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 /** 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
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
56 static const uint8_t *run_value_bits[2] = {
57 run_value_bits_long, run_value_bits_short
62 * Quantize one coefficient.
63 * @return absolute value of the quantized coefficient
64 * @see 3GPP TS26.403 5.6.2 "Scalefactor determination"
66 static av_always_inline int quant(float coef, const float Q)
69 return sqrtf(a * sqrtf(a)) + 0.4054;
72 static void quantize_bands(int *out, const float *in, const float *scaled,
73 int size, float Q34, int is_signed, int maxval)
77 for (i = 0; i < size; i++) {
79 out[i] = (int)FFMIN(qc + 0.4054, (double)maxval);
80 if (is_signed && in[i] < 0.0f) {
86 static void abs_pow34_v(float *out, const float *in, const int size)
88 #ifndef USE_REALLY_FULL_SEARCH
90 for (i = 0; i < size; i++) {
91 float a = fabsf(in[i]);
92 out[i] = sqrtf(a * sqrtf(a));
94 #endif /* USE_REALLY_FULL_SEARCH */
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};
101 * Calculate rate distortion cost for quantizing with given codebook
103 * @return quantization distortion
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)
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;
120 const int dim = BT_PAIR ? 2 : 4;
122 const int range = aac_cb_range[cb];
123 const int maxval = aac_cb_maxval[cb];
127 for (i = 0; i < size; i++)
131 return cost * lambda;
134 abs_pow34_v(s->scoefs, in, size);
137 quantize_bands(s->qcoefs, in, scaled, size, Q34, !BT_UNSIGNED, maxval);
143 for (i = 0; i < size; i += dim) {
145 int *quants = s->qcoefs + i;
149 for (j = 0; j < dim; j++) {
151 curidx += quants[j] + off;
153 curbits = ff_aac_spectral_bits[cb-1][curidx];
154 vec = &ff_aac_codebook_vectors[cb-1][curidx*dim];
156 for (j = 0; j < dim; j++) {
157 float t = fabsf(in[i+j]);
159 if (BT_ESC && vec[j] == 64.0f) { //FIXME: slow
160 if (t >= CLIPPED_ESCAPE) {
161 di = t - CLIPPED_ESCAPE;
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;
176 for (j = 0; j < dim; j++) {
177 float di = in[i+j] - vec[j]*IQ;
181 cost += rd * lambda + curbits;
186 put_bits(pb, ff_aac_spectral_bits[cb-1][curidx], ff_aac_spectral_codes[cb-1][curidx]);
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);
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);
197 put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
198 put_bits(pb, len, coef & ((1 << len) - 1));
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, \
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); \
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)
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,
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,
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, \
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,
262 return quantize_and_encode_band_cost(s, NULL, in, scaled, size, scale_idx,
263 cb, lambda, uplim, bits);
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)
270 quantize_and_encode_band_cost(s, pb, in, NULL, size, scale_idx, cb, lambda,
274 static float find_max_val(int group_len, int swb_size, const float *scaled) {
277 for (w2 = 0; w2 < group_len; w2++) {
278 for (i = 0; i < swb_size; i++) {
279 maxval = FFMAX(maxval, scaled[w2*128+i]);
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));
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;
301 * structure used in optimal codebook search
303 typedef struct BandCodingPath {
304 int prev_idx; ///< pointer to the previous path point
305 float cost; ///< path cost
310 * Encode band info for single window group bands.
