2 * AAC coefficients encoder
3 * Copyright (C) 2008-2009 Konstantin Shishkov
5 * This file is part of FFmpeg.
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.
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.
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
24 * AAC coefficients encoder
27 /***********************************
29 * speedup quantizer selection
30 * add sane pulse detection
31 ***********************************/
33 #include "libavutil/libm.h" // brought forward to work around cygwin header breakage
36 #include "libavutil/mathematics.h"
43 /** Frequency in Hz for lower limit of noise substitution **/
44 #define NOISE_LOW_LIMIT 4000
46 /** Total number of usable codebooks **/
49 /** Total number of codebooks, including special ones **/
52 /** bits needed to code codebook run value for long windows */
53 static const uint8_t run_value_bits_long[64] = {
54 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
55 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 10,
56 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
57 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 15
60 /** bits needed to code codebook run value for short windows */
61 static const uint8_t run_value_bits_short[16] = {
62 3, 3, 3, 3, 3, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 9
65 static const uint8_t * const run_value_bits[2] = {
66 run_value_bits_long, run_value_bits_short
69 /** Map to convert values from BandCodingPath index to a codebook index **/
70 static const uint8_t aac_cb_out_map[CB_TOT_ALL] = {0,1,2,3,4,5,6,7,8,9,10,11,13,14,15};
71 /** Inverse map to convert from codebooks to BandCodingPath indices **/
72 static const uint8_t aac_cb_in_map[CB_TOT_ALL+1] = {0,1,2,3,4,5,6,7,8,9,10,11,0,12,13,14};
75 * Quantize one coefficient.
76 * @return absolute value of the quantized coefficient
77 * @see 3GPP TS26.403 5.6.2 "Scalefactor determination"
79 static av_always_inline int quant(float coef, const float Q)
82 return sqrtf(a * sqrtf(a)) + 0.4054;
85 static void quantize_bands(int *out, const float *in, const float *scaled,
86 int size, float Q34, int is_signed, int maxval)
90 for (i = 0; i < size; i++) {
92 out[i] = (int)FFMIN(qc + 0.4054, (double)maxval);
93 if (is_signed && in[i] < 0.0f) {
99 static void abs_pow34_v(float *out, const float *in, const int size)
101 #ifndef USE_REALLY_FULL_SEARCH
103 for (i = 0; i < size; i++) {
104 float a = fabsf(in[i]);
105 out[i] = sqrtf(a * sqrtf(a));
107 #endif /* USE_REALLY_FULL_SEARCH */
110 static const uint8_t aac_cb_range [12] = {0, 3, 3, 3, 3, 9, 9, 8, 8, 13, 13, 17};
111 static const uint8_t aac_cb_maxval[12] = {0, 1, 1, 2, 2, 4, 4, 7, 7, 12, 12, 16};
114 * Calculate rate distortion cost for quantizing with given codebook
116 * @return quantization distortion
118 static av_always_inline float quantize_and_encode_band_cost_template(
119 struct AACEncContext *s,
120 PutBitContext *pb, const float *in,
121 const float *scaled, int size, int scale_idx,
122 int cb, const float lambda, const float uplim,
123 int *bits, int BT_ZERO, int BT_UNSIGNED,
124 int BT_PAIR, int BT_ESC, int BT_NOISE, int BT_STEREO)
126 const int q_idx = POW_SF2_ZERO - scale_idx + SCALE_ONE_POS - SCALE_DIV_512;
127 const float Q = ff_aac_pow2sf_tab [q_idx];
128 const float Q34 = ff_aac_pow34sf_tab[q_idx];
129 const float IQ = ff_aac_pow2sf_tab [POW_SF2_ZERO + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
130 const float CLIPPED_ESCAPE = 165140.0f*IQ;
133 const int dim = BT_PAIR ? 2 : 4;
137 if (BT_ZERO || BT_NOISE || BT_STEREO) {
138 for (i = 0; i < size; i++)
142 return cost * lambda;
145 abs_pow34_v(s->scoefs, in, size);
148 quantize_bands(s->qcoefs, in, scaled, size, Q34, !BT_UNSIGNED, aac_cb_maxval[cb]);
152 off = aac_cb_maxval[cb];
154 for (i = 0; i < size; i += dim) {
156 int *quants = s->qcoefs + i;
160 for (j = 0; j < dim; j++) {
161 curidx *= aac_cb_range[cb];
162 curidx += quants[j] + off;
164 curbits = ff_aac_spectral_bits[cb-1][curidx];
165 vec = &ff_aac_codebook_vectors[cb-1][curidx*dim];
167 for (j = 0; j < dim; j++) {
168 float t = fabsf(in[i+j]);
170 if (BT_ESC && vec[j] == 64.0f) { //FIXME: slow
171 if (t >= CLIPPED_ESCAPE) {
172 di = t - CLIPPED_ESCAPE;
175 int c = av_clip_uintp2(quant(t, Q), 13);
176 di = t - c*cbrtf(c)*IQ;
177 curbits += av_log2(c)*2 - 4 + 1;
187 for (j = 0; j < dim; j++) {
188 float di = in[i+j] - vec[j]*IQ;
192 cost += rd * lambda + curbits;
