2 * AAC coefficients encoder
3 * Copyright (C) 2008-2009 Konstantin Shishkov
5 * This file is part of Libav.
7 * Libav 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 * Libav 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 Libav; 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
42 /** bits needed to code codebook run value for long windows */
43 static const uint8_t run_value_bits_long[64] = {
44 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
45 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 10,
46 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
47 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 15
50 /** bits needed to code codebook run value for short windows */
51 static const uint8_t run_value_bits_short[16] = {
52 3, 3, 3, 3, 3, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 9
55 static const uint8_t *run_value_bits[2] = {
56 run_value_bits_long, run_value_bits_short
61 * Quantize one coefficient.
62 * @return absolute value of the quantized coefficient
63 * @see 3GPP TS26.403 5.6.2 "Scalefactor determination"
65 static av_always_inline int quant(float coef, const float Q)
68 return sqrtf(a * sqrtf(a)) + 0.4054;
71 static void quantize_bands(int *out, const float *in, const float *scaled,
72 int size, float Q34, int is_signed, int maxval)
76 for (i = 0; i < size; i++) {
78 out[i] = (int)FFMIN(qc + 0.4054, (double)maxval);
79 if (is_signed && in[i] < 0.0f) {
85 static void abs_pow34_v(float *out, const float *in, const int size)
87 #ifndef USE_REALLY_FULL_SEARCH
89 for (i = 0; i < size; i++) {
90 float a = fabsf(in[i]);
91 out[i] = sqrtf(a * sqrtf(a));
93 #endif /* USE_REALLY_FULL_SEARCH */
96 static const uint8_t aac_cb_range [12] = {0, 3, 3, 3, 3, 9, 9, 8, 8, 13, 13, 17};
97 static const uint8_t aac_cb_maxval[12] = {0, 1, 1, 2, 2, 4, 4, 7, 7, 12, 12, 16};
100 * Calculate rate distortion cost for quantizing with given codebook
102 * @return quantization distortion
104 static av_always_inline float quantize_and_encode_band_cost_template(
105 struct AACEncContext *s,
106 PutBitContext *pb, const float *in,
107 const float *scaled, int size, int scale_idx,
108 int cb, const float lambda, const float uplim,
109 int *bits, int BT_ZERO, int BT_UNSIGNED,
110 int BT_PAIR, int BT_ESC)
112 const float IQ = ff_aac_pow2sf_tab[POW_SF2_ZERO + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
113 const float Q = ff_aac_pow2sf_tab[POW_SF2_ZERO - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
114 const float CLIPPED_ESCAPE = 165140.0f*IQ;
117 const int dim = BT_PAIR ? 2 : 4;
119 const float Q34 = sqrtf(Q * sqrtf(Q));
120 const int range = aac_cb_range[cb];
121 const int maxval = aac_cb_maxval[cb];
125 for (i = 0; i < size; i++)
129 return cost * lambda;
132 abs_pow34_v(s->scoefs, in, size);
135 quantize_bands(s->qcoefs, in, scaled, size, Q34, !BT_UNSIGNED, maxval);
141 for (i = 0; i < size; i += dim) {
143 int *quants = s->qcoefs + i;
147 for (j = 0; j < dim; j++) {
149 curidx += quants[j] + off;
151 curbits = ff_aac_spectral_bits[cb-1][curidx];
152 vec = &ff_aac_codebook_vectors[cb-1][curidx*dim];
154 for (j = 0; j < dim; j++) {
155 float t = fabsf(in[i+j]);
157 if (BT_ESC && vec[j] == 64.0f) { //FIXME: slow
158 if (t >= CLIPPED_ESCAPE) {
159 di = t - CLIPPED_ESCAPE;
162 int c = av_clip(quant(t, Q), 0, 8191);
163 di = t - c*cbrtf(c)*IQ;
164 curbits += av_log2(c)*2 - 4 + 1;
174 for (j = 0; j < dim; j++) {
175 float di = in[i+j] - vec[j]*IQ;
179 cost += rd * lambda + curbits;
184 put_bits(pb, ff_aac_spectral_bits[cb-1][curidx], ff_aac_spectral_codes[cb-1][curidx]);
186 for (j = 0; j < dim; j++)