312 static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
313 int win, int group_len, const float lambda)
315 BandCodingPath path[120][12];
316 int w, swb, cb, start, size;
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;
326 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
328 for (cb = 0; cb < 12; cb++) {
329 path[0][cb].cost = 0.0f;
330 path[0][cb].prev_idx = -1;
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;
342 float minrd = next_minrd;
343 int mincb = next_mincb;
344 next_minrd = INFINITY;
346 for (cb = 0; cb < 12; cb++) {
347 float cost_stay_here, cost_get_here;
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);
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;
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;
370 if (path[swb+1][cb].cost < next_minrd) {
371 next_minrd = path[swb+1][cb].cost;
376 start += sce->ics.swb_sizes[swb];
379 //convert resulting path from backward-linked list
382 for (cb = 1; cb < 12; cb++)
383 if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
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;
394 //perform actual band info encoding
396 for (i = stack_len - 1; i >= 0; i--) {
397 put_bits(&s->pb, 4, stackcb[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];
405 while (count >= run_esc) {
406 put_bits(&s->pb, run_bits, run_esc);
409 put_bits(&s->pb, run_bits, count);
413 static void codebook_trellis_rate(AACEncContext *s, SingleChannelElement *sce,
414 int win, int group_len, const float lambda)
416 BandCodingPath path[120][12];
417 int w, swb, cb, start, size;
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;
427 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
429 for (cb = 0; cb < 12; cb++) {
430 path[0][cb].cost = run_bits+4;
431 path[0][cb].prev_idx = -1;
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;
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;
451 next_minbits = path[swb+1][0].cost;
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;
459 float minbits = next_minbits;
460 int mincb = next_mincb;
461 int startcb = sce->band_type[win*16+swb];
462 next_minbits = INFINITY;
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;
469 for (cb = startcb; cb < 12; cb++) {
470 float cost_stay_here, cost_get_here;
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,
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;
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;
492 if (path[swb+1][cb].cost < next_minbits) {
493 next_minbits = path[swb+1][cb].cost;
498 start += sce->ics.swb_sizes[swb];
501 //convert resulting path from backward-linked list
504 for (cb = 1; cb < 12; cb++)
505 if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
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;
517 //perform actual band info encoding
519 for (i = stack_len - 1; i >= 0; i--) {
520 put_bits(&s->pb, 4, stackcb[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];
528 while (count >= run_esc) {
529 put_bits(&s->pb, run_bits, run_esc);
532 put_bits(&s->pb, run_bits, count);
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);
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);
546 typedef struct TrellisPath {
551 #define TRELLIS_STAGES 121
552 #define TRELLIS_STATES (SCALE_MAX_DIFF+1)
554 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
555 SingleChannelElement *sce,
558 int q, w, w2, g, start = 0;
561 TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
562 int bandaddr[TRELLIS_STAGES];
565 float q0f = FLT_MAX, q1f = 0.0f, qnrgf = 0.0f;
566 int q0, q1, qcnt = 0;
568 for (i = 0; i < 1024; i++) {
569 float t = fabsf(sce->coeffs[i]);
579 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
580 memset(sce->zeroes, 1, sizeof(sce->zeroes));
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 //av_log(NULL, AV_LOG_ERROR, "q0 %d, q1 %d\n", q0, q1);
592 //minimum scalefactor index is when maximum nonzero coefficient after quantizing is not clipped
593 int qnrg = av_clip_uint8(log2f(sqrtf(qnrgf/qcnt))*4 - 31 + SCALE_ONE_POS - SCALE_DIV_512);
596 //av_log(NULL, AV_LOG_ERROR, "q0 %d, q1 %d\n", q0, q1);
600 } else if (q1 > q1high) {
605 //av_log(NULL, AV_LOG_ERROR, "q0 %d, q1 %d\n", q0, q1);
607 for (i = 0; i < TRELLIS_STATES; i++) {
608 paths[0][i].cost = 0.0f;
609 paths[0][i].prev = -1;
611 for (j = 1; j < TRELLIS_STAGES; j++) {
612 for (i = 0; i < TRELLIS_STATES; i++) {
613 paths[j][i].cost = INFINITY;
614 paths[j][i].prev = -2;
