clip(quant(t, Q), 0, 8191);
165 di = t - c*cbrtf(c)*IQ;
166 curbits += av_log2(c)*2 - 4 + 1;
167 }
168 } else {
169 di = t - vec[j]*IQ;
170 }
171 if (vec[j] != 0.0f)
172 curbits++;
173 rd += di*di;
174 }
175 } else {
176 for (j = 0; j < dim; j++) {
177 float di = in[i+j] - vec[j]*IQ;
178 rd += di*di;
179 }
180 }
181 cost += rd * lambda + curbits;
182 resbits += curbits;
183 if (cost >= uplim)
184 return uplim;
185 if (pb) {
186 put_bits(pb, ff_aac_spectral_bits[cb-1][curidx], ff_aac_spectral_codes[cb-1][curidx]);
187 if (BT_UNSIGNED)
188 for (j = 0; j < dim; j++)
189 if (ff_aac_codebook_vectors[cb-1][curidx*dim+j] != 0.0f)
190 put_bits(pb, 1, in[i+j] < 0.0f);
191 if (BT_ESC) {
192 for (j = 0; j < 2; j++) {
193 if (ff_aac_codebook_vectors[cb-1][curidx*2+j] == 64.0f) {
194 int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
195 int len = av_log2(coef);
196
197 put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
198 put_bits(pb, len, coef & ((1 << len) - 1));
199 }
200 }
201 }
202 }
203 }
204
205 if (bits)
206 *bits = resbits;
207 return cost;
208 }
209
210 #define QUANTIZE_AND_ENCODE_BAND_COST_FUNC(NAME, BT_ZERO, BT_UNSIGNED, BT_PAIR, BT_ESC) \
211 static float quantize_and_encode_band_cost_ ## NAME( \
212 struct AACEncContext *s, \
213 PutBitContext *pb, const float *in, \
214 const float *scaled, int size, int scale_idx, \
215 int cb, const float lambda, const float uplim, \
216 int *bits) { \
217 return quantize_and_encode_band_cost_template( \
218 s, pb, in, scaled, size, scale_idx, \
219 BT_ESC ? ESC_BT : cb, lambda, uplim, bits, \
220 BT_ZERO, BT_UNSIGNED, BT_PAIR, BT_ESC); \
221 }
222
223 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(ZERO, 1, 0, 0, 0)
224 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(SQUAD, 0, 0, 0, 0)
225 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(UQUAD, 0, 1, 0, 0)
226 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(SPAIR, 0, 0, 1, 0)
227 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(UPAIR, 0, 1, 1, 0)
228 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(ESC, 0, 1, 1, 1)
229
230 static float (*const quantize_and_encode_band_cost_arr[])(
231 struct AACEncContext *s,
232 PutBitContext *pb, const float *in,
233 const float *scaled, int size, int scale_idx,
234 int cb, const float lambda, const float uplim,
235 int *bits) = {
236 quantize_and_encode_band_cost_ZERO,
237 quantize_and_encode_band_cost_SQUAD,
238 quantize_and_encode_band_cost_SQUAD,
239 quantize_and_encode_band_cost_UQUAD,
240 quantize_and_encode_band_cost_UQUAD,
241 quantize_and_encode_band_cost_SPAIR,
242 quantize_and_encode_band_cost_SPAIR,
243 quantize_and_encode_band_cost_UPAIR,
244 quantize_and_encode_band_cost_UPAIR,
245 quantize_and_encode_band_cost_UPAIR,
246 quantize_and_encode_band_cost_UPAIR,
247 quantize_and_encode_band_cost_ESC,
248 };
249
250 #define quantize_and_encode_band_cost( \
251 s, pb, in, scaled, size, scale_idx, cb, \
252 lambda, uplim, bits) \
253 quantize_and_encode_band_cost_arr[cb]( \
254 s, pb, in, scaled, size, scale_idx, cb, \
255 lambda, uplim, bits)
256
257 static float quantize_band_cost(struct AACEncContext *s, const float *in,
258 const float *scaled, int size, int scale_idx,
259 int cb, const float lambda, const float uplim,
260 int *bits)
261 {
262 return quantize_and_encode_band_cost(s, NULL, in, scaled, size, scale_idx,
263 cb, lambda, uplim, bits);
264 }
265
266 static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
267 const float *in, int size, int scale_idx,
