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 ***********************************/
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
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
52 static const uint8_t *run_value_bits[2] = {
53 run_value_bits_long, run_value_bits_short
58 * Quantize one coefficient.
59 * @return absolute value of the quantized coefficient
60 * @see 3GPP TS26.403 5.6.2 "Scalefactor determination"
62 static av_always_inline int quant(float coef, const float Q)
65 return sqrtf(a * sqrtf(a)) + 0.4054;
68 static void quantize_bands(int *out, const float *in, const float *scaled,
69 int size, float Q34, int is_signed, int maxval)
73 for (i = 0; i < size; i++) {
75 out[i] = (int)FFMIN(qc + 0.4054, (double)maxval);
76 if (is_signed && in[i] < 0.0f) {
82 static void abs_pow34_v(float *out, const float *in, const int size)
84 #ifndef USE_REALLY_FULL_SEARCH
86 for (i = 0; i < size; i++) {
87 float a = fabsf(in[i]);
88 out[i] = sqrtf(a * sqrtf(a));
90 #endif /* USE_REALLY_FULL_SEARCH */
93 static const uint8_t aac_cb_range [12] = {0, 3, 3, 3, 3, 9, 9, 8, 8, 13, 13, 17};
94 static const uint8_t aac_cb_maxval[12] = {0, 1, 1, 2, 2, 4, 4, 7, 7, 12, 12, 16};
97 * Calculate rate distortion cost for quantizing with given codebook
99 * @return quantization distortion
101 static float quantize_and_encode_band_cost(struct AACEncContext *s,
102 PutBitContext *pb, const float *in,
103 const float *scaled, int size, int scale_idx,
104 int cb, const float lambda, const float uplim,
107 const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
108 const float Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
109 const float CLIPPED_ESCAPE = 165140.0f*IQ;
112 const int dim = cb < FIRST_PAIR_BT ? 4 : 2;
114 const float Q34 = sqrtf(Q * sqrtf(Q));
115 const int range = aac_cb_range[cb];
116 const int maxval = aac_cb_maxval[cb];
120 for (i = 0; i < size; i++)
124 return cost * lambda;
127 abs_pow34_v(s->scoefs, in, size);
130 quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
131 if (IS_CODEBOOK_UNSIGNED(cb)) {
136 for (i = 0; i < size; i += dim) {
138 int *quants = s->qcoefs + i;
142 for (j = 0; j < dim; j++) {
144 curidx += quants[j] + off;
146 curbits = ff_aac_spectral_bits[cb-1][curidx];
147 vec = &ff_aac_codebook_vectors[cb-1][curidx*dim];
148 if (IS_CODEBOOK_UNSIGNED(cb)) {
149 for (k = 0; k < dim; k++) {
150 float t = fabsf(in[i+k]);
152 if (vec[k] == 64.0f) { //FIXME: slow
153 if (t >= CLIPPED_ESCAPE) {
154 di = t - CLIPPED_ESCAPE;
157 int c = av_clip(quant(t, Q), 0, 8191);
158 di = t - c*cbrtf(c)*IQ;
159 curbits += av_log2(c)*2 - 4 + 1;
169 for (k = 0; k < dim; k++) {
170 float di = in[i+k] - vec[k]*IQ;
174 cost += rd * lambda + curbits;
179 put_bits(pb, ff_aac_spectral_bits[cb-1][curidx], ff_aac_spectral_codes[cb-1][curidx]);
180 if (IS_CODEBOOK_UNSIGNED(cb))
181 for (j = 0; j < dim; j++)
182 if (ff_aac_codebook_vectors[cb-1][curidx*dim+j] != 0.0f)
183 put_bits(pb, 1, in[i+j] < 0.0f);
185 for (j = 0; j < 2; j++) {
186 if (ff_aac_codebook_vectors[cb-1][curidx*2+j] == 64.0f) {
187 int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
188 int len = av_log2(coef);
190 put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
191 put_bits(pb, len, coef & ((1 << len) - 1));
202 static float quantize_band_cost(struct AACEncContext *s, const float *in,
203 const float *scaled, int size, int scale_idx,
204 int cb, const float lambda, const float uplim,
207 return quantize_and_encode_band_cost(s, NULL, in, scaled, size, scale_idx,
208 cb, lambda, uplim, bits);
211 static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
212 const float *in, int size, int scale_idx,
213 int cb, const float lambda)
215 quantize_and_encode_band_cost(s, pb, in, NULL, size, scale_idx, cb, lambda,
220 * structure used in optimal codebook search
222 typedef struct BandCodingPath {
223 int prev_idx; ///< pointer to the previous path point
224 float cost; ///< path cost
229 * Encode band info for single window group bands.
