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
37 #include "libavutil/mathematics.h"
44 #include "aacenctab.h"
45 #include "aacenc_utils.h"
46 #include "aacenc_quantization.h"
48 #include "aacenc_is.h"
49 #include "aacenc_tns.h"
50 #include "aacenc_ltp.h"
51 #include "aacenc_pred.h"
53 #include "libavcodec/aaccoder_twoloop.h"
55 /* Parameter of f(x) = a*(lambda/100), defines the maximum fourier spread
56 * beyond which no PNS is used (since the SFBs contain tone rather than noise) */
57 #define NOISE_SPREAD_THRESHOLD 0.9f
59 /* Parameter of f(x) = a*(100/lambda), defines how much PNS is allowed to
60 * replace low energy non zero bands */
61 #define NOISE_LAMBDA_REPLACE 1.948f
63 #include "libavcodec/aaccoder_trellis.h"
66 * structure used in optimal codebook search
68 typedef struct BandCodingPath {
69 int prev_idx; ///< pointer to the previous path point
70 float cost; ///< path cost
75 * Encode band info for single window group bands.
77 static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
78 int win, int group_len, const float lambda)
80 BandCodingPath path[120][CB_TOT_ALL];
81 int w, swb, cb, start, size;
83 const int max_sfb = sce->ics.max_sfb;
84 const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
85 const int run_esc = (1 << run_bits) - 1;
87 int stackrun[120], stackcb[120], stack_len;
88 float next_minrd = INFINITY;
91 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
93 for (cb = 0; cb < CB_TOT_ALL; cb++) {
94 path[0][cb].cost = 0.0f;
95 path[0][cb].prev_idx = -1;
98 for (swb = 0; swb < max_sfb; swb++) {
99 size = sce->ics.swb_sizes[swb];
100 if (sce->zeroes[win*16 + swb]) {
101 for (cb = 0; cb < CB_TOT_ALL; cb++) {
102 path[swb+1][cb].prev_idx = cb;
103 path[swb+1][cb].cost = path[swb][cb].cost;
104 path[swb+1][cb].run = path[swb][cb].run + 1;
107 float minrd = next_minrd;
108 int mincb = next_mincb;
109 next_minrd = INFINITY;
111 for (cb = 0; cb < CB_TOT_ALL; cb++) {
112 float cost_stay_here, cost_get_here;
114 if (cb >= 12 && sce->band_type[win*16+swb] < aac_cb_out_map[cb] ||
115 cb < aac_cb_in_map[sce->band_type[win*16+swb]] && sce->band_type[win*16+swb] > aac_cb_out_map[cb]) {
116 path[swb+1][cb].prev_idx = -1;
117 path[swb+1][cb].cost = INFINITY;
118 path[swb+1][cb].run = path[swb][cb].run + 1;
121 for (w = 0; w < group_len; w++) {
122 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(win+w)*16+swb];
123 rd += quantize_band_cost(s, &sce->coeffs[start + w*128],
124 &s->scoefs[start + w*128], size,
125 sce->sf_idx[(win+w)*16+swb], aac_cb_out_map[cb],
126 lambda / band->threshold, INFINITY, NULL, NULL, 0);
128 cost_stay_here = path[swb][cb].cost + rd;
129 cost_get_here = minrd + rd + run_bits + 4;
130 if ( run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
131 != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
132 cost_stay_here += run_bits;
133 if (cost_get_here < cost_stay_here) {
134 path[swb+1][cb].prev_idx = mincb;
135 path[swb+1][cb].cost = cost_get_here;
136 path[swb+1][cb].run = 1;
138 path[swb+1][cb].prev_idx = cb;
139 path[swb+1][cb].cost = cost_stay_here;
140 path[swb+1][cb].run = path[swb][cb].run + 1;
142 if (path[swb+1][cb].cost < next_minrd) {
143 next_minrd = path[swb+1][cb].cost;
148 start += sce->ics.swb_sizes[swb];
151 //convert resulting path from backward-linked list
154 for (cb = 1; cb < CB_TOT_ALL; cb++)
155 if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
159 av_assert1(idx >= 0);
161 stackrun[stack_len] = path[ppos][cb].run;
162 stackcb [stack_len] = cb;
163 idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