268 int cb, const float lambda)
269 {
270 quantize_and_encode_band_cost(s, pb, in, NULL, size, scale_idx, cb, lambda,
271 INFINITY, NULL);
272 }
273
274 static float find_max_val(int group_len, int swb_size, const float *scaled) {
275 float maxval = 0.0f;
276 int w2, i;
277 for (w2 = 0; w2 < group_len; w2++) {
278 for (i = 0; i < swb_size; i++) {
279 maxval = FFMAX(maxval, scaled[w2*128+i]);
280 }
281 }
282 return maxval;
283 }
284
285 static int find_min_book(float maxval, int sf) {
286 float Q = ff_aac_pow2sf_tab[POW_SF2_ZERO - sf + SCALE_ONE_POS - SCALE_DIV_512];
287 float Q34 = sqrtf(Q * sqrtf(Q));
288 int qmaxval, cb;
289 qmaxval = maxval * Q34 + 0.4054f;
290 if (qmaxval == 0) cb = 0;
291 else if (qmaxval == 1) cb = 1;
292 else if (qmaxval == 2) cb = 3;
293 else if (qmaxval <= 4) cb = 5;
294 else if (qmaxval <= 7) cb = 7;
295 else if (qmaxval <= 12) cb = 9;
296 else cb = 11;
297 return cb;
298 }
299
300 /**
301 * structure used in optimal codebook search
302 */
303 typedef struct BandCodingPath {
304 int prev_idx; ///< pointer to the previous path point
305 float cost; ///< path cost
306 int run;
307 } BandCodingPath;
308
309 /**
310 * Encode band info for single window group bands.
311 */
312 static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
313 int win, int group_len, const float lambda)
314 {
315 BandCodingPath path[120][12];
316 int w, swb, cb, start, size;
317 int i, j;
318 const int max_sfb = sce->ics.max_sfb;
319 const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
320 const int run_esc = (1 << run_bits) - 1;
321 int idx, ppos, count;
322 int stackrun[120], stackcb[120], stack_len;
323 float next_minrd = INFINITY;
324 int next_mincb = 0;
325
326 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
327 start = win*128;
328 for (cb = 0; cb < 12; cb++) {
329 path[0][cb].cost = 0.0f;
330 path[0][cb].prev_idx = -1;
331 path[0][cb].run = 0;
332 }
333 for (swb = 0; swb < max_sfb; swb++) {
334 size = sce->ics.swb_sizes[swb];
335 if (sce->zeroes[win*16 + swb]) {
336 for (cb = 0; cb < 12; cb++) {
337 path[swb+1][cb].prev_idx = cb;
338 path[swb+1][cb].cost = path[swb][cb].cost;
339 path[swb+1][cb].run = path[swb][cb].run + 1;
340 }
341 } else {
342 float minrd = next_minrd;
343 int mincb = next_mincb;
344 next_minrd = INFINITY;
345 next_mincb = 0;
346 for (cb = 0; cb < 12; cb++) {
347 float cost_stay_here, cost_get_here;
348 float rd = 0.0f;
349 for (w = 0; w < group_len; w++) {
350 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(win+w)*16+swb];
351 rd += quantize_band_cost(s, sce->coeffs + start + w*128,
352 s->scoefs + start + w*128, size,
353 sce->sf_idx[(win+w)*16+swb], cb,
354 lambda / band->threshold, INFINITY, NULL);
355 }
356 cost_stay_here = path[swb][cb].cost + rd;
357 cost_get_here = minrd + rd + run_bits + 4;
358 if ( run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
359 != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
360 cost_stay_here += run_bits;
361 if (cost_get_here < cost_stay_here) {
362 path[swb+1][cb].prev_idx = mincb;
363 path[swb+1][cb].cost = cost_get_here;
364 path[swb+1][cb].run = 1;
365 } else {
366 path[swb+1][cb].prev_idx = cb;
367 path[swb+1][cb].cost = cost_stay_here;
368 path[swb+1][cb].run = path[swb][cb].run + 1;
369 }
370 if (path[swb+1][cb].cost < next_minrd) {
371 next_minrd = path[swb+1][cb].cost;
372 next_mincb = cb;
373 }
374 }
375 }
376 start += sce->ics.swb_sizes[swb];
377 }
378
379 //convert resulting path from backward-linked list
380 stack_len = 0;