231 static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
232 int win, int group_len, const float lambda)
234 BandCodingPath path[120][12];
235 int w, swb, cb, start, start2, size;
237 const int max_sfb = sce->ics.max_sfb;
238 const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
239 const int run_esc = (1 << run_bits) - 1;
240 int idx, ppos, count;
241 int stackrun[120], stackcb[120], stack_len;
242 float next_minrd = INFINITY;
245 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
247 for (cb = 0; cb < 12; cb++) {
248 path[0][cb].cost = 0.0f;
249 path[0][cb].prev_idx = -1;
252 for (swb = 0; swb < max_sfb; swb++) {
254 size = sce->ics.swb_sizes[swb];
255 if (sce->zeroes[win*16 + swb]) {
256 for (cb = 0; cb < 12; cb++) {
257 path[swb+1][cb].prev_idx = cb;
258 path[swb+1][cb].cost = path[swb][cb].cost;
259 path[swb+1][cb].run = path[swb][cb].run + 1;
262 float minrd = next_minrd;
263 int mincb = next_mincb;
264 next_minrd = INFINITY;
266 for (cb = 0; cb < 12; cb++) {
267 float cost_stay_here, cost_get_here;
269 for (w = 0; w < group_len; w++) {
270 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(win+w)*16+swb];
271 rd += quantize_band_cost(s, sce->coeffs + start + w*128,
272 s->scoefs + start + w*128, size,
273 sce->sf_idx[(win+w)*16+swb], cb,
274 lambda / band->threshold, INFINITY, NULL);
276 cost_stay_here = path[swb][cb].cost + rd;
277 cost_get_here = minrd + rd + run_bits + 4;
278 if ( run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
279 != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
280 cost_stay_here += run_bits;
281 if (cost_get_here < cost_stay_here) {
282 path[swb+1][cb].prev_idx = mincb;
283 path[swb+1][cb].cost = cost_get_here;
284 path[swb+1][cb].run = 1;
286 path[swb+1][cb].prev_idx = cb;
287 path[swb+1][cb].cost = cost_stay_here;
288 path[swb+1][cb].run = path[swb][cb].run + 1;
290 if (path[swb+1][cb].cost < next_minrd) {
291 next_minrd = path[swb+1][cb].cost;
296 start += sce->ics.swb_sizes[swb];
299 //convert resulting path from backward-linked list
302 for (cb = 1; cb < 12; cb++)
303 if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
308 stackrun[stack_len] = path[ppos][cb].run;
309 stackcb [stack_len] = cb;
310 idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
311 ppos -= path[ppos][cb].run;
314 //perform actual band info encoding
316 for (i = stack_len - 1; i >= 0; i--) {
317 put_bits(&s->pb, 4, stackcb[i]);
319 memset(sce->zeroes + win*16 + start, !stackcb[i], count);
320 //XXX: memset when band_type is also uint8_t
321 for (j = 0; j < count; j++) {
322 sce->band_type[win*16 + start] = stackcb[i];
325 while (count >= run_esc) {
326 put_bits(&s->pb, run_bits, run_esc);
329 put_bits(&s->pb, run_bits, count);
333 static void codebook_trellis_rate(AACEncContext *s, SingleChannelElement *sce,
334 int win, int group_len, const float lambda)
336 BandCodingPath path[120][12];
337 int w, swb, cb, start, start2, size;
339 const int max_sfb = sce->ics.max_sfb;
340 const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
341 const int run_esc = (1 << run_bits) - 1;
342 int idx, ppos, count;
343 int stackrun[120], stackcb[120], stack_len;
344 float next_minrd = INFINITY;
347 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
349 for (cb = 0; cb < 12; cb++) {
350 path[0][cb].cost = run_bits+4;
351 path[0][cb].prev_idx = -1;
354 for (swb = 0; swb < max_sfb; swb++) {
356 size = sce->ics.swb_sizes[swb];
357 if (sce->zeroes[win*16 + swb]) {
358 for (cb = 0; cb < 12; cb++) {
359 path[swb+1][cb].prev_idx = cb;
360 path[swb+1][cb].cost = path[swb][cb].cost;
361 path[swb+1][cb].run = path[swb][cb].run + 1;
364 float minrd = next_minrd;
365 int mincb = next_mincb;
366 int startcb = sce->band_type[win*16+swb];
367 next_minrd = INFINITY;
369 for (cb = 0; cb < startcb; cb++) {
370 path[swb+1][cb].cost = 61450;
371 path[swb+1][cb].prev_idx = -1;
372 path[swb+1][cb].run = 0;
374 for (cb = startcb; cb < 12; cb++) {
375 float cost_stay_here, cost_get_here;
377 for (w = 0; w < group_len; w++) {
378 rd += quantize_band_cost(s, sce->coeffs + start + w*128,