197 put_bits(pb, ff_aac_spectral_bits[cb-1][curidx], ff_aac_spectral_codes[cb-1][curidx]);
199 for (j = 0; j < dim; j++)
200 if (ff_aac_codebook_vectors[cb-1][curidx*dim+j] != 0.0f)
201 put_bits(pb, 1, in[i+j] < 0.0f);
203 for (j = 0; j < 2; j++) {
204 if (ff_aac_codebook_vectors[cb-1][curidx*2+j] == 64.0f) {
205 int coef = av_clip_uintp2(quant(fabsf(in[i+j]), Q), 13);
206 int len = av_log2(coef);
208 put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
209 put_sbits(pb, len, coef);
221 static float quantize_and_encode_band_cost_NONE(struct AACEncContext *s, PutBitContext *pb,
222 const float *in, const float *scaled,
223 int size, int scale_idx, int cb,
224 const float lambda, const float uplim,
230 #define QUANTIZE_AND_ENCODE_BAND_COST_FUNC(NAME, BT_ZERO, BT_UNSIGNED, BT_PAIR, BT_ESC, BT_NOISE, BT_STEREO) \
231 static float quantize_and_encode_band_cost_ ## NAME( \
232 struct AACEncContext *s, \
233 PutBitContext *pb, const float *in, \
234 const float *scaled, int size, int scale_idx, \
235 int cb, const float lambda, const float uplim, \
237 return quantize_and_encode_band_cost_template( \
238 s, pb, in, scaled, size, scale_idx, \
239 BT_ESC ? ESC_BT : cb, lambda, uplim, bits, \
240 BT_ZERO, BT_UNSIGNED, BT_PAIR, BT_ESC, BT_NOISE, BT_STEREO); \
243 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(ZERO, 1, 0, 0, 0, 0, 0)
244 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(SQUAD, 0, 0, 0, 0, 0, 0)
245 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(UQUAD, 0, 1, 0, 0, 0, 0)
246 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(SPAIR, 0, 0, 1, 0, 0, 0)
247 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(UPAIR, 0, 1, 1, 0, 0, 0)
248 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(ESC, 0, 1, 1, 1, 0, 0)
249 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(NOISE, 0, 0, 0, 0, 1, 0)
250 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(STEREO,0, 0, 0, 0, 0, 1)
252 static float (*const quantize_and_encode_band_cost_arr[])(
253 struct AACEncContext *s,
254 PutBitContext *pb, const float *in,
255 const float *scaled, int size, int scale_idx,
256 int cb, const float lambda, const float uplim,
258 quantize_and_encode_band_cost_ZERO,
259 quantize_and_encode_band_cost_SQUAD,
260 quantize_and_encode_band_cost_SQUAD,
261 quantize_and_encode_band_cost_UQUAD,
262 quantize_and_encode_band_cost_UQUAD,
263 quantize_and_encode_band_cost_SPAIR,
264 quantize_and_encode_band_cost_SPAIR,
265 quantize_and_encode_band_cost_UPAIR,
266 quantize_and_encode_band_cost_UPAIR,
267 quantize_and_encode_band_cost_UPAIR,
268 quantize_and_encode_band_cost_UPAIR,
269 quantize_and_encode_band_cost_ESC,
270 quantize_and_encode_band_cost_NONE, /* CB 12 doesn't exist */
271 quantize_and_encode_band_cost_NOISE,
272 quantize_and_encode_band_cost_STEREO,
273 quantize_and_encode_band_cost_STEREO,
276 #define quantize_and_encode_band_cost( \
277 s, pb, in, scaled, size, scale_idx, cb, \
278 lambda, uplim, bits) \
279 quantize_and_encode_band_cost_arr[cb]( \
280 s, pb, in, scaled, size, scale_idx, cb, \
283 static float quantize_band_cost(struct AACEncContext *s, const float *in,
284 const float *scaled, int size, int scale_idx,
285 int cb, const float lambda, const float uplim,
288 return quantize_and_encode_band_cost(s, NULL, in, scaled, size, scale_idx,
289 cb, lambda, uplim, bits);
292 static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
293 const float *in, int size, int scale_idx,
294 int cb, const float lambda)
296 quantize_and_encode_band_cost(s, pb, in, NULL, size, scale_idx, cb, lambda,
300 static float find_max_val(int group_len, int swb_size, const float *scaled) {
303 for (w2 = 0; w2 < group_len; w2++) {
304 for (i = 0; i < swb_size; i++) {
305 maxval = FFMAX(maxval, scaled[w2*128+i]);
311 static int find_min_book(float maxval, int sf) {
312 float Q = ff_aac_pow2sf_tab[POW_SF2_ZERO - sf + SCALE_ONE_POS - SCALE_DIV_512];
313 float Q34 = sqrtf(Q * sqrtf(Q));
315 qmaxval = maxval * Q34 + 0.4054f;
316 if (qmaxval == 0) cb = 0;
317 else if (qmaxval == 1) cb = 1;
318 else if (qmaxval == 2) cb = 3;
319 else if (qmaxval <= 4) cb = 5;
320 else if (qmaxval <= 7) cb = 7;
321 else if (qmaxval <= 12) cb = 9;
327 * structure used in optimal codebook search
329 typedef struct BandCodingPath {
330 int prev_idx; ///< pointer to the previous path point
331 float cost; ///< path cost
336 * Encode band info for single window group bands.