618 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
619 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
621 for (g = 0; g < sce->ics.num_swb; g++) {
622 const float *coefs = sce->coeffs + start;
626 bandaddr[idx] = w * 16 + g;
629 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
630 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
631 if (band->energy <= band->threshold || band->threshold == 0.0f) {
632 sce->zeroes[(w+w2)*16+g] = 1;
635 sce->zeroes[(w+w2)*16+g] = 0;
637 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
638 float t = fabsf(coefs[w2*128+i]);
640 qmin = FFMIN(qmin, t);
641 qmax = FFMAX(qmax, t);
645 int minscale, maxscale;
646 float minrd = INFINITY;
648 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
649 minscale = coef2minsf(qmin);
650 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
651 maxscale = coef2maxsf(qmax);
652 minscale = av_clip(minscale - q0, 0, TRELLIS_STATES - 1);
653 maxscale = av_clip(maxscale - q0, 0, TRELLIS_STATES);
654 maxval = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], s->scoefs+start);
655 for (q = minscale; q < maxscale; q++) {
657 int cb = find_min_book(maxval, sce->sf_idx[w*16+g]);
658 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
659 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
660 dist += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
661 q + q0, cb, lambda / band->threshold, INFINITY, NULL);
663 minrd = FFMIN(minrd, dist);
665 for (i = 0; i < q1 - q0; i++) {
667 cost = paths[idx - 1][i].cost + dist
668 + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
669 if (cost < paths[idx][q].cost) {
670 paths[idx][q].cost = cost;
671 paths[idx][q].prev = i;
676 for (q = 0; q < q1 - q0; q++) {
677 paths[idx][q].cost = paths[idx - 1][q].cost + 1;
678 paths[idx][q].prev = q;
681 sce->zeroes[w*16+g] = !nz;
682 start += sce->ics.swb_sizes[g];
687 mincost = paths[idx][0].cost;
689 for (i = 1; i < TRELLIS_STATES; i++) {
690 if (paths[idx][i].cost < mincost) {
691 mincost = paths[idx][i].cost;
696 sce->sf_idx[bandaddr[idx]] = minq + q0;
697 minq = paths[idx][minq].prev;
700 //set the same quantizers inside window groups
701 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
702 for (g = 0; g < sce->ics.num_swb; g++)
703 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
704 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
708 * two-loop quantizers search taken from ISO 13818-7 Appendix C
710 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
712 SingleChannelElement *sce,
715 int start = 0, i, w, w2, g;
716 int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels;
717 float dists[128] = { 0 }, uplims[128];
719 int fflag, minscaler;
722 float minthr = INFINITY;
724 // for values above this the decoder might end up in an endless loop
725 // due to always having more bits than what can be encoded.
726 destbits = FFMIN(destbits, 5800);
727 //XXX: some heuristic to determine initial quantizers will reduce search time
728 //determine zero bands and upper limits
729 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
730 for (g = 0; g < sce->ics.num_swb; g++) {
733 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
734 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
735 uplim += band->threshold;
736 if (band->energy <= band->threshold || band->threshold == 0.0f) {
737 sce->zeroes[(w+w2)*16+g] = 1;
742 uplims[w*16+g] = uplim *512;
743 sce->zeroes[w*16+g] = !nz;
745 minthr = FFMIN(minthr, uplim);
749 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
750 for (g = 0; g < sce->ics.num_swb; g++) {
751 if (sce->zeroes[w*16+g]) {
752 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
755 sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2f(uplims[w*16+g]/minthr)*4,59);
761 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
763 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
765 for (g = 0; g < sce->ics.num_swb; g++) {
766 const float *scaled = s->scoefs + start;
767 maxvals[w*16+g] = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled);
768 start += sce->ics.swb_sizes[g];
772 //perform two-loop search
773 //outer loop - improve quality
776 minscaler = sce->sf_idx[0];
777 //inner loop - quantize spectrum to fit into given number of bits
778 qstep = its ? 1 : 32;
783 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
785 for (g = 0; g < sce->ics.num_swb; g++) {
786 const float *coefs = sce->coeffs + start;