379 s->scoefs + start + w*128, size,
380 sce->sf_idx[(win+w)*16+swb], cb,
383 cost_stay_here = path[swb][cb].cost + rd;
384 cost_get_here = minrd + rd + run_bits + 4;
385 if ( run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
386 != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
387 cost_stay_here += run_bits;
388 if (cost_get_here < cost_stay_here) {
389 path[swb+1][cb].prev_idx = mincb;
390 path[swb+1][cb].cost = cost_get_here;
391 path[swb+1][cb].run = 1;
393 path[swb+1][cb].prev_idx = cb;
394 path[swb+1][cb].cost = cost_stay_here;
395 path[swb+1][cb].run = path[swb][cb].run + 1;
397 if (path[swb+1][cb].cost < next_minrd) {
398 next_minrd = path[swb+1][cb].cost;
403 start += sce->ics.swb_sizes[swb];
406 //convert resulting path from backward-linked list
409 for (cb = 1; cb < 12; cb++)
410 if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
414 if (idx < 0) abort();
416 stackrun[stack_len] = path[ppos][cb].run;
417 stackcb [stack_len] = cb;
418 idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
419 ppos -= path[ppos][cb].run;
422 //perform actual band info encoding
424 for (i = stack_len - 1; i >= 0; i--) {
425 put_bits(&s->pb, 4, stackcb[i]);
427 memset(sce->zeroes + win*16 + start, !stackcb[i], count);
428 //XXX: memset when band_type is also uint8_t
429 for (j = 0; j < count; j++) {
430 sce->band_type[win*16 + start] = stackcb[i];
433 while (count >= run_esc) {
434 put_bits(&s->pb, run_bits, run_esc);
437 put_bits(&s->pb, run_bits, count);
441 typedef struct TrellisPath {
448 #define TRELLIS_STAGES 121
449 #define TRELLIS_STATES 256
451 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
452 SingleChannelElement *sce,
455 int q, w, w2, g, start = 0;
458 TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
459 int bandaddr[TRELLIS_STAGES];
463 for (i = 0; i < TRELLIS_STATES; i++) {
464 paths[0][i].cost = 0.0f;
465 paths[0][i].prev = -1;
466 paths[0][i].min_val = i;
467 paths[0][i].max_val = i;
469 for (j = 1; j < TRELLIS_STAGES; j++) {
470 for (i = 0; i < TRELLIS_STATES; i++) {
471 paths[j][i].cost = INFINITY;
472 paths[j][i].prev = -2;
473 paths[j][i].min_val = INT_MAX;
474 paths[j][i].max_val = 0;
478 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
479 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
481 for (g = 0; g < sce->ics.num_swb; g++) {
482 const float *coefs = sce->coeffs + start;
486 bandaddr[idx] = w * 16 + g;
489 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
490 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
491 if (band->energy <= band->threshold || band->threshold == 0.0f) {
492 sce->zeroes[(w+w2)*16+g] = 1;
495 sce->zeroes[(w+w2)*16+g] = 0;
497 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
498 float t = fabsf(coefs[w2*128+i]);
500 qmin = FFMIN(qmin, t);
501 qmax = FFMAX(qmax, t);
505 int minscale, maxscale;
506 float minrd = INFINITY;
507 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
508 minscale = av_clip_uint8(log2(qmin)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
509 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
510 maxscale = av_clip_uint8(log2(qmax)*4 + 6 + SCALE_ONE_POS - SCALE_DIV_512);
511 for (q = minscale; q < maxscale; q++) {
512 float dists[12], dist;
513 memset(dists, 0, sizeof(dists));
514 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
515 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
517 for (cb = 0; cb <= ESC_BT; cb++)
518 dists[cb] += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
519 q, cb, lambda / band->threshold, INFINITY, NULL);
522 for (i = 1; i <= ESC_BT; i++)
523 dist = FFMIN(dist, dists[i]);
524 minrd = FFMIN(minrd, dist);
526 for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, TRELLIS_STATES); i++) {
529 if (isinf(paths[idx - 1][i].cost))
531 cost = paths[idx - 1][i].cost + dist
532 + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
533 minv = FFMIN(paths[idx - 1][i].min_val, q);
534 maxv = FFMAX(paths[idx - 1][i].max_val, q);
535 if (cost < paths[idx][q].cost && maxv-minv < SCALE_MAX_DIFF) {
536 paths[idx][q].cost = cost;