338 static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
339 int win, int group_len, const float lambda)
341 BandCodingPath path[120][CB_TOT_ALL];
342 int w, swb, cb, start, size;
344 const int max_sfb = sce->ics.max_sfb;
345 const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
346 const int run_esc = (1 << run_bits) - 1;
347 int idx, ppos, count;
348 int stackrun[120], stackcb[120], stack_len;
349 float next_minrd = INFINITY;
352 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
354 for (cb = 0; cb < CB_TOT_ALL; cb++) {
355 path[0][cb].cost = 0.0f;
356 path[0][cb].prev_idx = -1;
359 for (swb = 0; swb < max_sfb; swb++) {
360 size = sce->ics.swb_sizes[swb];
361 if (sce->zeroes[win*16 + swb]) {
362 for (cb = 0; cb < CB_TOT_ALL; cb++) {
363 path[swb+1][cb].prev_idx = cb;
364 path[swb+1][cb].cost = path[swb][cb].cost;
365 path[swb+1][cb].run = path[swb][cb].run + 1;
368 float minrd = next_minrd;
369 int mincb = next_mincb;
370 next_minrd = INFINITY;
372 for (cb = 0; cb < CB_TOT_ALL; cb++) {
373 float cost_stay_here, cost_get_here;
375 if (cb >= 12 && sce->band_type[win*16+swb] < aac_cb_out_map[cb] ||
376 cb < aac_cb_in_map[sce->band_type[win*16+swb]] && sce->band_type[win*16+swb] > aac_cb_out_map[cb]) {
377 path[swb+1][cb].prev_idx = -1;
378 path[swb+1][cb].cost = INFINITY;
379 path[swb+1][cb].run = path[swb][cb].run + 1;
382 for (w = 0; w < group_len; w++) {
383 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(win+w)*16+swb];
384 rd += quantize_band_cost(s, sce->coeffs + start + w*128,
385 s->scoefs + start + w*128, size,
386 sce->sf_idx[(win+w)*16+swb], aac_cb_out_map[cb],
387 lambda / band->threshold, INFINITY, NULL);
389 cost_stay_here = path[swb][cb].cost + rd;
390 cost_get_here = minrd + rd + run_bits + 4;
391 if ( run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
392 != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
393 cost_stay_here += run_bits;
394 if (cost_get_here < cost_stay_here) {
395 path[swb+1][cb].prev_idx = mincb;
396 path[swb+1][cb].cost = cost_get_here;
397 path[swb+1][cb].run = 1;
399 path[swb+1][cb].prev_idx = cb;
400 path[swb+1][cb].cost = cost_stay_here;
401 path[swb+1][cb].run = path[swb][cb].run + 1;
403 if (path[swb+1][cb].cost < next_minrd) {
404 next_minrd = path[swb+1][cb].cost;
409 start += sce->ics.swb_sizes[swb];
412 //convert resulting path from backward-linked list
415 for (cb = 1; cb < CB_TOT_ALL; cb++)
416 if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
420 av_assert1(idx >= 0);
422 stackrun[stack_len] = path[ppos][cb].run;
423 stackcb [stack_len] = cb;
424 idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
425 ppos -= path[ppos][cb].run;
428 //perform actual band info encoding
430 for (i = stack_len - 1; i >= 0; i--) {
431 cb = aac_cb_out_map[stackcb[i]];
432 put_bits(&s->pb, 4, cb);
434 memset(sce->zeroes + win*16 + start, !cb, count);
435 //XXX: memset when band_type is also uint8_t
436 for (j = 0; j < count; j++) {
437 sce->band_type[win*16 + start] = cb;
440 while (count >= run_esc) {
441 put_bits(&s->pb, run_bits, run_esc);
444 put_bits(&s->pb, run_bits, count);
448 static void codebook_trellis_rate(AACEncContext *s, SingleChannelElement *sce,
449 int win, int group_len, const float lambda)
451 BandCodingPath path[120][CB_TOT_ALL];
452 int w, swb, cb, start, size;
454 const int max_sfb = sce->ics.max_sfb;
455 const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
456 const int run_esc = (1 << run_bits) - 1;
457 int idx, ppos, count;
458 int stackrun[120], stackcb[120], stack_len;
459 float next_minbits = INFINITY;
462 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
464 for (cb = 0; cb < CB_TOT_ALL; cb++) {
465 path[0][cb].cost = run_bits+4;
466 path[0][cb].prev_idx = -1;
469 for (swb = 0; swb < max_sfb; swb++) {
470 size = sce->ics.swb_sizes[swb];
471 if (sce->zeroes[win*16 + swb]) {
472 float cost_stay_here = path[swb][0].cost;
473 float cost_get_here = next_minbits + run_bits + 4;
474 if ( run_value_bits[sce->ics.num_windows == 8][path[swb][0].run]