187 if (ff_aac_codebook_vectors[cb-1][curidx*dim+j] != 0.0f)
188 put_bits(pb, 1, in[i+j] < 0.0f);
190 for (j = 0; j < 2; j++) {
191 if (ff_aac_codebook_vectors[cb-1][curidx*2+j] == 64.0f) {
192 int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
193 int len = av_log2(coef);
195 put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
196 put_bits(pb, len, coef & ((1 << len) - 1));
208 #define QUANTIZE_AND_ENCODE_BAND_COST_FUNC(NAME, BT_ZERO, BT_UNSIGNED, BT_PAIR, BT_ESC) \
209 static float quantize_and_encode_band_cost_ ## NAME( \
210 struct AACEncContext *s, \
211 PutBitContext *pb, const float *in, \
212 const float *scaled, int size, int scale_idx, \
213 int cb, const float lambda, const float uplim, \
215 return quantize_and_encode_band_cost_template( \
216 s, pb, in, scaled, size, scale_idx, \
217 BT_ESC ? ESC_BT : cb, lambda, uplim, bits, \
218 BT_ZERO, BT_UNSIGNED, BT_PAIR, BT_ESC); \
221 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(ZERO, 1, 0, 0, 0)
222 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(SQUAD, 0, 0, 0, 0)
223 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(UQUAD, 0, 1, 0, 0)
224 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(SPAIR, 0, 0, 1, 0)
225 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(UPAIR, 0, 1, 1, 0)
226 QUANTIZE_AND_ENCODE_BAND_COST_FUNC(ESC, 0, 1, 1, 1)
228 static float (*const quantize_and_encode_band_cost_arr[])(
229 struct AACEncContext *s,
230 PutBitContext *pb, const float *in,
231 const float *scaled, int size, int scale_idx,
232 int cb, const float lambda, const float uplim,
234 quantize_and_encode_band_cost_ZERO,
235 quantize_and_encode_band_cost_SQUAD,
236 quantize_and_encode_band_cost_SQUAD,
237 quantize_and_encode_band_cost_UQUAD,
238 quantize_and_encode_band_cost_UQUAD,
239 quantize_and_encode_band_cost_SPAIR,
240 quantize_and_encode_band_cost_SPAIR,
241 quantize_and_encode_band_cost_UPAIR,
242 quantize_and_encode_band_cost_UPAIR,
243 quantize_and_encode_band_cost_UPAIR,
244 quantize_and_encode_band_cost_UPAIR,
245 quantize_and_encode_band_cost_ESC,
248 #define quantize_and_encode_band_cost( \
249 s, pb, in, scaled, size, scale_idx, cb, \
250 lambda, uplim, bits) \
251 quantize_and_encode_band_cost_arr[cb]( \
252 s, pb, in, scaled, size, scale_idx, cb, \
255 static float quantize_band_cost(struct AACEncContext *s, const float *in,
256 const float *scaled, int size, int scale_idx,
257 int cb, const float lambda, const float uplim,
260 return quantize_and_encode_band_cost(s, NULL, in, scaled, size, scale_idx,
261 cb, lambda, uplim, bits);
264 static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
265 const float *in, int size, int scale_idx,
266 int cb, const float lambda)
268 quantize_and_encode_band_cost(s, pb, in, NULL, size, scale_idx, cb, lambda,
272 static float find_max_val(int group_len, int swb_size, const float *scaled) {
275 for (w2 = 0; w2 < group_len; w2++) {
276 for (i = 0; i < swb_size; i++) {
277 maxval = FFMAX(maxval, scaled[w2*128+i]);
283 static int find_min_book(float maxval, int sf) {
284 float Q = ff_aac_pow2sf_tab[POW_SF2_ZERO - sf + SCALE_ONE_POS - SCALE_DIV_512];
285 float Q34 = sqrtf(Q * sqrtf(Q));
287 qmaxval = maxval * Q34 + 0.4054f;
288 if (qmaxval == 0) cb = 0;
289 else if (qmaxval == 1) cb = 1;
290 else if (qmaxval == 2) cb = 3;
291 else if (qmaxval <= 4) cb = 5;
292 else if (qmaxval <= 7) cb = 7;
293 else if (qmaxval <= 12) cb = 9;
299 * structure used in optimal codebook search
301 typedef struct BandCodingPath {
302 int prev_idx; ///< pointer to the previous path point
303 float cost; ///< path cost
308 * Encode band info for single window group bands.