164 ppos -= path[ppos][cb].run;
167 //perform actual band info encoding
169 for (i = stack_len - 1; i >= 0; i--) {
170 cb = aac_cb_out_map[stackcb[i]];
171 put_bits(&s->pb, 4, cb);
173 memset(sce->zeroes + win*16 + start, !cb, count);
174 //XXX: memset when band_type is also uint8_t
175 for (j = 0; j < count; j++) {
176 sce->band_type[win*16 + start] = cb;
179 while (count >= run_esc) {
180 put_bits(&s->pb, run_bits, run_esc);
183 put_bits(&s->pb, run_bits, count);
188 typedef struct TrellisPath {
193 #define TRELLIS_STAGES 121
194 #define TRELLIS_STATES (SCALE_MAX_DIFF+1)
196 static void set_special_band_scalefactors(AACEncContext *s, SingleChannelElement *sce)
199 int minscaler_n = sce->sf_idx[0], minscaler_i = sce->sf_idx[0];
202 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
204 for (g = 0; g < sce->ics.num_swb; g++) {
205 if (sce->band_type[w*16+g] == INTENSITY_BT || sce->band_type[w*16+g] == INTENSITY_BT2) {
206 sce->sf_idx[w*16+g] = av_clip(roundf(log2f(sce->is_ener[w*16+g])*2), -155, 100);
207 minscaler_i = FFMIN(minscaler_i, sce->sf_idx[w*16+g]);
209 } else if (sce->band_type[w*16+g] == NOISE_BT) {
210 sce->sf_idx[w*16+g] = av_clip(3+ceilf(log2f(sce->pns_ener[w*16+g])*2), -100, 155);
211 minscaler_n = FFMIN(minscaler_n, sce->sf_idx[w*16+g]);
214 start += sce->ics.swb_sizes[g];
221 /* Clip the scalefactor indices */
222 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
223 for (g = 0; g < sce->ics.num_swb; g++) {
224 if (sce->band_type[w*16+g] == INTENSITY_BT || sce->band_type[w*16+g] == INTENSITY_BT2) {
225 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler_i, minscaler_i + SCALE_MAX_DIFF);
226 } else if (sce->band_type[w*16+g] == NOISE_BT) {
227 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler_n, minscaler_n + SCALE_MAX_DIFF);
233 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
234 SingleChannelElement *sce,
237 int q, w, w2, g, start = 0;
240 TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
241 int bandaddr[TRELLIS_STAGES];
244 float q0f = FLT_MAX, q1f = 0.0f, qnrgf = 0.0f;
245 int q0, q1, qcnt = 0;
247 for (i = 0; i < 1024; i++) {
248 float t = fabsf(sce->coeffs[i]);
258 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
259 memset(sce->zeroes, 1, sizeof(sce->zeroes));
263 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
264 q0 = av_clip(coef2minsf(q0f), 0, SCALE_MAX_POS-1);
265 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
266 q1 = av_clip(coef2maxsf(q1f), 1, SCALE_MAX_POS);
270 //minimum scalefactor index is when maximum nonzero coefficient after quantizing is not clipped
271 int qnrg = av_clip_uint8(log2f(sqrtf(qnrgf/qcnt))*4 - 31 + SCALE_ONE_POS - SCALE_DIV_512);
277 } else if (q1 > q1high) {
282 // q0 == q1 isn't really a legal situation
284 // the following is indirect but guarantees q1 != q0 && q1 near q0
285 q1 = av_clip(q0+1, 1, SCALE_MAX_POS);
286 q0 = av_clip(q1-1, 0, SCALE_MAX_POS - 1);
289 for (i = 0; i < TRELLIS_STATES; i++) {
290 paths[0][i].cost = 0.0f;
291 paths[0][i].prev = -1;
293 for (j = 1; j < TRELLIS_STAGES; j++) {
294 for (i = 0; i < TRELLIS_STATES; i++) {
295 paths[j][i].cost = INFINITY;
296 paths[j][i].prev = -2;
300 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
301 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
303 for (g = 0; g < sce->ics.num_swb; g++) {
304 const float *coefs = &sce->coeffs[start];
308 bandaddr[idx] = w * 16 + g;
311 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
312 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