475 != run_value_bits[sce->ics.num_windows == 8][path[swb][0].run+1])
476 cost_stay_here += run_bits;
477 if (cost_get_here < cost_stay_here) {
478 path[swb+1][0].prev_idx = next_mincb;
479 path[swb+1][0].cost = cost_get_here;
480 path[swb+1][0].run = 1;
482 path[swb+1][0].prev_idx = 0;
483 path[swb+1][0].cost = cost_stay_here;
484 path[swb+1][0].run = path[swb][0].run + 1;
486 next_minbits = path[swb+1][0].cost;
488 for (cb = 1; cb < CB_TOT_ALL; cb++) {
489 path[swb+1][cb].cost = 61450;
490 path[swb+1][cb].prev_idx = -1;
491 path[swb+1][cb].run = 0;
494 float minbits = next_minbits;
495 int mincb = next_mincb;
496 int startcb = sce->band_type[win*16+swb];
497 startcb = aac_cb_in_map[startcb];
498 next_minbits = INFINITY;
500 for (cb = 0; cb < startcb; cb++) {
501 path[swb+1][cb].cost = 61450;
502 path[swb+1][cb].prev_idx = -1;
503 path[swb+1][cb].run = 0;
505 for (cb = startcb; cb < CB_TOT_ALL; cb++) {
506 float cost_stay_here, cost_get_here;
508 if (cb >= 12 && sce->band_type[win*16+swb] != aac_cb_out_map[cb]) {
509 path[swb+1][cb].cost = 61450;
510 path[swb+1][cb].prev_idx = -1;
511 path[swb+1][cb].run = 0;
514 for (w = 0; w < group_len; w++) {
515 bits += quantize_band_cost(s, sce->coeffs + start + w*128,
516 s->scoefs + start + w*128, size,
517 sce->sf_idx[(win+w)*16+swb],
521 cost_stay_here = path[swb][cb].cost + bits;
522 cost_get_here = minbits + bits + run_bits + 4;
523 if ( run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
524 != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
525 cost_stay_here += run_bits;
526 if (cost_get_here < cost_stay_here) {
527 path[swb+1][cb].prev_idx = mincb;
528 path[swb+1][cb].cost = cost_get_here;
529 path[swb+1][cb].run = 1;
531 path[swb+1][cb].prev_idx = cb;
532 path[swb+1][cb].cost = cost_stay_here;
533 path[swb+1][cb].run = path[swb][cb].run + 1;
535 if (path[swb+1][cb].cost < next_minbits) {
536 next_minbits = path[swb+1][cb].cost;
541 start += sce->ics.swb_sizes[swb];
544 //convert resulting path from backward-linked list
547 for (cb = 1; cb < CB_TOT_ALL; cb++)
548 if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
552 av_assert1(idx >= 0);
554 stackrun[stack_len] = path[ppos][cb].run;
555 stackcb [stack_len] = cb;
556 idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
557 ppos -= path[ppos][cb].run;
560 //perform actual band info encoding
562 for (i = stack_len - 1; i >= 0; i--) {
563 cb = aac_cb_out_map[stackcb[i]];
564 put_bits(&s->pb, 4, cb);
566 memset(sce->zeroes + win*16 + start, !cb, count);
567 //XXX: memset when band_type is also uint8_t
568 for (j = 0; j < count; j++) {
569 sce->band_type[win*16 + start] = cb;
572 while (count >= run_esc) {
573 put_bits(&s->pb, run_bits, run_esc);
576 put_bits(&s->pb, run_bits, count);
580 /** Return the minimum scalefactor where the quantized coef does not clip. */
581 static av_always_inline uint8_t coef2minsf(float coef) {
582 return av_clip_uint8(log2f(coef)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
585 /** Return the maximum scalefactor where the quantized coef is not zero. */
586 static av_always_inline uint8_t coef2maxsf(float coef) {
587 return av_clip_uint8(log2f(coef)*4 + 6 + SCALE_ONE_POS - SCALE_DIV_512);
590 typedef struct TrellisPath {
595 #define TRELLIS_STAGES 121
596 #define TRELLIS_STATES (SCALE_MAX_DIFF+1)
598 static void set_special_band_scalefactors(AACEncContext *s, SingleChannelElement *sce)
601 int minscaler_n = sce->sf_idx[0], minscaler_i = sce->sf_idx[0];
604 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
606 for (g = 0; g < sce->ics.num_swb; g++) {
607 if (sce->band_type[w*16+g] == INTENSITY_BT || sce->band_type[w*16+g] == INTENSITY_BT2) {
608 sce->sf_idx[w*16+g] = av_clip(ceilf(log2f(sce->is_ener[w*16+g])*2), -155, 100);
609 minscaler_i = FFMIN(minscaler_i, sce->sf_idx[w*16+g]);
611 } else if (sce->band_type[w*16+g] == NOISE_BT) {
612 sce->sf_idx[w*16+g] = av_clip(4+log2f(sce->pns_ener[w*16+g])*2, -100, 155);
613 minscaler_n = FFMIN(minscaler_n, sce->sf_idx[w*16+g]);