787 const float *scaled = s->scoefs + start;
792 if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
793 start += sce->ics.swb_sizes[g];
796 minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
797 cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
798 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
800 dist += quantize_band_cost(s, coefs + w2*128,
802 sce->ics.swb_sizes[g],
810 dists[w*16+g] = dist - bits;
812 bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
815 start += sce->ics.swb_sizes[g];
816 prev = sce->sf_idx[w*16+g];
819 if (tbits > destbits) {
820 for (i = 0; i < 128; i++)
821 if (sce->sf_idx[i] < 218 - qstep)
822 sce->sf_idx[i] += qstep;
824 for (i = 0; i < 128; i++)
825 if (sce->sf_idx[i] > 60 - qstep)
826 sce->sf_idx[i] -= qstep;
829 if (!qstep && tbits > destbits*1.02 && sce->sf_idx[0] < 217)
834 minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
835 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
836 for (g = 0; g < sce->ics.num_swb; g++) {
837 int prevsc = sce->sf_idx[w*16+g];
838 if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60) {
839 if (find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]-1))
840 sce->sf_idx[w*16+g]--;
841 else //Try to make sure there is some energy in every band
842 sce->sf_idx[w*16+g]-=2;
844 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
845 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
846 if (sce->sf_idx[w*16+g] != prevsc)
848 sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
852 } while (fflag && its < 10);
855 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
856 SingleChannelElement *sce,
859 int start = 0, i, w, w2, g;
860 float uplim[128], maxq[128];
862 float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
863 int last = 0, lastband = 0, curband = 0;
864 float avg_energy = 0.0;
865 if (sce->ics.num_windows == 1) {
867 for (i = 0; i < 1024; i++) {
868 if (i - start >= sce->ics.swb_sizes[curband]) {
869 start += sce->ics.swb_sizes[curband];
872 if (sce->coeffs[i]) {
873 avg_energy += sce->coeffs[i] * sce->coeffs[i];
879 for (w = 0; w < 8; w++) {
880 const float *coeffs = sce->coeffs + w*128;
882 for (i = 0; i < 128; i++) {
883 if (i - start >= sce->ics.swb_sizes[curband]) {
884 start += sce->ics.swb_sizes[curband];
888 avg_energy += coeffs[i] * coeffs[i];
889 last = FFMAX(last, i);
890 lastband = FFMAX(lastband, curband);
897 if (avg_energy == 0.0f) {
898 for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
899 sce->sf_idx[i] = SCALE_ONE_POS;
902 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
904 for (g = 0; g < sce->ics.num_swb; g++) {
905 float *coefs = sce->coeffs + start;
906 const int size = sce->ics.swb_sizes[g];
907 int start2 = start, end2 = start + size, peakpos = start;
908 float maxval = -1, thr = 0.0f, t;
913 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
914 memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
917 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
918 for (i = 0; i < size; i++) {
919 float t = coefs[w2*128+i]*coefs[w2*128+i];
920 maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
922 if (sce->ics.num_windows == 1 && maxval < t) {
928 if (sce->ics.num_windows == 1) {
929 start2 = FFMAX(peakpos - 2, start2);
930 end2 = FFMIN(peakpos + 3, end2);
936 thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
937 t = 1.0 - (1.0 * start2 / last);
938 uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
941 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
942 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
943 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
945 for (g = 0; g < sce->ics.num_swb; g++) {
946 const float *coefs = sce->coeffs + start;
947 const float *scaled = s->scoefs + start;
948 const int size = sce->ics.swb_sizes[g];
949 int scf, prev_scf, step;
950 int min_scf = -1, max_scf = 256;
952 if (maxq[w*16+g] < 21.544) {
953 sce->zeroes[w*16+g] = 1;
957 sce->zeroes[w*16+g] = 0;
958 scf = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2f(1/maxq[w*16+g])*16/3, 60, 218);
964 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
966 dist += quantize_band_cost(s, coefs + w2*128,
968 sce->ics.swb_sizes[g],
976 dist *= 1.0f / 512.0f / lambda;
977 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[POW_SF2_ZERO - scf + SCALE_ONE_POS - SCALE_DIV_512]);
978 if (quant_max >= 8191) { // too much, return to the previous quantizer
979 sce->sf_idx[w*16+g] = prev_scf;