313 if (band->energy <= band->threshold || band->threshold == 0.0f) {
314 sce->zeroes[(w+w2)*16+g] = 1;
317 sce->zeroes[(w+w2)*16+g] = 0;
319 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
320 float t = fabsf(coefs[w2*128+i]);
322 qmin = FFMIN(qmin, t);
323 qmax = FFMAX(qmax, t);
327 int minscale, maxscale;
328 float minrd = INFINITY;
330 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
331 minscale = coef2minsf(qmin);
332 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
333 maxscale = coef2maxsf(qmax);
334 minscale = av_clip(minscale - q0, 0, TRELLIS_STATES - 1);
335 maxscale = av_clip(maxscale - q0, 0, TRELLIS_STATES);
336 if (minscale == maxscale) {
337 maxscale = av_clip(minscale+1, 1, TRELLIS_STATES);
338 minscale = av_clip(maxscale-1, 0, TRELLIS_STATES - 1);
340 maxval = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], s->scoefs+start);
341 for (q = minscale; q < maxscale; q++) {
343 int cb = find_min_book(maxval, sce->sf_idx[w*16+g]);
344 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
345 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
346 dist += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
347 q + q0, cb, lambda / band->threshold, INFINITY, NULL, NULL, 0);
349 minrd = FFMIN(minrd, dist);
351 for (i = 0; i < q1 - q0; i++) {
353 cost = paths[idx - 1][i].cost + dist
354 + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
355 if (cost < paths[idx][q].cost) {
356 paths[idx][q].cost = cost;
357 paths[idx][q].prev = i;
362 for (q = 0; q < q1 - q0; q++) {
363 paths[idx][q].cost = paths[idx - 1][q].cost + 1;
364 paths[idx][q].prev = q;
367 sce->zeroes[w*16+g] = !nz;
368 start += sce->ics.swb_sizes[g];
373 mincost = paths[idx][0].cost;
375 for (i = 1; i < TRELLIS_STATES; i++) {
376 if (paths[idx][i].cost < mincost) {
377 mincost = paths[idx][i].cost;
382 sce->sf_idx[bandaddr[idx]] = minq + q0;
383 minq = FFMAX(paths[idx][minq].prev, 0);
386 //set the same quantizers inside window groups
387 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
388 for (g = 0; g < sce->ics.num_swb; g++)
389 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
390 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
394 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
395 SingleChannelElement *sce,
398 int start = 0, i, w, w2, g;
399 float uplim[128], maxq[128];
401 float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
402 int last = 0, lastband = 0, curband = 0;
403 float avg_energy = 0.0;
404 if (sce->ics.num_windows == 1) {
406 for (i = 0; i < 1024; i++) {
407 if (i - start >= sce->ics.swb_sizes[curband]) {
408 start += sce->ics.swb_sizes[curband];
411 if (sce->coeffs[i]) {
412 avg_energy += sce->coeffs[i] * sce->coeffs[i];
418 for (w = 0; w < 8; w++) {
419 const float *coeffs = &sce->coeffs[w*128];
421 for (i = 0; i < 128; i++) {
422 if (i - start >= sce->ics.swb_sizes[curband]) {
423 start += sce->ics.swb_sizes[curband];
427 avg_energy += coeffs[i] * coeffs[i];
428 last = FFMAX(last, i);
429 lastband = FFMAX(lastband, curband);
436 if (avg_energy == 0.0f) {
437 for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
438 sce->sf_idx[i] = SCALE_ONE_POS;
441 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
443 for (g = 0; g < sce->ics.num_swb; g++) {
444 float *coefs = &sce->coeffs[start];
445 const int size = sce->ics.swb_sizes[g];
446 int start2 = start, end2 = start + size, peakpos = start;
447 float maxval = -1, thr = 0.0f, t;
452 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
453 memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