310 static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
311 int win, int group_len, const float lambda)
313 BandCodingPath path[120][12];
314 int w, swb, cb, start, size;
316 const int max_sfb = sce->ics.max_sfb;
317 const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
318 const int run_esc = (1 << run_bits) - 1;
319 int idx, ppos, count;
320 int stackrun[120], stackcb[120], stack_len;
321 float next_minrd = INFINITY;
324 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
326 for (cb = 0; cb < 12; cb++) {
327 path[0][cb].cost = 0.0f;
328 path[0][cb].prev_idx = -1;
331 for (swb = 0; swb < max_sfb; swb++) {
332 size = sce->ics.swb_sizes[swb];
333 if (sce->zeroes[win*16 + swb]) {
334 for (cb = 0; cb < 12; cb++) {
335 path[swb+1][cb].prev_idx = cb;
336 path[swb+1][cb].cost = path[swb][cb].cost;
337 path[swb+1][cb].run = path[swb][cb].run + 1;
340 float minrd = next_minrd;
341 int mincb = next_mincb;
342 next_minrd = INFINITY;
344 for (cb = 0; cb < 12; cb++) {
345 float cost_stay_here, cost_get_here;
347 for (w = 0; w < group_len; w++) {
348 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(win+w)*16+swb];
349 rd += quantize_band_cost(s, sce->coeffs + start + w*128,
350 s->scoefs + start + w*128, size,
351 sce->sf_idx[(win+w)*16+swb], cb,
352 lambda / band->threshold, INFINITY, NULL);
354 cost_stay_here = path[swb][cb].cost + rd;
355 cost_get_here = minrd + rd + run_bits + 4;
356 if ( run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
357 != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
358 cost_stay_here += run_bits;
359 if (cost_get_here < cost_stay_here) {
360 path[swb+1][cb].prev_idx = mincb;
361 path[swb+1][cb].cost = cost_get_here;
362 path[swb+1][cb].run = 1;
364 path[swb+1][cb].prev_idx = cb;
365 path[swb+1][cb].cost = cost_stay_here;
366 path[swb+1][cb].run = path[swb][cb].run + 1;
368 if (path[swb+1][cb].cost < next_minrd) {
369 next_minrd = path[swb+1][cb].cost;
374 start += sce->ics.swb_sizes[swb];
377 //convert resulting path from backward-linked list
380 for (cb = 1; cb < 12; cb++)
381 if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
386 stackrun[stack_len] = path[ppos][cb].run;
387 stackcb [stack_len] = cb;
388 idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
389 ppos -= path[ppos][cb].run;
392 //perform actual band info encoding
394 for (i = stack_len - 1; i >= 0; i--) {
395 put_bits(&s->pb, 4, stackcb[i]);
397 memset(sce->zeroes + win*16 + start, !stackcb[i], count);
398 //XXX: memset when band_type is also uint8_t
399 for (j = 0; j < count; j++) {
400 sce->band_type[win*16 + start] = stackcb[i];
403 while (count >= run_esc) {
404 put_bits(&s->pb, run_bits, run_esc);
407 put_bits(&s->pb, run_bits, count);
411 static void codebook_trellis_rate(AACEncContext *s, SingleChannelElement *sce,
412 int win, int group_len, const float lambda)
414 BandCodingPath path[120][12];
415 int w, swb, cb, start, size;
417 const int max_sfb = sce->ics.max_sfb;
418 const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
419 const int run_esc = (1 << run_bits) - 1;
420 int idx, ppos, count;
421 int stackrun[120], stackcb[120], stack_len;
422 float next_minrd = INFINITY;
425 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
427 for (cb = 0; cb < 12; cb++) {
428 path[0][cb].cost = run_bits+4;
429 path[0][cb].prev_idx = -1;
432 for (swb = 0; swb < max_sfb; swb++) {
433 size = sce->ics.swb_sizes[swb];
434 if (sce->zeroes[win*16 + swb]) {
435 float cost_stay_here = path[swb][0].cost;
436 float cost_get_here = next_minrd + run_bits + 4;
437 if ( run_value_bits[sce->ics.num_windows == 8][path[swb][0].run]
438 != run_value_bits[sce->ics.num_windows == 8][path[swb][0].run+1])
439 cost_stay_here += run_bits;
440 if (cost_get_here < cost_stay_here) {
441 path[swb+1][0].prev_idx = next_mincb;
442 path[swb+1][0].cost = cost_get_here;
443 path[swb+1][0].run = 1;