537 paths[idx][q].prev = i;
538 paths[idx][q].min_val = minv;
539 paths[idx][q].max_val = maxv;
544 for (q = 0; q < TRELLIS_STATES; q++) {
545 if (!isinf(paths[idx - 1][q].cost)) {
546 paths[idx][q].cost = paths[idx - 1][q].cost + 1;
547 paths[idx][q].prev = q;
548 paths[idx][q].min_val = FFMIN(paths[idx - 1][q].min_val, q);
549 paths[idx][q].max_val = FFMAX(paths[idx - 1][q].max_val, q);
552 for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, TRELLIS_STATES); i++) {
555 if (isinf(paths[idx - 1][i].cost))
557 cost = paths[idx - 1][i].cost + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
558 minv = FFMIN(paths[idx - 1][i].min_val, q);
559 maxv = FFMAX(paths[idx - 1][i].max_val, q);
560 if (cost < paths[idx][q].cost && maxv-minv < SCALE_MAX_DIFF) {
561 paths[idx][q].cost = cost;
562 paths[idx][q].prev = i;
563 paths[idx][q].min_val = minv;
564 paths[idx][q].max_val = maxv;
569 sce->zeroes[w*16+g] = !nz;
570 start += sce->ics.swb_sizes[g];
575 mincost = paths[idx][0].cost;
577 for (i = 1; i < TRELLIS_STATES; i++) {
578 if (paths[idx][i].cost < mincost) {
579 mincost = paths[idx][i].cost;
584 sce->sf_idx[bandaddr[idx]] = minq;
585 minq = paths[idx][minq].prev;
588 //set the same quantizers inside window groups
589 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
590 for (g = 0; g < sce->ics.num_swb; g++)
591 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
592 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
596 * two-loop quantizers search taken from ISO 13818-7 Appendix C
598 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
600 SingleChannelElement *sce,
603 int start = 0, i, w, w2, g;
604 int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels;
605 float dists[128], uplims[128];
606 int fflag, minscaler;
609 float minthr = INFINITY;
611 //XXX: some heuristic to determine initial quantizers will reduce search time
612 memset(dists, 0, sizeof(dists));
613 //determine zero bands and upper limits
614 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
615 for (g = 0; g < sce->ics.num_swb; g++) {
618 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
619 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
620 uplim += band->threshold;
621 if (band->energy <= band->threshold || band->threshold == 0.0f) {
622 sce->zeroes[(w+w2)*16+g] = 1;
627 uplims[w*16+g] = uplim *512;
628 sce->zeroes[w*16+g] = !nz;
630 minthr = FFMIN(minthr, uplim);
631 allz = FFMAX(allz, nz);
634 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
635 for (g = 0; g < sce->ics.num_swb; g++) {
636 if (sce->zeroes[w*16+g]) {
637 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
640 sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2(uplims[w*16+g]/minthr)*4,59);
646 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
647 //perform two-loop search
648 //outer loop - improve quality
651 minscaler = sce->sf_idx[0];
652 //inner loop - quantize spectrum to fit into given number of bits
653 qstep = its ? 1 : 32;
658 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
660 for (g = 0; g < sce->ics.num_swb; g++) {
661 const float *coefs = sce->coeffs + start;
662 const float *scaled = s->scoefs + start;
665 float mindist = INFINITY;
668 if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
669 start += sce->ics.swb_sizes[g];
672 minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
677 float Q = ff_aac_pow2sf_tab[200 - sce->sf_idx[w*16+g] + SCALE_ONE_POS - SCALE_DIV_512];
678 float Q34 = sqrtf(Q * sqrtf(Q));
680 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
681 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
682 maxval = FFMAX(maxval, scaled[w2*128+i]);
685 qmaxval = maxval * Q34 + 0.4054;
686 if (qmaxval == 0) cb = 0;
687 else if (qmaxval == 1) cb = 1;
688 else if (qmaxval == 2) cb = 3;
689 else if (qmaxval <= 4) cb = 5;
690 else if (qmaxval <= 7) cb = 7;
691 else if (qmaxval <= 12) cb = 9;
693 sce->band_type[w*16+g] = cb;
694 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
696 dist += quantize_band_cost(s, coefs + w2*128,
698 sce->ics.swb_sizes[g],