456 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
457 for (i = 0; i < size; i++) {
458 float t = coefs[w2*128+i]*coefs[w2*128+i];
459 maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
461 if (sce->ics.num_windows == 1 && maxval < t) {
467 if (sce->ics.num_windows == 1) {
468 start2 = FFMAX(peakpos - 2, start2);
469 end2 = FFMIN(peakpos + 3, end2);
475 thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
476 t = 1.0 - (1.0 * start2 / last);
477 uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
480 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
481 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
482 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
484 for (g = 0; g < sce->ics.num_swb; g++) {
485 const float *coefs = &sce->coeffs[start];
486 const float *scaled = &s->scoefs[start];
487 const int size = sce->ics.swb_sizes[g];
488 int scf, prev_scf, step;
489 int min_scf = -1, max_scf = 256;
491 if (maxq[w*16+g] < 21.544) {
492 sce->zeroes[w*16+g] = 1;
496 sce->zeroes[w*16+g] = 0;
497 scf = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2f(1/maxq[w*16+g])*16/3, 60, 218);
502 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
504 dist += quantize_band_cost(s, coefs + w2*128,
506 sce->ics.swb_sizes[g],
515 dist *= 1.0f / 512.0f / lambda;
516 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[POW_SF2_ZERO - scf + SCALE_ONE_POS - SCALE_DIV_512], ROUND_STANDARD);
517 if (quant_max >= 8191) { // too much, return to the previous quantizer
518 sce->sf_idx[w*16+g] = prev_scf;
522 curdiff = fabsf(dist - uplim[w*16+g]);
526 step = log2f(curdiff);
527 if (dist > uplim[w*16+g])
530 scf = av_clip_uint8(scf);
531 step = scf - prev_scf;
532 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
533 sce->sf_idx[w*16+g] = av_clip(scf, min_scf, max_scf);
544 minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
545 for (i = 1; i < 128; i++) {
547 sce->sf_idx[i] = sce->sf_idx[i-1];
549 minq = FFMIN(minq, sce->sf_idx[i]);
553 minq = FFMIN(minq, SCALE_MAX_POS);
554 maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
555 for (i = 126; i >= 0; i--) {
557 sce->sf_idx[i] = sce->sf_idx[i+1];
558 sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
562 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
563 SingleChannelElement *sce,
569 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
570 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
571 for (g = 0; g < sce->ics.num_swb; g++) {
572 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
573 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
574 if (band->energy <= band->threshold) {
575 sce->sf_idx[(w+w2)*16+g] = 218;
576 sce->zeroes[(w+w2)*16+g] = 1;
578 sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2f(band->threshold), 80, 218);
579 sce->zeroes[(w+w2)*16+g] = 0;
581 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
585 for (i = 0; i < 128; i++) {
586 sce->sf_idx[i] = 140;
587 //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
589 //set the same quantizers inside window groups
590 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
591 for (g = 0; g < sce->ics.num_swb; g++)
592 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
593 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
596 static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChannelElement *sce)
600 int wlen = 1024 / sce->ics.num_windows;
601 int bandwidth, cutoff;
602 float *PNS = &s->scoefs[0*128], *PNS34 = &s->scoefs[1*128];
603 float *NOR34 = &s->scoefs[3*128];
604 uint8_t nextband[128];