445 path[swb+1][0].prev_idx = 0;
446 path[swb+1][0].cost = cost_stay_here;
447 path[swb+1][0].run = path[swb][0].run + 1;
449 next_minrd = path[swb+1][0].cost;
451 for (cb = 1; cb < 12; cb++) {
452 path[swb+1][cb].cost = 61450;
453 path[swb+1][cb].prev_idx = -1;
454 path[swb+1][cb].run = 0;
457 float minrd = next_minrd;
458 int mincb = next_mincb;
459 int startcb = sce->band_type[win*16+swb];
460 next_minrd = INFINITY;
462 for (cb = 0; cb < startcb; cb++) {
463 path[swb+1][cb].cost = 61450;
464 path[swb+1][cb].prev_idx = -1;
465 path[swb+1][cb].run = 0;
467 for (cb = startcb; cb < 12; cb++) {
468 float cost_stay_here, cost_get_here;
470 for (w = 0; w < group_len; w++) {
471 rd += quantize_band_cost(s, sce->coeffs + start + w*128,
472 s->scoefs + start + w*128, size,
473 sce->sf_idx[(win+w)*16+swb], cb,
476 cost_stay_here = path[swb][cb].cost + rd;
477 cost_get_here = minrd + rd + run_bits + 4;
478 if ( run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
479 != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
480 cost_stay_here += run_bits;
481 if (cost_get_here < cost_stay_here) {
482 path[swb+1][cb].prev_idx = mincb;
483 path[swb+1][cb].cost = cost_get_here;
484 path[swb+1][cb].run = 1;
486 path[swb+1][cb].prev_idx = cb;
487 path[swb+1][cb].cost = cost_stay_here;
488 path[swb+1][cb].run = path[swb][cb].run + 1;
490 if (path[swb+1][cb].cost < next_minrd) {
491 next_minrd = path[swb+1][cb].cost;
496 start += sce->ics.swb_sizes[swb];
499 //convert resulting path from backward-linked list
502 for (cb = 1; cb < 12; cb++)
503 if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
509 stackrun[stack_len] = path[ppos][cb].run;
510 stackcb [stack_len] = cb;
511 idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
512 ppos -= path[ppos][cb].run;
515 //perform actual band info encoding
517 for (i = stack_len - 1; i >= 0; i--) {
518 put_bits(&s->pb, 4, stackcb[i]);
520 memset(sce->zeroes + win*16 + start, !stackcb[i], count);
521 //XXX: memset when band_type is also uint8_t
522 for (j = 0; j < count; j++) {
523 sce->band_type[win*16 + start] = stackcb[i];
526 while (count >= run_esc) {
527 put_bits(&s->pb, run_bits, run_esc);
530 put_bits(&s->pb, run_bits, count);
534 /** Return the minimum scalefactor where the quantized coef does not clip. */
535 static av_always_inline uint8_t coef2minsf(float coef) {
536 return av_clip_uint8(log2f(coef)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
539 /** Return the maximum scalefactor where the quantized coef is not zero. */
540 static av_always_inline uint8_t coef2maxsf(float coef) {
541 return av_clip_uint8(log2f(coef)*4 + 6 + SCALE_ONE_POS - SCALE_DIV_512);
544 typedef struct TrellisPath {
549 #define TRELLIS_STAGES 121
550 #define TRELLIS_STATES (SCALE_MAX_DIFF+1)
552 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
553 SingleChannelElement *sce,
556 int q, w, w2, g, start = 0;
559 TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
560 int bandaddr[TRELLIS_STAGES];
563 float q0f = FLT_MAX, q1f = 0.0f, qnrgf = 0.0f;
564 int q0, q1, qcnt = 0;
566 for (i = 0; i < 1024; i++) {
567 float t = fabsf(sce->coeffs[i]);
577 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
578 memset(sce->zeroes, 1, sizeof(sce->zeroes));
582 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
583 q0 = coef2minsf(q0f);
584 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
585 q1 = coef2maxsf(q1f);
586 //av_log(NULL, AV_LOG_ERROR, "q0 %d, q1 %d\n", q0, q1);
590 //minimum scalefactor index is when maximum nonzero coefficient after quantizing is not clipped
591 int qnrg = av_clip_uint8(log2f(sqrtf(qnrgf/qcnt))*4 - 31 + SCALE_ONE_POS - SCALE_DIV_512);
594 //av_log(NULL, AV_LOG_ERROR, "q0 %d, q1 %d\n", q0, q1);
598 } else if (q1 > q1high) {
603 //av_log(NULL, AV_LOG_ERROR, "q0 %d, q1 %d\n", q0, q1);
605 for (i = 0; i < TRELLIS_STATES; i++) {
606 paths[0][i].cost = 0.0f;
