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 prevscaler_n = -255, prevscaler_i = 0;
202 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
203 for (g = 0; g < sce->ics.num_swb; g++) {
204 if (sce->zeroes[w*16+g])
206 if (sce->band_type[w*16+g] == INTENSITY_BT || sce->band_type[w*16+g] == INTENSITY_BT2) {
207 sce->sf_idx[w*16+g] = av_clip(roundf(log2f(sce->is_ener[w*16+g])*2), -155, 100);
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 if (prevscaler_n == -255)
212 prevscaler_n = sce->sf_idx[w*16+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->zeroes[w*16+g])
226 if (sce->band_type[w*16+g] == INTENSITY_BT || sce->band_type[w*16+g] == INTENSITY_BT2) {
227 sce->sf_idx[w*16+g] = prevscaler_i = av_clip(sce->sf_idx[w*16+g], prevscaler_i - SCALE_MAX_DIFF, prevscaler_i + SCALE_MAX_DIFF);
228 } else if (sce->band_type[w*16+g] == NOISE_BT) {
229 sce->sf_idx[w*16+g] = prevscaler_n = av_clip(sce->sf_idx[w*16+g], prevscaler_n - SCALE_MAX_DIFF, prevscaler_n + SCALE_MAX_DIFF);
235 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
236 SingleChannelElement *sce,
239 int q, w, w2, g, start = 0;
242 TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
243 int bandaddr[TRELLIS_STAGES];
246 float q0f = FLT_MAX, q1f = 0.0f, qnrgf = 0.0f;
247 int q0, q1, qcnt = 0;
249 for (i = 0; i < 1024; i++) {
250 float t = fabsf(sce->coeffs[i]);
260 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
261 memset(sce->zeroes, 1, sizeof(sce->zeroes));
265 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
266 q0 = av_clip(coef2minsf(q0f), 0, SCALE_MAX_POS-1);
267 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
268 q1 = av_clip(coef2maxsf(q1f), 1, SCALE_MAX_POS);
272 //minimum scalefactor index is when maximum nonzero coefficient after quantizing is not clipped
273 int qnrg = av_clip_uint8(log2f(sqrtf(qnrgf/qcnt))*4 - 31 + SCALE_ONE_POS - SCALE_DIV_512);
279 } else if (q1 > q1high) {
284 // q0 == q1 isn't really a legal situation
286 // the following is indirect but guarantees q1 != q0 && q1 near q0
287 q1 = av_clip(q0+1, 1, SCALE_MAX_POS);
288 q0 = av_clip(q1-1, 0, SCALE_MAX_POS - 1);
291 for (i = 0; i < TRELLIS_STATES; i++) {
292 paths[0][i].cost = 0.0f;
293 paths[0][i].prev = -1;
295 for (j = 1; j < TRELLIS_STAGES; j++) {
296 for (i = 0; i < TRELLIS_STATES; i++) {
297 paths[j][i].cost = INFINITY;
298 paths[j][i].prev = -2;
302 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
303 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
305 for (g = 0; g < sce->ics.num_swb; g++) {
306 const float *coefs = &sce->coeffs[start];
310 bandaddr[idx] = w * 16 + g;
313 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
314 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
315 if (band->energy <= band->threshold || band->threshold == 0.0f) {
316 sce->zeroes[(w+w2)*16+g] = 1;
319 sce->zeroes[(w+w2)*16+g] = 0;
321 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
322 float t = fabsf(coefs[w2*128+i]);
324 qmin = FFMIN(qmin, t);
325 qmax = FFMAX(qmax, t);
329 int minscale, maxscale;
330 float minrd = INFINITY;
332 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
333 minscale = coef2minsf(qmin);
334 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
335 maxscale = coef2maxsf(qmax);
336 minscale = av_clip(minscale - q0, 0, TRELLIS_STATES - 1);
337 maxscale = av_clip(maxscale - q0, 0, TRELLIS_STATES);
338 if (minscale == maxscale) {
339 maxscale = av_clip(minscale+1, 1, TRELLIS_STATES);
340 minscale = av_clip(maxscale-1, 0, TRELLIS_STATES - 1);
342 maxval = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], s->scoefs+start);
343 for (q = minscale; q < maxscale; q++) {
345 int cb = find_min_book(maxval, sce->sf_idx[w*16+g]);
346 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
347 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
348 dist += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
349 q + q0, cb, lambda / band->threshold, INFINITY, NULL, NULL, 0);
351 minrd = FFMIN(minrd, dist);
353 for (i = 0; i < q1 - q0; i++) {
355 cost = paths[idx - 1][i].cost + dist
356 + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
357 if (cost < paths[idx][q].cost) {
358 paths[idx][q].cost = cost;
359 paths[idx][q].prev = i;
364 for (q = 0; q < q1 - q0; q++) {
365 paths[idx][q].cost = paths[idx - 1][q].cost + 1;
366 paths[idx][q].prev = q;
369 sce->zeroes[w*16+g] = !nz;
370 start += sce->ics.swb_sizes[g];
375 mincost = paths[idx][0].cost;
377 for (i = 1; i < TRELLIS_STATES; i++) {
378 if (paths[idx][i].cost < mincost) {
379 mincost = paths[idx][i].cost;
384 sce->sf_idx[bandaddr[idx]] = minq + q0;
385 minq = FFMAX(paths[idx][minq].prev, 0);
388 //set the same quantizers inside window groups
389 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
390 for (g = 0; g < sce->ics.num_swb; g++)
391 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
392 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
396 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
397 SingleChannelElement *sce,
400 int start = 0, i, w, w2, g;
401 float uplim[128], maxq[128];
403 float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
404 int last = 0, lastband = 0, curband = 0;
405 float avg_energy = 0.0;
406 if (sce->ics.num_windows == 1) {
408 for (i = 0; i < 1024; i++) {