605 const float lambda = s->lambda;
606 const float freq_mult = avctx->sample_rate*0.5f/wlen;
607 const float thr_mult = NOISE_LAMBDA_REPLACE*(100.0f/lambda);
608 const float spread_threshold = FFMIN(0.75f, NOISE_SPREAD_THRESHOLD*FFMAX(0.5f, lambda/100.f));
609 const float dist_bias = av_clipf(4.f * 120 / lambda, 0.25f, 4.0f);
610 const float pns_transient_energy_r = FFMIN(0.7f, lambda / 140.f);
612 int refbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
613 / ((avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : avctx->channels)
616 /** Keep this in sync with twoloop's cutoff selection */
617 float rate_bandwidth_multiplier = 1.5f;
618 int prev = -1000, prev_sf = -1;
619 int frame_bit_rate = (avctx->flags & CODEC_FLAG_QSCALE)
620 ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
621 : (avctx->bit_rate / avctx->channels);
623 frame_bit_rate *= 1.15f;
625 if (avctx->cutoff > 0) {
626 bandwidth = avctx->cutoff;
628 bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
631 cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
633 memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type));
634 ff_init_nextband_map(sce, nextband);
635 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
637 for (g = 0; g < sce->ics.num_swb; g++) {
639 float dist1 = 0.0f, dist2 = 0.0f, noise_amp;
640 float pns_energy = 0.0f, pns_tgt_energy, energy_ratio, dist_thresh;
641 float sfb_energy = 0.0f, threshold = 0.0f, spread = 2.0f;
642 float min_energy = -1.0f, max_energy = 0.0f;
643 const int start = wstart+sce->ics.swb_offset[g];
644 const float freq = (start-wstart)*freq_mult;
645 const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
646 if (freq < NOISE_LOW_LIMIT || (start-wstart) >= cutoff)
648 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
649 band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
650 sfb_energy += band->energy;
651 spread = FFMIN(spread, band->spread);
652 threshold += band->threshold;
654 min_energy = max_energy = band->energy;
656 min_energy = FFMIN(min_energy, band->energy);
657 max_energy = FFMAX(max_energy, band->energy);
661 /* Ramps down at ~8000Hz and loosens the dist threshold */
662 dist_thresh = av_clipf(2.5f*NOISE_LOW_LIMIT/freq, 0.5f, 2.5f) * dist_bias;
664 /* PNS is acceptable when all of these are true:
665 * 1. high spread energy (noise-like band)
666 * 2. near-threshold energy (high PE means the random nature of PNS content will be noticed)
667 * 3. on short window groups, all windows have similar energy (variations in energy would be destroyed by PNS)
669 * At this stage, point 2 is relaxed for zeroed bands near the noise threshold (hole avoidance is more important)
671 if ((!sce->zeroes[w*16+g] && !ff_sfdelta_can_remove_band(sce, nextband, prev_sf, w*16+g)) ||
672 ((sce->zeroes[w*16+g] || !sce->band_alt[w*16+g]) && sfb_energy < threshold*sqrtf(1.0f/freq_boost)) || spread < spread_threshold ||
673 (!sce->zeroes[w*16+g] && sce->band_alt[w*16+g] && sfb_energy > threshold*thr_mult*freq_boost) ||
674 min_energy < pns_transient_energy_r * max_energy ) {
675 sce->pns_ener[w*16+g] = sfb_energy;
676 if (!sce->zeroes[w*16+g])
677 prev_sf = sce->sf_idx[w*16+g];
681 pns_tgt_energy = sfb_energy*FFMIN(1.0f, spread*spread);
682 noise_sfi = av_clip(roundf(log2f(pns_tgt_energy)*2), -100, 155); /* Quantize */
683 noise_amp = -ff_aac_pow2sf_tab[noise_sfi + POW_SF2_ZERO]; /* Dequantize */
685 int noise_sfdiff = noise_sfi - prev + SCALE_DIFF_ZERO;
686 if (noise_sfdiff < 0 || noise_sfdiff > 2*SCALE_MAX_DIFF) {
687 if (!sce->zeroes[w*16+g])
688 prev_sf = sce->sf_idx[w*16+g];