616 start += sce->ics.swb_sizes[g];
623 /* Clip the scalefactor indices */
624 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
625 for (g = 0; g < sce->ics.num_swb; g++) {
626 if (sce->band_type[w*16+g] == INTENSITY_BT || sce->band_type[w*16+g] == INTENSITY_BT2) {
627 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler_i, minscaler_i + SCALE_MAX_DIFF);
628 } else if (sce->band_type[w*16+g] == NOISE_BT) {
629 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler_n, minscaler_n + SCALE_MAX_DIFF);
635 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
636 SingleChannelElement *sce,
639 int q, w, w2, g, start = 0;
642 TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
643 int bandaddr[TRELLIS_STAGES];
646 float q0f = FLT_MAX, q1f = 0.0f, qnrgf = 0.0f;
647 int q0, q1, qcnt = 0;
649 for (i = 0; i < 1024; i++) {
650 float t = fabsf(sce->coeffs[i]);
660 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
661 memset(sce->zeroes, 1, sizeof(sce->zeroes));
665 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
666 q0 = coef2minsf(q0f);
667 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
668 q1 = coef2maxsf(q1f);
672 //minimum scalefactor index is when maximum nonzero coefficient after quantizing is not clipped
673 int qnrg = av_clip_uint8(log2f(sqrtf(qnrgf/qcnt))*4 - 31 + SCALE_ONE_POS - SCALE_DIV_512);
679 } else if (q1 > q1high) {
685 for (i = 0; i < TRELLIS_STATES; i++) {
686 paths[0][i].cost = 0.0f;
687 paths[0][i].prev = -1;
689 for (j = 1; j < TRELLIS_STAGES; j++) {
690 for (i = 0; i < TRELLIS_STATES; i++) {
691 paths[j][i].cost = INFINITY;
692 paths[j][i].prev = -2;
696 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
697 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 const float *coefs = sce->coeffs + start;
704 bandaddr[idx] = w * 16 + g;
707 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
708 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
709 if (band->energy <= band->threshold || band->threshold == 0.0f) {
710 sce->zeroes[(w+w2)*16+g] = 1;
713 sce->zeroes[(w+w2)*16+g] = 0;
715 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
716 float t = fabsf(coefs[w2*128+i]);
718 qmin = FFMIN(qmin, t);
719 qmax = FFMAX(qmax, t);
723 int minscale, maxscale;
724 float minrd = INFINITY;
726 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
727 minscale = coef2minsf(qmin);
728 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
729 maxscale = coef2maxsf(qmax);
730 minscale = av_clip(minscale - q0, 0, TRELLIS_STATES - 1);
731 maxscale = av_clip(maxscale - q0, 0, TRELLIS_STATES);
732 maxval = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], s->scoefs+start);
733 for (q = minscale; q < maxscale; q++) {
735 int cb = find_min_book(maxval, sce->sf_idx[w*16+g]);
736 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
737 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
738 dist += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
739 q + q0, cb, lambda / band->threshold, INFINITY, NULL);
741 minrd = FFMIN(minrd, dist);
743 for (i = 0; i < q1 - q0; i++) {
745 cost = paths[idx - 1][i].cost + dist
746 + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
747 if (cost < paths[idx][q].cost) {
748 paths[idx][q].cost = cost;
749 paths[idx][q].prev = i;
754 for (q = 0; q < q1 - q0; q++) {
755 paths[idx][q].cost = paths[idx - 1][q].cost + 1;
756 paths[idx][q].prev = q;
759 sce->zeroes[w*16+g] = !nz;
760 start += sce->ics.swb_sizes[g];
765 mincost = paths[idx][0].cost;
767 for (i = 1; i < TRELLIS_STATES; i++) {
768 if (paths[idx][i].cost < mincost) {
769 mincost = paths[idx][i].cost;
774 sce->sf_idx[bandaddr[idx]] = minq + q0;
775 minq = paths[idx][minq].prev;
778 //set the same quantizers inside window groups
779 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
780 for (g = 0; g < sce->ics.num_swb; g++)
781 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
782 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
786 * two-loop quantizers search taken from ISO 13818-7 Appendix C