607 paths[0][i].prev = -1;
609 for (j = 1; j < TRELLIS_STAGES; j++) {
610 for (i = 0; i < TRELLIS_STATES; i++) {
611 paths[j][i].cost = INFINITY;
612 paths[j][i].prev = -2;
616 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
617 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
619 for (g = 0; g < sce->ics.num_swb; g++) {
620 const float *coefs = sce->coeffs + start;
624 bandaddr[idx] = w * 16 + g;
627 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
628 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
629 if (band->energy <= band->threshold || band->threshold == 0.0f) {
630 sce->zeroes[(w+w2)*16+g] = 1;
633 sce->zeroes[(w+w2)*16+g] = 0;
635 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
636 float t = fabsf(coefs[w2*128+i]);
638 qmin = FFMIN(qmin, t);
639 qmax = FFMAX(qmax, t);
643 int minscale, maxscale;
644 float minrd = INFINITY;
646 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
647 minscale = coef2minsf(qmin);
648 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
649 maxscale = coef2maxsf(qmax);
650 minscale = av_clip(minscale - q0, 0, TRELLIS_STATES - 1);
651 maxscale = av_clip(maxscale - q0, 0, TRELLIS_STATES);
652 maxval = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], s->scoefs+start);
653 for (q = minscale; q < maxscale; q++) {
655 int cb = find_min_book(maxval, sce->sf_idx[w*16+g]);
656 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
657 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
658 dist += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
659 q + q0, cb, lambda / band->threshold, INFINITY, NULL);
661 minrd = FFMIN(minrd, dist);
663 for (i = 0; i < q1 - q0; i++) {
665 cost = paths[idx - 1][i].cost + dist
666 + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
667 if (cost < paths[idx][q].cost) {
668 paths[idx][q].cost = cost;
669 paths[idx][q].prev = i;
674 for (q = 0; q < q1 - q0; q++) {
675 paths[idx][q].cost = paths[idx - 1][q].cost + 1;
676 paths[idx][q].prev = q;
679 sce->zeroes[w*16+g] = !nz;
680 start += sce->ics.swb_sizes[g];
685 mincost = paths[idx][0].cost;
687 for (i = 1; i < TRELLIS_STATES; i++) {
688 if (paths[idx][i].cost < mincost) {
689 mincost = paths[idx][i].cost;
694 sce->sf_idx[bandaddr[idx]] = minq + q0;
695 minq = paths[idx][minq].prev;
698 //set the same quantizers inside window groups
699 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
700 for (g = 0; g < sce->ics.num_swb; g++)
701 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
702 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
706 * two-loop quantizers search taken from ISO 13818-7 Appendix C
708 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
710 SingleChannelElement *sce,
713 int start = 0, i, w, w2, g;
714 int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels;
715 float dists[128], uplims[128];
717 int fflag, minscaler;
720 float minthr = INFINITY;
722 //XXX: some heuristic to determine initial quantizers will reduce search time
723 memset(dists, 0, sizeof(dists));
724 //determine zero bands and upper limits
725 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
726 for (g = 0; g < sce->ics.num_swb; g++) {
729 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
730 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
731 uplim += band->threshold;
732 if (band->energy <= band->threshold || band->threshold == 0.0f) {
733 sce->zeroes[(w+w2)*16+g] = 1;
738 uplims[w*16+g] = uplim *512;
739 sce->zeroes[w*16+g] = !nz;
741 minthr = FFMIN(minthr, uplim);
745 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
746 for (g = 0; g < sce->ics.num_swb; g++) {
747 if (sce->zeroes[w*16+g]) {
748 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
751 sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2f(uplims[w*16+g]/minthr)*4,59);
757 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
759 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