709 dists[w*16+g] = (mindist - minbits) / lambda;
712 bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
715 start += sce->ics.swb_sizes[g];
716 prev = sce->sf_idx[w*16+g];
719 if (tbits > destbits) {
720 for (i = 0; i < 128; i++)
721 if (sce->sf_idx[i] < 218 - qstep)
722 sce->sf_idx[i] += qstep;
724 for (i = 0; i < 128; i++)
725 if (sce->sf_idx[i] > 60 - qstep)
726 sce->sf_idx[i] -= qstep;
729 if (!qstep && tbits > destbits*1.02)
731 if (sce->sf_idx[0] >= 217)
736 minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
737 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
739 for (g = 0; g < sce->ics.num_swb; g++) {
740 int prevsc = sce->sf_idx[w*16+g];
741 if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60)
742 sce->sf_idx[w*16+g]--;
743 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
744 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
745 if (sce->sf_idx[w*16+g] != prevsc)
750 } while (fflag && its < 10);
753 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
754 SingleChannelElement *sce,
757 int start = 0, i, w, w2, g;
758 float uplim[128], maxq[128];
760 float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
761 int last = 0, lastband = 0, curband = 0;
762 float avg_energy = 0.0;
763 if (sce->ics.num_windows == 1) {
765 for (i = 0; i < 1024; i++) {
766 if (i - start >= sce->ics.swb_sizes[curband]) {
767 start += sce->ics.swb_sizes[curband];
770 if (sce->coeffs[i]) {
771 avg_energy += sce->coeffs[i] * sce->coeffs[i];
777 for (w = 0; w < 8; w++) {
778 const float *coeffs = sce->coeffs + w*128;
780 for (i = 0; i < 128; i++) {
781 if (i - start >= sce->ics.swb_sizes[curband]) {
782 start += sce->ics.swb_sizes[curband];
786 avg_energy += coeffs[i] * coeffs[i];
787 last = FFMAX(last, i);
788 lastband = FFMAX(lastband, curband);
795 if (avg_energy == 0.0f) {
796 for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
797 sce->sf_idx[i] = SCALE_ONE_POS;
800 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
802 for (g = 0; g < sce->ics.num_swb; g++) {
803 float *coefs = sce->coeffs + start;
804 const int size = sce->ics.swb_sizes[g];
805 int start2 = start, end2 = start + size, peakpos = start;
806 float maxval = -1, thr = 0.0f, t;
811 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
812 memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
815 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
816 for (i = 0; i < size; i++) {
817 float t = coefs[w2*128+i]*coefs[w2*128+i];
818 maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
820 if (sce->ics.num_windows == 1 && maxval < t) {
826 if (sce->ics.num_windows == 1) {
827 start2 = FFMAX(peakpos - 2, start2);
828 end2 = FFMIN(peakpos + 3, end2);
834 thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
835 t = 1.0 - (1.0 * start2 / last);
836 uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
839 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
840 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
841 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
843 for (g = 0; g < sce->ics.num_swb; g++) {
844 const float *coefs = sce->coeffs + start;
845 const float *scaled = s->scoefs + start;
846 const int size = sce->ics.swb_sizes[g];
847 int scf, prev_scf, step;
848 int min_scf = -1, max_scf = 256;
850 if (maxq[w*16+g] < 21.544) {
851 sce->zeroes[w*16+g] = 1;
855 sce->zeroes[w*16+g] = 0;
856 scf = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2(1/maxq[w*16+g])*16/3, 60, 218);
862 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
864 dist += quantize_band_cost(s, coefs + w2*128,
866 sce->ics.swb_sizes[g],
874 dist *= 1.0f / 512.0f / lambda;
875 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[200 - scf + SCALE_ONE_POS - SCALE_DIV_512]);
876 if (quant_max >= 8191) { // too much, return to the previous quantizer
877 sce->sf_idx[w*16+g] = prev_scf;
881 curdiff = fabsf(dist - uplim[w*16+g]);
885 step = log2(curdiff);
886 if (dist > uplim[w*16+g])
889 scf = av_clip_uint8(scf);
890 step = scf - prev_scf;
891 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
892 sce->sf_idx[w*16+g] = av_clip(scf, min_scf, max_scf);