692 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
693 float band_energy, scale, pns_senergy;
694 const int start_c = (w+w2)*128+sce->ics.swb_offset[g];
695 band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
696 for (i = 0; i < sce->ics.swb_sizes[g]; i+=2) {
698 av_bmg_get(&s->lfg, rnd);
699 PNS[i+0] = (float)rnd[0];
700 PNS[i+1] = (float)rnd[1];
702 band_energy = s->fdsp->scalarproduct_float(PNS, PNS, sce->ics.swb_sizes[g]);
703 scale = noise_amp/sqrtf(band_energy);
704 s->fdsp->vector_fmul_scalar(PNS, PNS, scale, sce->ics.swb_sizes[g]);
705 pns_senergy = s->fdsp->scalarproduct_float(PNS, PNS, sce->ics.swb_sizes[g]);
706 pns_energy += pns_senergy;
707 abs_pow34_v(NOR34, &sce->coeffs[start_c], sce->ics.swb_sizes[g]);
708 abs_pow34_v(PNS34, PNS, sce->ics.swb_sizes[g]);
709 dist1 += quantize_band_cost(s, &sce->coeffs[start_c],
711 sce->ics.swb_sizes[g],
712 sce->sf_idx[(w+w2)*16+g],
713 sce->band_alt[(w+w2)*16+g],
714 lambda/band->threshold, INFINITY, NULL, NULL, 0);
715 /* Estimate rd on average as 5 bits for SF, 4 for the CB, plus spread energy * lambda/thr */
716 dist2 += band->energy/(band->spread*band->spread)*lambda*dist_thresh/band->threshold;
718 if (g && sce->band_type[w*16+g-1] == NOISE_BT) {
723 energy_ratio = pns_tgt_energy/pns_energy; /* Compensates for quantization error */
724 sce->pns_ener[w*16+g] = energy_ratio*pns_tgt_energy;
725 if (sce->zeroes[w*16+g] || !sce->band_alt[w*16+g] || (energy_ratio > 0.85f && energy_ratio < 1.25f && dist2 < dist1)) {
726 sce->band_type[w*16+g] = NOISE_BT;
727 sce->zeroes[w*16+g] = 0;
730 if (!sce->zeroes[w*16+g])
731 prev_sf = sce->sf_idx[w*16+g];
737 static void mark_pns(AACEncContext *s, AVCodecContext *avctx, SingleChannelElement *sce)
741 int wlen = 1024 / sce->ics.num_windows;
742 int bandwidth, cutoff;
743 const float lambda = s->lambda;
744 const float freq_mult = avctx->sample_rate*0.5f/wlen;
745 const float spread_threshold = FFMIN(0.75f, NOISE_SPREAD_THRESHOLD*FFMAX(0.5f, lambda/100.f));
746 const float pns_transient_energy_r = FFMIN(0.7f, lambda / 140.f);
748 int refbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
749 / ((avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : avctx->channels)
752 /** Keep this in sync with twoloop's cutoff selection */
753 float rate_bandwidth_multiplier = 1.5f;
754 int frame_bit_rate = (avctx->flags & CODEC_FLAG_QSCALE)
755 ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
756 : (avctx->bit_rate / avctx->channels);
758 frame_bit_rate *= 1.15f;
760 if (avctx->cutoff > 0) {
761 bandwidth = avctx->cutoff;
763 bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
766 cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
768 memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type));
769 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
770 for (g = 0; g < sce->ics.num_swb; g++) {
771 float sfb_energy = 0.0f, threshold = 0.0f, spread = 2.0f;
772 float min_energy = -1.0f, max_energy = 0.0f;
773 const int start = sce->ics.swb_offset[g];
774 const float freq = start*freq_mult;
775 const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
776 if (freq < NOISE_LOW_LIMIT || start >= cutoff) {
777 sce->can_pns[w*16+g] = 0;
780 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
781 band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
782 sfb_energy += band->energy;
783 spread = FFMIN(spread, band->spread);
784 threshold += band->threshold;
786 min_energy = max_energy = band->energy;
788 min_energy = FFMIN(min_energy, band->energy);
789 max_energy = FFMAX(max_energy, band->energy);