761 for (g = 0; g < sce->ics.num_swb; g++) {
762 const float *scaled = s->scoefs + start;
763 maxvals[w*16+g] = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled);
764 start += sce->ics.swb_sizes[g];
768 //perform two-loop search
769 //outer loop - improve quality
772 minscaler = sce->sf_idx[0];
773 //inner loop - quantize spectrum to fit into given number of bits
774 qstep = its ? 1 : 32;
779 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
781 for (g = 0; g < sce->ics.num_swb; g++) {
782 const float *coefs = sce->coeffs + start;
783 const float *scaled = s->scoefs + start;
788 if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
789 start += sce->ics.swb_sizes[g];
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++) {
796 dist += quantize_band_cost(s, coefs + w2*128,
798 sce->ics.swb_sizes[g],
806 dists[w*16+g] = dist - bits;
808 bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
811 start += sce->ics.swb_sizes[g];
812 prev = sce->sf_idx[w*16+g];
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;
820 for (i = 0; i < 128; i++)
821 if (sce->sf_idx[i] > 60 - qstep)
822 sce->sf_idx[i] -= qstep;
825 if (!qstep && tbits > destbits*1.02 && sce->sf_idx[0] < 217)
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;
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)
844 sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
848 } while (fflag && its < 10);
851 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
852 SingleChannelElement *sce,
855 int start = 0, i, w, w2, g;
856 float uplim[128], maxq[128];
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) {
863 for (i = 0; i < 1024; i++) {
864 if (i - start >= sce->ics.swb_sizes[curband]) {
865 start += sce->ics.swb_sizes[curband];
868 if (sce->coeffs[i]) {
869 avg_energy += sce->coeffs[i] * sce->coeffs[i];
875 for (w = 0; w < 8; w++) {
876 const float *coeffs = sce->coeffs + w*128;
878 for (i = 0; i < 128; i++) {
879 if (i - start >= sce->ics.swb_sizes[curband]) {
880 start += sce->ics.swb_sizes[curband];
884 avg_energy += coeffs[i] * coeffs[i];
885 last = FFMAX(last, i);
886 lastband = FFMAX(lastband, curband);
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;
898 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
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;
909 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
910 memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
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]));
918 if (sce->ics.num_windows == 1 && maxval < t) {
924 if (sce->ics.num_windows == 1) {
925 start2 = FFMAX(peakpos - 2, start2);
926 end2 = FFMIN(peakpos + 3, end2);
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);
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]) {
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;
948 if (maxq[w*16+g] < 21.544) {
949 sce->zeroes[w*16+g] = 1;
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);
960 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
962 dist += quantize_band_cost(s, coefs + w2*128,
964 sce->ics.swb_sizes[g],
972 dist *= 1.0f / 512.0f / lambda;
973 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[POW_SF2_ZERO - scf + SCALE_ONE_POS - SCALE_DIV_512]);
974 if (quant_max >= 8191) { // too much, return to the previous quantizer
975 sce->sf_idx[w*16+g] = prev_scf;
979 curdiff = fabsf(dist - uplim[w*16+g]);
983 step = log2f(curdiff);
984 if (dist > uplim[w*16+g])
987 scf = av_clip_uint8(scf);
988 step = scf - prev_scf;
989 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
990 sce->sf_idx[w*16+g] = av_clip(scf, min_scf, max_scf);
1001 minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
1002 for (i = 1; i < 128; i++) {
1003 if (!sce->sf_idx[i])
1004 sce->sf_idx[i] = sce->sf_idx[i-1];
1006 minq = FFMIN(minq, sce->sf_idx[i]);
1008 if (minq == INT_MAX)
1010 minq = FFMIN(minq, SCALE_MAX_POS);