381 idx = 0;
382 for (cb = 1; cb < 12; cb++)
383 if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
384 idx = cb;
385 ppos = max_sfb;
386 while (ppos > 0) {
387 cb = idx;
388 stackrun[stack_len] = path[ppos][cb].run;
389 stackcb [stack_len] = cb;
390 idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
391 ppos -= path[ppos][cb].run;
392 stack_len++;
393 }
394 //perform actual band info encoding
395 start = 0;
396 for (i = stack_len - 1; i >= 0; i--) {
397 put_bits(&s->pb, 4, stackcb[i]);
398 count = stackrun[i];
399 memset(sce->zeroes + win*16 + start, !stackcb[i], count);
400 //XXX: memset when band_type is also uint8_t
401 for (j = 0; j < count; j++) {
402 sce->band_type[win*16 + start] = stackcb[i];
403 start++;
404 }
405 while (count >= run_esc) {
406 put_bits(&s->pb, run_bits, run_esc);
407 count -= run_esc;
408 }
409 put_bits(&s->pb, run_bits, count);
410 }
411 }
412
413 static void codebook_trellis_rate(AACEncContext *s, SingleChannelElement *sce,
414 int win, int group_len, const float lambda)
415 {
416 BandCodingPath path[120][12];
417 int w, swb, cb, start, size;
418 int i, j;
419 const int max_sfb = sce->ics.max_sfb;
420 const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
421 const int run_esc = (1 << run_bits) - 1;
422 int idx, ppos, count;
423 int stackrun[120], stackcb[120], stack_len;
424 float next_minbits = INFINITY;
425 int next_mincb = 0;
426
427 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
428 start = win*128;
429 for (cb = 0; cb < 12; cb++) {
430 path[0][cb].cost = run_bits+4;
431 path[0][cb].prev_idx = -1;
432 path[0][cb].run = 0;
433 }
434 for (swb = 0; swb < max_sfb; swb++) {
435 size = sce->ics.swb_sizes[swb];
436 if (sce->zeroes[win*16 + swb]) {
437 float cost_stay_here = path[swb][0].cost;
438 float cost_get_here = next_minbits + run_bits + 4;
439 if ( run_value_bits[sce->ics.num_windows == 8][path[swb][0].run]
440 != run_value_bits[sce->ics.num_windows == 8][path[swb][0].run+1])
441 cost_stay_here += run_bits;
442 if (cost_get_here < cost_stay_here) {
443 path[swb+1][0].prev_idx = next_mincb;
444 path[swb+1][0].cost = cost_get_here;
445 path[swb+1][0].run = 1;
446 } else {
447 path[swb+1][0].prev_idx = 0;
448 path[swb+1][0].cost = cost_stay_here;
449 path[swb+1][0].run = path[swb][0].run + 1;
450 }
451 next_minbits = path[swb+1][0].cost;
452 next_mincb = 0;
453 for (cb = 1; cb < 12; cb++) {
454 path[swb+1][cb].cost = 61450;
455 path[swb+1][cb].prev_idx = -1;
456 path[swb+1][cb].run = 0;
457 }
458 } else {
459 float minbits = next_minbits;
460 int mincb = next_mincb;
461 int startcb = sce->band_type[win*16+swb];
462 next_minbits = INFINITY;
463 next_mincb = 0;
464 for (cb = 0; cb < startcb; cb++) {
465 path[swb+1][cb].cost = 61450;
466 path[swb+1][cb].prev_idx = -1;
467 path[swb+1][cb].run = 0;
468 }
469 for (cb = startcb; cb < 12; cb++) {
470 float cost_stay_here, cost_get_here;
471 float bits = 0.0f;
472 for (w = 0; w < group_len; w++) {
473 bits += quantize_band_cost(s, sce->coeffs + start + w*128,
474 s->scoefs + start + w*128, size,
475 sce->sf_idx[(win+w)*16+swb], cb,
476 0, INFINITY, NULL);
477 }
478 cost_stay_here = path[swb][cb].cost + bits;
479 cost_get_here = minbits + bits + run_bits + 4;
480 if ( run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
481 != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
482 cost_stay_here += run_bits;
483 if (cost_get_here < cost_stay_here) {
484 path[swb+1][cb].prev_idx = mincb;
485 path[swb+1][cb].cost = cost_get_here;