793 /* PNS is acceptable when all of these are true:
794 * 1. high spread energy (noise-like band)
795 * 2. near-threshold energy (high PE means the random nature of PNS content will be noticed)
796 * 3. on short window groups, all windows have similar energy (variations in energy would be destroyed by PNS)
798 sce->pns_ener[w*16+g] = sfb_energy;
799 if (sfb_energy < threshold*sqrtf(1.5f/freq_boost) || spread < spread_threshold || min_energy < pns_transient_energy_r * max_energy) {
800 sce->can_pns[w*16+g] = 0;
802 sce->can_pns[w*16+g] = 1;
808 static void search_for_ms(AACEncContext *s, ChannelElement *cpe)
810 int start = 0, i, w, w2, g, sid_sf_boost, prev_mid, prev_side;
811 uint8_t nextband0[128], nextband1[128];
812 float M[128], S[128];
813 float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
814 const float lambda = s->lambda;
815 const float mslambda = FFMIN(1.0f, lambda / 120.f);
816 SingleChannelElement *sce0 = &cpe->ch[0];
817 SingleChannelElement *sce1 = &cpe->ch[1];
818 if (!cpe->common_window)
821 /** Scout out next nonzero bands */
822 ff_init_nextband_map(sce0, nextband0);
823 ff_init_nextband_map(sce1, nextband1);
825 prev_mid = sce0->sf_idx[0];
826 prev_side = sce1->sf_idx[0];
827 for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
829 for (g = 0; g < sce0->ics.num_swb; g++) {
830 float bmax = bval2bmax(g * 17.0f / sce0->ics.num_swb) / 0.0045f;
831 cpe->ms_mask[w*16+g] = 0;
832 if (!sce0->zeroes[w*16+g] && !sce1->zeroes[w*16+g]) {
833 float Mmax = 0.0f, Smax = 0.0f;
835 /* Must compute mid/side SF and book for the whole window group */
836 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
837 for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
838 M[i] = (sce0->coeffs[start+(w+w2)*128+i]
839 + sce1->coeffs[start+(w+w2)*128+i]) * 0.5;
841 - sce1->coeffs[start+(w+w2)*128+i];
843 abs_pow34_v(M34, M, sce0->ics.swb_sizes[g]);
844 abs_pow34_v(S34, S, sce0->ics.swb_sizes[g]);
845 for (i = 0; i < sce0->ics.swb_sizes[g]; i++ ) {
846 Mmax = FFMAX(Mmax, M34[i]);
847 Smax = FFMAX(Smax, S34[i]);
851 for (sid_sf_boost = 0; sid_sf_boost < 4; sid_sf_boost++) {
852 float dist1 = 0.0f, dist2 = 0.0f;
858 minidx = FFMIN(sce0->sf_idx[w*16+g], sce1->sf_idx[w*16+g]);
859 mididx = av_clip(minidx, 0, SCALE_MAX_POS - SCALE_DIV_512);
860 sididx = av_clip(minidx - sid_sf_boost * 3, 0, SCALE_MAX_POS - SCALE_DIV_512);
861 if (!cpe->is_mask[w*16+g] && sce0->band_type[w*16+g] != NOISE_BT && sce1->band_type[w*16+g] != NOISE_BT
862 && ( !ff_sfdelta_can_replace(sce0, nextband0, prev_mid, mididx, w*16+g)
863 || !ff_sfdelta_can_replace(sce1, nextband1, prev_side, sididx, w*16+g))) {
864 /* scalefactor range violation, bad stuff, will decrease quality unacceptably */
868 midcb = find_min_book(Mmax, mididx);
869 sidcb = find_min_book(Smax, sididx);
871 /* No CB can be zero */
872 midcb = FFMAX(1,midcb);
873 sidcb = FFMAX(1,sidcb);
875 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
876 FFPsyBand *band0 = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g];
877 FFPsyBand *band1 = &s->psy.ch[s->cur_channel+1].psy_bands[(w+w2)*16+g];
878 float minthr = FFMIN(band0->threshold, band1->threshold);
880 for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
881 M[i] = (sce0->coeffs[start+(w+w2)*128+i]
882 + sce1->coeffs[start+(w+w2)*128+i]) * 0.5;
884 - sce1->coeffs[start+(w+w2)*128+i];
887 abs_pow34_v(L34, sce0->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
888 abs_pow34_v(R34, sce1->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