788 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
790 SingleChannelElement *sce,
793 int start = 0, i, w, w2, g;
794 int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels * (lambda / 120.f);
795 const float freq_mult = avctx->sample_rate/(1024.0f/sce->ics.num_windows)/2.0f;
796 float dists[128] = { 0 }, uplims[128] = { 0 };
798 int noise_sf[128] = { 0 };
799 int fflag, minscaler, minscaler_n;
802 float minthr = INFINITY;
804 // for values above this the decoder might end up in an endless loop
805 // due to always having more bits than what can be encoded.
806 destbits = FFMIN(destbits, 5800);
807 //XXX: some heuristic to determine initial quantizers will reduce search time
808 //determine zero bands and upper limits
809 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
811 for (g = 0; g < sce->ics.num_swb; g++) {
813 float uplim = 0.0f, energy = 0.0f;
814 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
815 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
816 uplim += band->threshold;
817 energy += band->energy;
818 if (band->energy <= band->threshold || band->threshold == 0.0f) {
819 sce->zeroes[(w+w2)*16+g] = 1;
824 uplims[w*16+g] = uplim *512;
825 if (s->options.pns && start*freq_mult > NOISE_LOW_LIMIT && energy < uplim * 1.2f) {
826 noise_sf[w*16+g] = av_clip(4+FFMIN(log2f(energy)*2,255), -100, 155);
827 sce->band_type[w*16+g] = NOISE_BT;
829 } else { /** Band type will be determined by the twoloop algorithm */
830 sce->band_type[w*16+g] = 0;
832 sce->zeroes[w*16+g] = !nz;
834 minthr = FFMIN(minthr, uplim);
836 start += sce->ics.swb_sizes[g];
839 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
840 for (g = 0; g < sce->ics.num_swb; g++) {
841 if (sce->zeroes[w*16+g]) {
842 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
845 sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2f(uplims[w*16+g]/minthr)*4,59);
851 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
853 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
855 for (g = 0; g < sce->ics.num_swb; g++) {
856 const float *scaled = s->scoefs + start;
857 maxvals[w*16+g] = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled);
858 start += sce->ics.swb_sizes[g];
862 //perform two-loop search
863 //outer loop - improve quality
866 minscaler = sce->sf_idx[0];
867 minscaler_n = sce->sf_idx[0];
868 //inner loop - quantize spectrum to fit into given number of bits
869 qstep = its ? 1 : 32;
873 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
875 for (g = 0; g < sce->ics.num_swb; g++) {
876 const float *coefs = sce->coeffs + start;
877 const float *scaled = s->scoefs + start;
882 if (sce->band_type[w*16+g] == NOISE_BT) {
883 minscaler_n = FFMIN(minscaler_n, noise_sf[w*16+g]);
884 start += sce->ics.swb_sizes[g];
886 } else if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
887 start += sce->ics.swb_sizes[g];
890 minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
891 cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
892 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
894 dist += quantize_band_cost(s, coefs + w2*128,
896 sce->ics.swb_sizes[g],
904 dists[w*16+g] = dist - bits;
906 bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
909 start += sce->ics.swb_sizes[g];
910 prev = sce->sf_idx[w*16+g];
913 if (tbits > destbits) {
914 for (i = 0; i < 128; i++)
915 if (sce->sf_idx[i] < 218 - qstep)
916 sce->sf_idx[i] += qstep;
918 for (i = 0; i < 128; i++)
919 if (sce->sf_idx[i] > 60 - qstep)
920 sce->sf_idx[i] -= qstep;
923 if (!qstep && tbits > destbits*1.02 && sce->sf_idx[0] < 217)
928 minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
930 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
931 for (g = 0; g < sce->ics.num_swb; g++)
932 if (sce->band_type[w*16+g] == NOISE_BT)
933 sce->sf_idx[w*16+g] = av_clip(noise_sf[w*16+g], minscaler_n, minscaler_n + SCALE_MAX_DIFF);
935 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
936 for (g = 0; g < sce->ics.num_swb; g++) {
937 int prevsc = sce->sf_idx[w*16+g];
938 if (sce->band_type[w*16+g] == NOISE_BT)