486 path[swb+1][cb].run = 1;
487 } else {
488 path[swb+1][cb].prev_idx = cb;
489 path[swb+1][cb].cost = cost_stay_here;
490 path[swb+1][cb].run = path[swb][cb].run + 1;
491 }
492 if (path[swb+1][cb].cost < next_minbits) {
493 next_minbits = path[swb+1][cb].cost;
494 next_mincb = cb;
495 }
496 }
497 }
498 start += sce->ics.swb_sizes[swb];
499 }
500
501 //convert resulting path from backward-linked list
502 stack_len = 0;
503 idx = 0;
504 for (cb = 1; cb < 12; cb++)
505 if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
506 idx = cb;
507 ppos = max_sfb;
508 while (ppos > 0) {
509 av_assert1(idx >= 0);
510 cb = idx;
511 stackrun[stack_len] = path[ppos][cb].run;
512 stackcb [stack_len] = cb;
513 idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
514 ppos -= path[ppos][cb].run;
515 stack_len++;
516 }
517 //perform actual band info encoding
518 start = 0;
519 for (i = stack_len - 1; i >= 0; i--) {
520 put_bits(&s->pb, 4, stackcb[i]);
521 count = stackrun[i];
522 memset(sce->zeroes + win*16 + start, !stackcb[i], count);
523 //XXX: memset when band_type is also uint8_t
524 for (j = 0; j < count; j++) {
525 sce->band_type[win*16 + start] = stackcb[i];
526 start++;
527 }
528 while (count >= run_esc) {
529 put_bits(&s->pb, run_bits, run_esc);
530 count -= run_esc;
531 }
532 put_bits(&s->pb, run_bits, count);
533 }
534 }
535
536 /** Return the minimum scalefactor where the quantized coef does not clip. */
537 static av_always_inline uint8_t coef2minsf(float coef) {
538 return av_clip_uint8(log2f(coef)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
539 }
540
541 /** Return the maximum scalefactor where the quantized coef is not zero. */
542 static av_always_inline uint8_t coef2maxsf(float coef) {
543 return av_clip_uint8(log2f(coef)*4 + 6 + SCALE_ONE_POS - SCALE_DIV_512);
544 }
545
546 typedef struct TrellisPath {
547 float cost;
548 int prev;
549 } TrellisPath;
550
551 #define TRELLIS_STAGES 121
552 #define TRELLIS_STATES (SCALE_MAX_DIFF+1)
553
554 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
555 SingleChannelElement *sce,
556 const float lambda)
557 {
558 int q, w, w2, g, start = 0;
559 int i, j;
560 int idx;
561 TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
562 int bandaddr[TRELLIS_STAGES];
563 int minq;
564 float mincost;
565 float q0f = FLT_MAX, q1f = 0.0f, qnrgf = 0.0f;
566 int q0, q1, qcnt = 0;
567
568 for (i = 0; i < 1024; i++) {
569 float t = fabsf(sce->coeffs[i]);
570 if (t > 0.0f) {
571 q0f = FFMIN(q0f, t);
572 q1f = FFMAX(q1f, t);
573 qnrgf += t*t;
574 qcnt++;
575 }
576 }
577
578 if (!qcnt) {
579 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
580 memset(sce->zeroes, 1, sizeof(sce->zeroes));
581 return;
582 }
583
584 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
585 q0 = coef2minsf(q0f);
586 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
587 q1 = coef2maxsf(q1f);
588 if (q1 - q0 > 60) {
589 int q0low = q0;
590 int q1high = q1;
591 //minimum scalefactor index is when maximum nonzero coefficient after quantizing is not clipped
592 int qnrg = av_clip_uint8(log2f(sqrtf(qnrgf/qcnt))*4 - 31 + SCALE_ONE_POS - SCALE_DIV_512);
593 q1 = qnrg + 30;
594 q0 = qnrg - 30;
595 if (q0 < q0low) {
596 q1 += q0low - q0;
597 q0 = q0low;
598 } else if (q1 > q1high) {
599 q0 -= q1 - q1high;
600 q1 = q1high;
601 }
602 }
603
604 for (i = 0; i < TRELLIS_STATES; i++) {
605 paths[0][i].cost = 0.0f;
606 paths[0][i].prev = -1;
607 }
608 for (j = 1; j < TRELLIS_STAGES; j++) {
609 for (i = 0; i < TRELLIS_STATES; i++) {