889 abs_pow34_v(M34, M, sce0->ics.swb_sizes[g]);
890 abs_pow34_v(S34, S, sce0->ics.swb_sizes[g]);
891 dist1 += quantize_band_cost(s, &sce0->coeffs[start + (w+w2)*128],
893 sce0->ics.swb_sizes[g],
894 sce0->sf_idx[(w+w2)*16+g],
895 sce0->band_type[(w+w2)*16+g],
896 lambda / band0->threshold, INFINITY, &b1, NULL, 0);
897 dist1 += quantize_band_cost(s, &sce1->coeffs[start + (w+w2)*128],
899 sce1->ics.swb_sizes[g],
900 sce1->sf_idx[(w+w2)*16+g],
901 sce1->band_type[(w+w2)*16+g],
902 lambda / band1->threshold, INFINITY, &b2, NULL, 0);
903 dist2 += quantize_band_cost(s, M,
905 sce0->ics.swb_sizes[g],
906 sce0->sf_idx[(w+w2)*16+g],
907 sce0->band_type[(w+w2)*16+g],
908 lambda / minthr, INFINITY, &b3, NULL, 0);
909 dist2 += quantize_band_cost(s, S,
911 sce1->ics.swb_sizes[g],
912 sce1->sf_idx[(w+w2)*16+g],
913 sce1->band_type[(w+w2)*16+g],
914 mslambda / (minthr * bmax), INFINITY, &b4, NULL, 0);
920 cpe->ms_mask[w*16+g] = dist2 <= dist1 && B1 < B0;
921 if (cpe->ms_mask[w*16+g]) {
922 /* Setting the M/S mask is useful with I/S or PNS, but only the flag */
923 if (!cpe->is_mask[w*16+g] && sce0->band_type[w*16+g] != NOISE_BT && sce1->band_type[w*16+g] != NOISE_BT) {
924 sce0->sf_idx[w*16+g] = mididx;
925 sce1->sf_idx[w*16+g] = sididx;
926 sce0->band_type[w*16+g] = midcb;
927 sce1->band_type[w*16+g] = sidcb;
930 } else if (B1 > B0) {
931 /* More boost won't fix this */
936 if (!sce0->zeroes[w*16+g] && sce0->band_type[w*16+g] < RESERVED_BT)
937 prev_mid = sce0->sf_idx[w*16+g];
938 if (!sce1->zeroes[w*16+g] && !cpe->is_mask[w*16+g] && sce1->band_type[w*16+g] < RESERVED_BT)
939 prev_side = sce1->sf_idx[w*16+g];
940 start += sce0->ics.swb_sizes[g];
945 AACCoefficientsEncoder ff_aac_coders[AAC_CODER_NB] = {
947 search_for_quantizers_faac,
948 encode_window_bands_info,
949 quantize_and_encode_band,
950 ff_aac_encode_tns_info,
951 ff_aac_encode_ltp_info,
952 ff_aac_encode_main_pred,
953 ff_aac_adjust_common_pred,
954 ff_aac_adjust_common_ltp,
955 ff_aac_apply_main_pred,
958 ff_aac_ltp_insert_new_frame,
959 set_special_band_scalefactors,
962 ff_aac_search_for_tns,
963 ff_aac_search_for_ltp,
965 ff_aac_search_for_is,
966 ff_aac_search_for_pred,
969 search_for_quantizers_anmr,
970 encode_window_bands_info,
971 quantize_and_encode_band,
972 ff_aac_encode_tns_info,
973 ff_aac_encode_ltp_info,
974 ff_aac_encode_main_pred,
975 ff_aac_adjust_common_pred,
976 ff_aac_adjust_common_ltp,
977 ff_aac_apply_main_pred,
980 ff_aac_ltp_insert_new_frame,
981 set_special_band_scalefactors,
984 ff_aac_search_for_tns,
985 ff_aac_search_for_ltp,
987 ff_aac_search_for_is,
988 ff_aac_search_for_pred,
990 [AAC_CODER_TWOLOOP] = {
991 search_for_quantizers_twoloop,
992 codebook_trellis_rate,
993 quantize_and_encode_band,
994 ff_aac_encode_tns_info,
995 ff_aac_encode_ltp_info,
996 ff_aac_encode_main_pred,
997 ff_aac_adjust_common_pred,
998 ff_aac_adjust_common_ltp,
999 ff_aac_apply_main_pred,
1002 ff_aac_ltp_insert_new_frame,
1003 set_special_band_scalefactors,
1006 ff_aac_search_for_tns,
1007 ff_aac_search_for_ltp,
1009 ff_aac_search_for_is,
1010 ff_aac_search_for_pred,
1012 [AAC_CODER_FAST] = {
1013 search_for_quantizers_fast,
1014 encode_window_bands_info,
1015 quantize_and_encode_band,
1016 ff_aac_encode_tns_info,
1017 ff_aac_encode_ltp_info,
1018 ff_aac_encode_main_pred,
1019 ff_aac_adjust_common_pred,
1020 ff_aac_adjust_common_ltp,
1021 ff_aac_apply_main_pred,
1024 ff_aac_ltp_insert_new_frame,
1025 set_special_band_scalefactors,
1028 ff_aac_search_for_tns,
1029 ff_aac_search_for_ltp,
1031 ff_aac_search_for_is,
1032 ff_aac_search_for_pred,