940 if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60) {
941 if (find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]-1))
942 sce->sf_idx[w*16+g]--;
943 else //Try to make sure there is some energy in every band
944 sce->sf_idx[w*16+g]-=2;
946 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
947 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
948 if (sce->sf_idx[w*16+g] != prevsc)
950 sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
954 } while (fflag && its < 10);
957 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
958 SingleChannelElement *sce,
961 int start = 0, i, w, w2, g;
962 float uplim[128], maxq[128];
964 float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
965 int last = 0, lastband = 0, curband = 0;
966 float avg_energy = 0.0;
967 if (sce->ics.num_windows == 1) {
969 for (i = 0; i < 1024; i++) {
970 if (i - start >= sce->ics.swb_sizes[curband]) {
971 start += sce->ics.swb_sizes[curband];
974 if (sce->coeffs[i]) {
975 avg_energy += sce->coeffs[i] * sce->coeffs[i];
981 for (w = 0; w < 8; w++) {
982 const float *coeffs = sce->coeffs + w*128;
984 for (i = 0; i < 128; i++) {
985 if (i - start >= sce->ics.swb_sizes[curband]) {
986 start += sce->ics.swb_sizes[curband];
990 avg_energy += coeffs[i] * coeffs[i];
991 last = FFMAX(last, i);
992 lastband = FFMAX(lastband, curband);
999 if (avg_energy == 0.0f) {
1000 for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
1001 sce->sf_idx[i] = SCALE_ONE_POS;
1004 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
1006 for (g = 0; g < sce->ics.num_swb; g++) {
1007 float *coefs = sce->coeffs + start;
1008 const int size = sce->ics.swb_sizes[g];
1009 int start2 = start, end2 = start + size, peakpos = start;
1010 float maxval = -1, thr = 0.0f, t;
1011 maxq[w*16+g] = 0.0f;
1013 maxq[w*16+g] = 0.0f;
1015 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
1016 memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
1019 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
1020 for (i = 0; i < size; i++) {
1021 float t = coefs[w2*128+i]*coefs[w2*128+i];
1022 maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
1024 if (sce->ics.num_windows == 1 && maxval < t) {
1030 if (sce->ics.num_windows == 1) {
1031 start2 = FFMAX(peakpos - 2, start2);
1032 end2 = FFMIN(peakpos + 3, end2);
1038 thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
1039 t = 1.0 - (1.0 * start2 / last);
1040 uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
1043 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
1044 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
1045 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 const float *coefs = sce->coeffs + start;
1049 const float *scaled = s->scoefs + start;
1050 const int size = sce->ics.swb_sizes[g];
1051 int scf, prev_scf, step;
1052 int min_scf = -1, max_scf = 256;
1054 if (maxq[w*16+g] < 21.544) {
1055 sce->zeroes[w*16+g] = 1;
1059 sce->zeroes[w*16+g] = 0;
1060 scf = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2f(1/maxq[w*16+g])*16/3, 60, 218);
1065 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
1067 dist += quantize_band_cost(s, coefs + w2*128,
1069 sce->ics.swb_sizes[g],
1077 dist *= 1.0f / 512.0f / lambda;
1078 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[POW_SF2_ZERO - scf + SCALE_ONE_POS - SCALE_DIV_512]);
1079 if (quant_max >= 8191) { // too much, return to the previous quantizer
1080 sce->sf_idx[w*16+g] = prev_scf;
1084 curdiff = fabsf(dist - uplim[w*16+g]);
1085 if (curdiff <= 1.0f)
1088 step = log2f(curdiff);
1089 if (dist > uplim[w*16+g])
1092 scf = av_clip_uint8(scf);
1093 step = scf - prev_scf;
1094 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
1095 sce->sf_idx[w*16+g] = av_clip(scf, min_scf, max_scf);
1106 minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
1107 for (i = 1; i < 128; i++) {
1108 if (!sce->sf_idx[i])
1109 sce->sf_idx[i] = sce->sf_idx[i-1];
1111 minq = FFMIN(minq, sce->sf_idx[i]);
1113 if (minq == INT_MAX)
1115 minq = FFMIN(minq, SCALE_MAX_POS);