610 paths[j][i].cost = INFINITY;
611 paths[j][i].prev = -2;
612 }
613 }
614 idx = 1;
615 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
616 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
617 start = w*128;
618 for (g = 0; g < sce->ics.num_swb; g++) {
619 const float *coefs = sce->coeffs + start;
620 float qmin, qmax;
621 int nz = 0;
622
623 bandaddr[idx] = w * 16 + g;
624 qmin = INT_MAX;
625 qmax = 0.0f;
626 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
627 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
628 if (band->energy <= band->threshold || band->threshold == 0.0f) {
629 sce->zeroes[(w+w2)*16+g] = 1;
630 continue;
631 }
632 sce->zeroes[(w+w2)*16+g] = 0;
633 nz = 1;
634 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
635 float t = fabsf(coefs[w2*128+i]);
636 if (t > 0.0f)
637 qmin = FFMIN(qmin, t);
638 qmax = FFMAX(qmax, t);
639 }
640 }
641 if (nz) {
642 int minscale, maxscale;
643 float minrd = INFINITY;
644 float maxval;
645 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
646 minscale = coef2minsf(qmin);
647 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
648 maxscale = coef2maxsf(qmax);
649 minscale = av_clip(minscale - q0, 0, TRELLIS_STATES - 1);
650 maxscale = av_clip(maxscale - q0, 0, TRELLIS_STATES);
651 maxval = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], s->scoefs+start);
652 for (q = minscale; q < maxscale; q++) {
653 float dist = 0;
654 int cb = find_min_book(maxval, sce->sf_idx[w*16+g]);
655 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
656 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
657 dist += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
658 q + q0, cb, lambda / band->threshold, INFINITY, NULL);
659 }
660 minrd = FFMIN(minrd, dist);
661
662 for (i = 0; i < q1 - q0; i++) {
663 float cost;
664 cost = paths[idx - 1][i].cost + dist
665 + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
666 if (cost < paths[idx][q].cost) {
667 paths[idx][q].cost = cost;
668 paths[idx][q].prev = i;
669 }
670 }
671 }
672 } else {
673 for (q = 0; q < q1 - q0; q++) {
674 paths[idx][q].cost = paths[idx - 1][q].cost + 1;
675 paths[idx][q].prev = q;
676 }
677 }
678 sce->zeroes[w*16+g] = !nz;
679 start += sce->ics.swb_sizes[g];
680 idx++;
681 }
682 }
683 idx--;
684 mincost = paths[idx][0].cost;
685 minq = 0;
686 for (i = 1; i < TRELLIS_STATES; i++) {
687 if (paths[idx][i].cost < mincost) {
688 mincost = paths[idx][i].cost;
689 minq = i;
690 }
691 }
692 while (idx) {
693 sce->sf_idx[bandaddr[idx]] = minq + q0;
694 minq = paths[idx][minq].prev;
695 idx--;
696 }
697 //set the same quantizers inside window groups
698 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
699 for (g = 0; g < sce->ics.num_swb; g++)
700 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
701 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
702 }
703
704 /**
705 * two-loop quantizers search taken from ISO 13818-7 Appendix C
706 */
707 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
708 AACEncContext *s,
709 SingleChannelElement *sce,
710 const float lambda)
711 {
712 int start = 0, i, w, w2, g;
713 int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels * (lambda / 120.f);
714 float dists[128] = { 0 }, uplims[128];
715 float maxvals[128];
716 int fflag, minscaler;
717 int its = 0;
718 int allz = 0;
719 float minthr = INFINITY;
720
721 // for values above this the decoder might end up in an endless loop
722 // due to always having more bits than what can be encoded.