903 minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
904 for (i = 1; i < 128; i++) {
906 sce->sf_idx[i] = sce->sf_idx[i-1];
908 minq = FFMIN(minq, sce->sf_idx[i]);
912 minq = FFMIN(minq, SCALE_MAX_POS);
913 maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
914 for (i = 126; i >= 0; i--) {
916 sce->sf_idx[i] = sce->sf_idx[i+1];
917 sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
921 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
922 SingleChannelElement *sce,
925 int start = 0, i, w, w2, g;
928 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
929 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 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
933 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
934 if (band->energy <= band->threshold) {
935 sce->sf_idx[(w+w2)*16+g] = 218;
936 sce->zeroes[(w+w2)*16+g] = 1;
938 sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2(band->threshold), 80, 218);
939 sce->zeroes[(w+w2)*16+g] = 0;
941 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
945 for (i = 0; i < 128; i++) {
946 sce->sf_idx[i] = 140;
947 //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
949 //set the same quantizers inside window groups
950 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
951 for (g = 0; g < sce->ics.num_swb; g++)
952 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
953 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
956 static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
959 int start = 0, i, w, w2, g;
960 float M[128], S[128];
961 float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
962 SingleChannelElement *sce0 = &cpe->ch[0];
963 SingleChannelElement *sce1 = &cpe->ch[1];
964 if (!cpe->common_window)
966 for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
967 for (g = 0; g < sce0->ics.num_swb; g++) {
968 if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
969 float dist1 = 0.0f, dist2 = 0.0f;
970 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
971 FFPsyBand *band0 = &s->psy.psy_bands[(s->cur_channel+0)*PSY_MAX_BANDS+(w+w2)*16+g];
972 FFPsyBand *band1 = &s->psy.psy_bands[(s->cur_channel+1)*PSY_MAX_BANDS+(w+w2)*16+g];
973 float minthr = FFMIN(band0->threshold, band1->threshold);
974 float maxthr = FFMAX(band0->threshold, band1->threshold);
975 for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
976 M[i] = (sce0->coeffs[start+w2*128+i]
977 + sce1->coeffs[start+w2*128+i]) * 0.5;
978 S[i] = sce0->coeffs[start+w2*128+i]
979 - sce1->coeffs[start+w2*128+i];
981 abs_pow34_v(L34, sce0->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
982 abs_pow34_v(R34, sce1->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
983 abs_pow34_v(M34, M, sce0->ics.swb_sizes[g]);
984 abs_pow34_v(S34, S, sce0->ics.swb_sizes[g]);
985 dist1 += quantize_band_cost(s, sce0->coeffs + start + w2*128,
987 sce0->ics.swb_sizes[g],
988 sce0->sf_idx[(w+w2)*16+g],
989 sce0->band_type[(w+w2)*16+g],
990 lambda / band0->threshold, INFINITY, NULL);
991 dist1 += quantize_band_cost(s, sce1->coeffs + start + w2*128,
993 sce1->ics.swb_sizes[g],
994 sce1->sf_idx[(w+w2)*16+g],
995 sce1->band_type[(w+w2)*16+g],
996 lambda / band1->threshold, INFINITY, NULL);
997 dist2 += quantize_band_cost(s, M,
999 sce0->ics.swb_sizes[g],
1000 sce0->sf_idx[(w+w2)*16+g],
1001 sce0->band_type[(w+w2)*16+g],
1002 lambda / maxthr, INFINITY, NULL);
1003 dist2 += quantize_band_cost(s, S,
1005 sce1->ics.swb_sizes[g],
1006 sce1->sf_idx[(w+w2)*16+g],
1007 sce1->band_type[(w+w2)*16+g],
1008 lambda / minthr, INFINITY, NULL);
1010 cpe->ms_mask[w*16+g] = dist2 < dist1;
1012 start += sce0->ics.swb_sizes[g];
1017 AACCoefficientsEncoder ff_aac_coders[] = {
1019 search_for_quantizers_faac,
1020 encode_window_bands_info,
1021 quantize_and_encode_band,
1025 search_for_quantizers_anmr,
1026 encode_window_bands_info,
1027 quantize_and_encode_band,
1031 search_for_quantizers_twoloop,
1032 codebook_trellis_rate,
1033 quantize_and_encode_band,
1037 search_for_quantizers_fast,
1038 encode_window_bands_info,
1039 quantize_and_encode_band,