1116 maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
1117 for (i = 126; i >= 0; i--) {
1118 if (!sce->sf_idx[i])
1119 sce->sf_idx[i] = sce->sf_idx[i+1];
1120 sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
1124 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
1125 SingleChannelElement *sce,
1131 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
1132 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
1133 for (g = 0; g < sce->ics.num_swb; g++) {
1134 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
1135 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
1136 if (band->energy <= band->threshold) {
1137 sce->sf_idx[(w+w2)*16+g] = 218;
1138 sce->zeroes[(w+w2)*16+g] = 1;
1140 sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2f(band->threshold), 80, 218);
1141 sce->zeroes[(w+w2)*16+g] = 0;
1143 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
1147 for (i = 0; i < 128; i++) {
1148 sce->sf_idx[i] = 140;
1149 //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
1151 //set the same quantizers inside window groups
1152 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
1153 for (g = 0; g < sce->ics.num_swb; g++)
1154 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
1155 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
1158 static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
1161 int start = 0, i, w, w2, g;
1162 float M[128], S[128];
1163 float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
1164 SingleChannelElement *sce0 = &cpe->ch[0];
1165 SingleChannelElement *sce1 = &cpe->ch[1];
1166 if (!cpe->common_window)
1168 for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
1169 for (g = 0; g < sce0->ics.num_swb; g++) {
1170 if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
1171 float dist1 = 0.0f, dist2 = 0.0f;
1172 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
1173 FFPsyBand *band0 = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g];
1174 FFPsyBand *band1 = &s->psy.ch[s->cur_channel+1].psy_bands[(w+w2)*16+g];
1175 float minthr = FFMIN(band0->threshold, band1->threshold);
1176 float maxthr = FFMAX(band0->threshold, band1->threshold);
1177 for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
1178 M[i] = (sce0->pcoeffs[start+w2*128+i]
1179 + sce1->pcoeffs[start+w2*128+i]) * 0.5;
1181 - sce1->pcoeffs[start+w2*128+i];
1183 abs_pow34_v(L34, sce0->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
1184 abs_pow34_v(R34, sce1->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
1185 abs_pow34_v(M34, M, sce0->ics.swb_sizes[g]);
1186 abs_pow34_v(S34, S, sce0->ics.swb_sizes[g]);
1187 dist1 += quantize_band_cost(s, sce0->coeffs + start + w2*128,
1189 sce0->ics.swb_sizes[g],
1190 sce0->sf_idx[(w+w2)*16+g],
1191 sce0->band_type[(w+w2)*16+g],
1192 lambda / band0->threshold, INFINITY, NULL);
1193 dist1 += quantize_band_cost(s, sce1->coeffs + start + w2*128,
1195 sce1->ics.swb_sizes[g],
1196 sce1->sf_idx[(w+w2)*16+g],
1197 sce1->band_type[(w+w2)*16+g],
1198 lambda / band1->threshold, INFINITY, NULL);
1199 dist2 += quantize_band_cost(s, M,
1201 sce0->ics.swb_sizes[g],
1202 sce0->sf_idx[(w+w2)*16+g],
1203 sce0->band_type[(w+w2)*16+g],
1204 lambda / maxthr, INFINITY, NULL);
1205 dist2 += quantize_band_cost(s, S,
1207 sce1->ics.swb_sizes[g],
1208 sce1->sf_idx[(w+w2)*16+g],
1209 sce1->band_type[(w+w2)*16+g],
1210 lambda / minthr, INFINITY, NULL);
1212 cpe->ms_mask[w*16+g] = dist2 < dist1;
1214 start += sce0->ics.swb_sizes[g];
1219 AACCoefficientsEncoder ff_aac_coders[AAC_CODER_NB] = {
1220 [AAC_CODER_FAAC] = {
1221 search_for_quantizers_faac,
1222 encode_window_bands_info,
1223 quantize_and_encode_band,
1226 [AAC_CODER_ANMR] = {
1227 search_for_quantizers_anmr,
1228 encode_window_bands_info,
1229 quantize_and_encode_band,
1232 [AAC_CODER_TWOLOOP] = {
1233 search_for_quantizers_twoloop,
1234 codebook_trellis_rate,
1235 quantize_and_encode_band,
1238 [AAC_CODER_FAST] = {
1239 search_for_quantizers_fast,
1240 encode_window_bands_info,
1241 quantize_and_encode_band,