723 destbits = FFMIN(destbits, 5800);
724 //XXX: some heuristic to determine initial quantizers will reduce search time
725 //determine zero bands and upper limits
726 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
727 for (g = 0; g < sce->ics.num_swb; g++) {
728 int nz = 0;
729 float uplim = 0.0f;
730 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
731 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
732 uplim += band->threshold;
733 if (band->energy <= band->threshold || band->threshold == 0.0f) {
734 sce->zeroes[(w+w2)*16+g] = 1;
735 continue;
736 }
737 nz = 1;
738 }
739 uplims[w*16+g] = uplim *512;
740 sce->zeroes[w*16+g] = !nz;
741 if (nz)
742 minthr = FFMIN(minthr, uplim);
743 allz |= nz;
744 }
745 }
746 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
747 for (g = 0; g < sce->ics.num_swb; g++) {
748 if (sce->zeroes[w*16+g]) {
749 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
750 continue;
751 }
752 sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2f(uplims[w*16+g]/minthr)*4,59);
753 }
754 }
755
756 if (!allz)
757 return;
758 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
759
760 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
761 start = w*128;
762 for (g = 0; g < sce->ics.num_swb; g++) {
763 const float *scaled = s->scoefs + start;
764 maxvals[w*16+g] = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled);
765 start += sce->ics.swb_sizes[g];
766 }
767 }
768
769 //perform two-loop search
770 //outer loop - improve quality
771 do {
772 int tbits, qstep;
773 minscaler = sce->sf_idx[0];
774 //inner loop - quantize spectrum to fit into given number of bits
775 qstep = its ? 1 : 32;
776 do {
777 int prev = -1;
778 tbits = 0;
779 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
780 start = w*128;
781 for (g = 0; g < sce->ics.num_swb; g++) {
782 const float *coefs = sce->coeffs + start;
783 const float *scaled = s->scoefs + start;
784 int bits = 0;
785 int cb;
786 float dist = 0.0f;
787
788 if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
789 start += sce->ics.swb_sizes[g];
790 continue;
791 }
792 minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
793 cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
794 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
795 int b;
796 dist += quantize_band_cost(s, coefs + w2*128,
797 scaled + w2*128,
798 sce->ics.swb_sizes[g],
799 sce->sf_idx[w*16+g],
800 cb,
801 1.0f,
802 INFINITY,
803 &b);
804 bits += b;
805 }
806 dists[w*16+g] = dist - bits;
807 if (prev != -1) {
808 bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
809 }
810 tbits += bits;
811 start += sce->ics.swb_sizes[g];
812 prev = sce->sf_idx[w*16+g];
813 }
814 }
815 if (tbits > destbits) {
816 for (i = 0; i < 128; i++)
817 if (sce->sf_idx[i] < 218 - qstep)
818 sce->sf_idx[i] += qstep;
819 } else {
820 for (i = 0; i < 128; i++)
821 if (sce->sf_idx[i] > 60 - qstep)
822 sce->sf_idx[i] -= qstep;
823 }
824 qstep >>= 1;
825 if (!qstep && tbits > destbits*1.02 && sce->sf_idx[0] < 217)
826 qstep = 1;
827 } while (qstep);
828
829 fflag = 0;
830 minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
831 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
832 for (g = 0; g < sce->ics.num_swb; g++) {
833 int prevsc = sce->sf_idx[w*16+g];
834 if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60) {
835 if (find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]-1))
836 sce->sf_idx[w*16+g]--;
837 else //Try to make sure there is some energy in every band
838 sce->sf_idx[w*16+g]-=2;
839 }
840 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
841 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
842 if (sce->sf_idx[w*16+g] != prevsc)
843 fflag = 1;
844 sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
845 }
846 }
847 its++;
848 } while (fflag && its < 10);
849 }
850
851 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
852 SingleChannelElement *sce,
853 const float lambda)
854 {
855 int start = 0, i, w, w2, g;
856 float uplim[128], maxq[128];
857 int minq, maxsf;
858 float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
859 int last = 0, lastband = 0, curband = 0;
860 float avg_energy = 0.0;
861 if (sce->ics.num_windows == 1) {
862 start = 0;
863 for (i = 0; i < 1024; i++) {