1011 maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
1012 for (i = 126; i >= 0; i--) {
1013 if (!sce->sf_idx[i])
1014 sce->sf_idx[i] = sce->sf_idx[i+1];
1015 sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
1019 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
1020 SingleChannelElement *sce,
1026 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
1027 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
1028 for (g = 0; g < sce->ics.num_swb; g++) {
1029 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
1030 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
1031 if (band->energy <= band->threshold) {
1032 sce->sf_idx[(w+w2)*16+g] = 218;
1033 sce->zeroes[(w+w2)*16+g] = 1;
1035 sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2f(band->threshold), 80, 218);
1036 sce->zeroes[(w+w2)*16+g] = 0;
1038 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
1042 for (i = 0; i < 128; i++) {
1043 sce->sf_idx[i] = 140;
1044 //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
1046 //set the same quantizers inside window groups
1047 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
1048 for (g = 0; g < sce->ics.num_swb; g++)
1049 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
1050 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
1053 static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
1056 int start = 0, i, w, w2, g;
1057 float M[128], S[128];
1058 float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
1059 SingleChannelElement *sce0 = &cpe->ch[0];
1060 SingleChannelElement *sce1 = &cpe->ch[1];
1061 if (!cpe->common_window)
1063 for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
1064 for (g = 0; g < sce0->ics.num_swb; g++) {
1065 if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
1066 float dist1 = 0.0f, dist2 = 0.0f;
1067 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
1068 FFPsyBand *band0 = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g];
1069 FFPsyBand *band1 = &s->psy.ch[s->cur_channel+1].psy_bands[(w+w2)*16+g];
1070 float minthr = FFMIN(band0->threshold, band1->threshold);
1071 float maxthr = FFMAX(band0->threshold, band1->threshold);
1072 for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
1073 M[i] = (sce0->coeffs[start+w2*128+i]
1074 + sce1->coeffs[start+w2*128+i]) * 0.5;
1076 - sce1->coeffs[start+w2*128+i];
1078 abs_pow34_v(L34, sce0->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
1079 abs_pow34_v(R34, sce1->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
1080 abs_pow34_v(M34, M, sce0->ics.swb_sizes[g]);
1081 abs_pow34_v(S34, S, sce0->ics.swb_sizes[g]);
1082 dist1 += quantize_band_cost(s, sce0->coeffs + start + w2*128,
1084 sce0->ics.swb_sizes[g],
1085 sce0->sf_idx[(w+w2)*16+g],
1086 sce0->band_type[(w+w2)*16+g],
1087 lambda / band0->threshold, INFINITY, NULL);
1088 dist1 += quantize_band_cost(s, sce1->coeffs + start + w2*128,
1090 sce1->ics.swb_sizes[g],
1091 sce1->sf_idx[(w+w2)*16+g],
1092 sce1->band_type[(w+w2)*16+g],
1093 lambda / band1->threshold, INFINITY, NULL);
1094 dist2 += quantize_band_cost(s, M,
1096 sce0->ics.swb_sizes[g],
1097 sce0->sf_idx[(w+w2)*16+g],
1098 sce0->band_type[(w+w2)*16+g],
1099 lambda / maxthr, INFINITY, NULL);
1100 dist2 += quantize_band_cost(s, S,
1102 sce1->ics.swb_sizes[g],
1103 sce1->sf_idx[(w+w2)*16+g],
1104 sce1->band_type[(w+w2)*16+g],
1105 lambda / minthr, INFINITY, NULL);
1107 cpe->ms_mask[w*16+g] = dist2 < dist1;
1109 start += sce0->ics.swb_sizes[g];
1114 AACCoefficientsEncoder ff_aac_coders[] = {
1116 search_for_quantizers_faac,
1117 encode_window_bands_info,
1118 quantize_and_encode_band,
1122 search_for_quantizers_anmr,
1123 encode_window_bands_info,
1124 quantize_and_encode_band,
1128 search_for_quantizers_twoloop,
1129 codebook_trellis_rate,
1130 quantize_and_encode_band,
1134 search_for_quantizers_fast,
1135 encode_window_bands_info,
1136 quantize_and_encode_band,