983 curdiff = fabsf(dist - uplim[w*16+g]);
987 step = log2f(curdiff);
988 if (dist > uplim[w*16+g])
991 scf = av_clip_uint8(scf);
992 step = scf - prev_scf;
993 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
994 sce->sf_idx[w*16+g] = av_clip(scf, min_scf, max_scf);
1005 minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
1006 for (i = 1; i < 128; i++) {
1007 if (!sce->sf_idx[i])
1008 sce->sf_idx[i] = sce->sf_idx[i-1];
1010 minq = FFMIN(minq, sce->sf_idx[i]);
1012 if (minq == INT_MAX)
1014 minq = FFMIN(minq, SCALE_MAX_POS);
1015 maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
1016 for (i = 126; i >= 0; i--) {
1017 if (!sce->sf_idx[i])
1018 sce->sf_idx[i] = sce->sf_idx[i+1];
1019 sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
1023 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
1024 SingleChannelElement *sce,
1030 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
1031 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
1032 for (g = 0; g < sce->ics.num_swb; g++) {
1033 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
1034 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
1035 if (band->energy <= band->threshold) {
1036 sce->sf_idx[(w+w2)*16+g] = 218;
1037 sce->zeroes[(w+w2)*16+g] = 1;
1039 sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2f(band->threshold), 80, 218);
1040 sce->zeroes[(w+w2)*16+g] = 0;
1042 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
1046 for (i = 0; i < 128; i++) {
1047 sce->sf_idx[i] = 140;
1048 //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
1050 //set the same quantizers inside window groups
1051 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
1052 for (g = 0; g < sce->ics.num_swb; g++)
1053 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
1054 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
1057 static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
1060 int start = 0, i, w, w2, g;
1061 float M[128], S[128];
1062 float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
1063 SingleChannelElement *sce0 = &cpe->ch[0];
1064 SingleChannelElement *sce1 = &cpe->ch[1];
1065 if (!cpe->common_window)
1067 for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
1068 for (g = 0; g < sce0->ics.num_swb; g++) {
1069 if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
1070 float dist1 = 0.0f, dist2 = 0.0f;
1071 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
1072 FFPsyBand *band0 = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g];
1073 FFPsyBand *band1 = &s->psy.ch[s->cur_channel+1].psy_bands[(w+w2)*16+g];
1074 float minthr = FFMIN(band0->threshold, band1->threshold);
1075 float maxthr = FFMAX(band0->threshold, band1->threshold);
1076 for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
1077 M[i] = (sce0->coeffs[start+w2*128+i]
1078 + sce1->coeffs[start+w2*128+i]) * 0.5;
1080 - sce1->coeffs[start+w2*128+i];
1082 abs_pow34_v(L34, sce0->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
1083 abs_pow34_v(R34, sce1->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
1084 abs_pow34_v(M34, M, sce0->ics.swb_sizes[g]);
1085 abs_pow34_v(S34, S, sce0->ics.swb_sizes[g]);
1086 dist1 += quantize_band_cost(s, sce0->coeffs + start + w2*128,
1088 sce0->ics.swb_sizes[g],
1089 sce0->sf_idx[(w+w2)*16+g],
1090 sce0->band_type[(w+w2)*16+g],
1091 lambda / band0->threshold, INFINITY, NULL);
1092 dist1 += quantize_band_cost(s, sce1->coeffs + start + w2*128,
1094 sce1->ics.swb_sizes[g],
1095 sce1->sf_idx[(w+w2)*16+g],
1096 sce1->band_type[(w+w2)*16+g],
1097 lambda / band1->threshold, INFINITY, NULL);
1098 dist2 += quantize_band_cost(s, M,
1100 sce0->ics.swb_sizes[g],
1101 sce0->sf_idx[(w+w2)*16+g],
1102 sce0->band_type[(w+w2)*16+g],
1103 lambda / maxthr, INFINITY, NULL);
1104 dist2 += quantize_band_cost(s, S,
1106 sce1->ics.swb_sizes[g],
1107 sce1->sf_idx[(w+w2)*16+g],
1108 sce1->band_type[(w+w2)*16+g],
1109 lambda / minthr, INFINITY, NULL);
1111 cpe->ms_mask[w*16+g] = dist2 < dist1;
1113 start += sce0->ics.swb_sizes[g];
1118 AACCoefficientsEncoder ff_aac_coders[AAC_CODER_NB] = {
1120 search_for_quantizers_faac,
1121 encode_window_bands_info,
1122 quantize_and_encode_band,
1126 search_for_quantizers_anmr,
1127 encode_window_bands_info,
1128 quantize_and_encode_band,
1132 search_for_quantizers_twoloop,
1133 codebook_trellis_rate,
1134 quantize_and_encode_band,
1138 search_for_quantizers_fast,
1139 encode_window_bands_info,
1140 quantize_and_encode_band,