864 if (i - start >= sce->ics.swb_sizes[curband]) {
865 start += sce->ics.swb_sizes[curband];
866 curband++;
867 }
868 if (sce->coeffs[i]) {
869 avg_energy += sce->coeffs[i] * sce->coeffs[i];
870 last = i;
871 lastband = curband;
872 }
873 }
874 } else {
875 for (w = 0; w < 8; w++) {
876 const float *coeffs = sce->coeffs + w*128;
877 curband = start = 0;
878 for (i = 0; i < 128; i++) {
879 if (i - start >= sce->ics.swb_sizes[curband]) {
880 start += sce->ics.swb_sizes[curband];
881 curband++;
882 }
883 if (coeffs[i]) {
884 avg_energy += coeffs[i] * coeffs[i];
885 last = FFMAX(last, i);
886 lastband = FFMAX(lastband, curband);
887 }
888 }
889 }
890 }
891 last++;
892 avg_energy /= last;
893 if (avg_energy == 0.0f) {
894 for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
895 sce->sf_idx[i] = SCALE_ONE_POS;
896 return;
897 }
898 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
899 start = w*128;
900 for (g = 0; g < sce->ics.num_swb; g++) {
901 float *coefs = sce->coeffs + start;
902 const int size = sce->ics.swb_sizes[g];
903 int start2 = start, end2 = start + size, peakpos = start;
904 float maxval = -1, thr = 0.0f, t;
905 maxq[w*16+g] = 0.0f;
906 if (g > lastband) {
907 maxq[w*16+g] = 0.0f;
908 start += size;
909 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
910 memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
911 continue;
912 }
913 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
914 for (i = 0; i < size; i++) {
915 float t = coefs[w2*128+i]*coefs[w2*128+i];
916 maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
917 thr += t;
918 if (sce->ics.num_windows == 1 && maxval < t) {
919 maxval = t;
920 peakpos = start+i;
921 }
922 }
923 }
924 if (sce->ics.num_windows == 1) {
925 start2 = FFMAX(peakpos - 2, start2);
926 end2 = FFMIN(peakpos + 3, end2);
927 } else {
928 start2 -= start;
929 end2 -= start;
930 }
931 start += size;
932 thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
933 t = 1.0 - (1.0 * start2 / last);
934 uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
935 }
936 }
937 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
938 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
939 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
940 start = w*128;
941 for (g = 0; g < sce->ics.num_swb; g++) {
942 const float *coefs = sce->coeffs + start;
943 const float *scaled = s->scoefs + start;
944 const int size = sce->ics.swb_sizes[g];
945 int scf, prev_scf, step;
946 int min_scf = -1, max_scf = 256;
947 float curdiff;
948 if (maxq[w*16+g] < 21.544) {
949 sce->zeroes[w*16+g] = 1;
950 start += size;
951 continue;
952 }
953 sce->zeroes[w*16+g] = 0;
954 scf = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2f(1/maxq[w*16+g])*16/3, 60, 218);
955 for (;;) {
956 float dist = 0.0f;
957 int quant_max;
958
959 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
960 int b;
961 dist += quantize_band_cost(s, coefs + w2*128,
962 scaled + w2*128,
963 sce->ics.swb_sizes[g],
964 scf,
965 ESC_BT,
966 lambda,
967 INFINITY,
968 &b);
969 dist -= b;
970 }
971 dist *= 1.0f / 512.0f / lambda;
972 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[POW_SF2_ZERO - scf + SCALE_ONE_POS - SCALE_DIV_512]);
973 if (quant_max >= 8191) { // too much, return to the previous quantizer
974 sce->sf_idx[w*16+g] = prev_scf;
975 break;
976 }
977 prev_scf = scf;
978 curdiff = fabsf(dist - uplim[w*16+g]);
979 if (curdiff <= 1.0f)
980 step = 0;
981 else
982 step = log2f(curdiff);
983 if (dist > uplim[w*16+g])
984 step = -step;
985 scf += step;
986 scf = av_clip_uint8(scf);
987 step = scf - prev_scf;
988 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
989 sce->sf_idx[w*16+g] = av_clip(scf, min_scf, max_scf);
990 break;
991 }
992 if (step > 0)
993 min_scf = prev_scf;
994 else
995 max_scf = prev_scf;
996 }
997 start += size;
998 }
999 }
1000 minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
1001 for (i = 1; i < 128; i++) {
1002 if (!sce->sf_idx[i])
1003 sce->sf_idx[i] = sce->sf_idx[i-1];
1004 else
1005 minq = FFMIN(minq, sce->sf_idx[i]);
1006 }
1007 if (minq == INT_MAX)
1008 minq = 0;
1009 minq = FFMIN(minq, SCALE_MAX_POS);
1010 maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
1011 for (i = 126; i >= 0; i--) {
1012 if (!sce->sf_idx[i])
1013 sce->sf_idx[i] = sce->sf_idx[i+1];
1014 sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
1015 }
1016 }
1017
1018 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
1019 SingleChannelElement *sce,
1020 const float lambda)
1021 {
1022 int i, w, w2, g;
1023 int minq = 255;
1024
1025 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
1026 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
1027 for (g = 0; g < sce->ics.num_swb; g++) {
1028 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
1029 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
1030 if (band->energy <= band->threshold) {
1031 sce->sf_idx[(w+w2)*16+g] = 218;
1032 sce->zeroes[(w+w2)*16+g] = 1;
1033 } else {
1034 sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2f(band->threshold), 80, 218);
1035 sce->zeroes[(w+w2)*16+g] = 0;
1036 }
1037 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
1038 }
1039 }
1040 }
1041 for (i = 0; i < 128; i++) {
1042 sce->sf_idx[i] = 140;
1043 //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
1044 }
1045 //set the same quantizers inside window groups
1046 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
1047 for (g = 0; g < sce->ics.num_swb; g++)
1048 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
1049 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
1050 }
1051
1052 static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
1053 const float lambda)
1054 {
1055 int start = 0, i, w, w2, g;
1056 float M[128], S[128];
1057 float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
1058 SingleChannelElement *sce0 = &cpe->ch[0];
1059 SingleChannelElement *sce1 = &cpe->ch[1];
1060 if (!cpe->common_window)
1061 return;
1062 for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
1063 for (g = 0; g < sce0->ics.num_swb; g++) {
1064 if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
1065 float dist1 = 0.0f, dist2 = 0.0f;
1066 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
1067 FFPsyBand *band0 = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g];
1068 FFPsyBand *band1 = &s->psy.ch[s->cur_channel+1].psy_bands[(w+w2)*16+g];
1069 float minthr = FFMIN(band0->threshold, band1->threshold);
1070 float maxthr = FFMAX(band0->threshold, band1->threshold);
1071 for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
1072 M[i] = (sce0->coeffs[start+w2*128+i]
1073 + sce1->coeffs[start+w2*128+i]) * 0.5;
1074 S[i] = M[i]
1075 - sce1->coeffs[start+w2*128+i];
1076 }
1077 abs_pow34_v(L34, sce0->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
1078 abs_pow34_v(R34, sce1->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
1079 abs_pow34_v(M34, M, sce0->ics.swb_sizes[g]);
1080 abs_pow34_v(S34, S, sce0->ics.swb_sizes[g]);
1081 dist1 += quantize_band_cost(s, sce0->coeffs + start + w2*128,
1082 L34,
1083 sce0->ics.swb_sizes[g],
1084 sce0->sf_idx[(w+w2)*16+g],
1085 sce0->band_type[(w+w2)*16+g],
1086 lambda / band0->threshold, INFINITY, NULL);
1087 dist1 += quantize_band_cost(s, sce1->coeffs + start + w2*128,
1088 R34,
1089 sce1->ics.swb_sizes[g],
1090 sce1->sf_idx[(w+w2)*16+g],
1091 sce1->band_type[(w+w2)*16+g],
1092 lambda / band1->threshold, INFINITY, NULL);
1093 dist2 += quantize_band_cost(s, M,
1094 M34,
1095 sce0->ics.swb_sizes[g],
1096 sce0->sf_idx[(w+w2)*16+g],
1097 sce0->band_type[(w+w2)*16+g],
1098 lambda / maxthr, INFINITY, NULL);
1099 dist2 += quantize_band_cost(s, S,
1100 S34,
1101 sce1->ics.swb_sizes[g],
1102 sce1->sf_idx[(w+w2)*16+g],
1103 sce1->band_type[(w+w2)*16+g],
1104 lambda / minthr, INFINITY, NULL);
1105 }
1106 cpe->ms_mask[w*16+g] = dist2 < dist1;
1107 }
1108 start += sce0->ics.swb_sizes[g];
1109 }
1110 }
1111 }
1112
1113 AACCoefficientsEncoder ff_aac_coders[AAC_CODER_NB] = {
1114 [AAC_CODER_FAAC] = {
1115 search_for_quantizers_faac,
1116 encode_window_bands_info,
1117 quantize_and_encode_band,
1118 search_for_ms,
1119 },
1120 [AAC_CODER_ANMR] = {
1121 search_for_quantizers_anmr,
1122 encode_window_bands_info,
1123 quantize_and_encode_band,
1124 search_for_ms,
1125 },
1126 [AAC_CODER_TWOLOOP] = {
1127 search_for_quantizers_twoloop,
1128 codebook_trellis_rate,
1129 quantize_and_encode_band,
1130 search_for_ms,
1131 },
1132 [AAC_CODER_FAST] = {
1133 search_for_quantizers_fast,
1134 encode_window_bands_info,
1135 quantize_and_encode_band,
1136 search_for_ms,
1137 },
1138 };