409 if (i - start >= sce->ics.swb_sizes[curband]) {
410 start += sce->ics.swb_sizes[curband];
413 if (sce->coeffs[i]) {
414 avg_energy += sce->coeffs[i] * sce->coeffs[i];
420 for (w = 0; w < 8; w++) {
421 const float *coeffs = &sce->coeffs[w*128];
423 for (i = 0; i < 128; i++) {
424 if (i - start >= sce->ics.swb_sizes[curband]) {
425 start += sce->ics.swb_sizes[curband];
429 avg_energy += coeffs[i] * coeffs[i];
430 last = FFMAX(last, i);
431 lastband = FFMAX(lastband, curband);
438 if (avg_energy == 0.0f) {
439 for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
440 sce->sf_idx[i] = SCALE_ONE_POS;
443 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
445 for (g = 0; g < sce->ics.num_swb; g++) {
446 float *coefs = &sce->coeffs[start];
447 const int size = sce->ics.swb_sizes[g];
448 int start2 = start, end2 = start + size, peakpos = start;
449 float maxval = -1, thr = 0.0f, t;
454 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
455 memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
458 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
459 for (i = 0; i < size; i++) {
460 float t = coefs[w2*128+i]*coefs[w2*128+i];
461 maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
463 if (sce->ics.num_windows == 1 && maxval < t) {
469 if (sce->ics.num_windows == 1) {
470 start2 = FFMAX(peakpos - 2, start2);
471 end2 = FFMIN(peakpos + 3, end2);
477 thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
478 t = 1.0 - (1.0 * start2 / last);
479 uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
482 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
483 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
484 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
486 for (g = 0; g < sce->ics.num_swb; g++) {
487 const float *coefs = &sce->coeffs[start];
488 const float *scaled = &s->scoefs[start];
489 const int size = sce->ics.swb_sizes[g];
490 int scf, prev_scf, step;
491 int min_scf = -1, max_scf = 256;
493 if (maxq[w*16+g] < 21.544) {
494 sce->zeroes[w*16+g] = 1;
498 sce->zeroes[w*16+g] = 0;
499 scf = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2f(1/maxq[w*16+g])*16/3, 60, 218);
504 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
506 dist += quantize_band_cost(s, coefs + w2*128,
508 sce->ics.swb_sizes[g],
517 dist *= 1.0f / 512.0f / lambda;
518 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[POW_SF2_ZERO - scf + SCALE_ONE_POS - SCALE_DIV_512], ROUND_STANDARD);
519 if (quant_max >= 8191) { // too much, return to the previous quantizer
520 sce->sf_idx[w*16+g] = prev_scf;
524 curdiff = fabsf(dist - uplim[w*16+g]);
528 step = log2f(curdiff);
529 if (dist > uplim[w*16+g])
532 scf = av_clip_uint8(scf);
533 step = scf - prev_scf;
534 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
535 sce->sf_idx[w*16+g] = av_clip(scf, min_scf, max_scf);
546 minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
547 for (i = 1; i < 128; i++) {
549 sce->sf_idx[i] = sce->sf_idx[i-1];
551 minq = FFMIN(minq, sce->sf_idx[i]);
555 minq = FFMIN(minq, SCALE_MAX_POS);
556 maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
557 for (i = 126; i >= 0; i--) {
559 sce->sf_idx[i] = sce->sf_idx[i+1];
560 sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
564 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
565 SingleChannelElement *sce,
571 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
572 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
573 for (g = 0; g < sce->ics.num_swb; g++) {
574 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
575 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
576 if (band->energy <= band->threshold) {
577 sce->sf_idx[(w+w2)*16+g] = 218;
578 sce->zeroes[(w+w2)*16+g] = 1;
580 sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2f(band->threshold), 80, 218);
581 sce->zeroes[(w+w2)*16+g] = 0;
583 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
587 for (i = 0; i < 128; i++) {
588 sce->sf_idx[i] = 140;
589 //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
591 //set the same quantizers inside window groups
592 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
593 for (g = 0; g < sce->ics.num_swb; g++)
594 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
595 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
598 static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChannelElement *sce)
602 int wlen = 1024 / sce->ics.num_windows;
603 int bandwidth, cutoff;
604 float *PNS = &s->scoefs[0*128], *PNS34 = &s->scoefs[1*128];
605 float *NOR34 = &s->scoefs[3*128];
606 uint8_t nextband[128];
607 const float lambda = s->lambda;
608 const float freq_mult = avctx->sample_rate*0.5f/wlen;
609 const float thr_mult = NOISE_LAMBDA_REPLACE*(100.0f/lambda);
610 const float spread_threshold = FFMIN(0.75f, NOISE_SPREAD_THRESHOLD*FFMAX(0.5f, lambda/100.f));
611 const float dist_bias = av_clipf(4.f * 120 / lambda, 0.25f, 4.0f);
612 const float pns_transient_energy_r = FFMIN(0.7f, lambda / 140.f);
614 int refbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
615 / ((avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : avctx->channels)
618 /** Keep this in sync with twoloop's cutoff selection */
619 float rate_bandwidth_multiplier = 1.5f;
620 int prev = -1000, prev_sf = -1;
621 int frame_bit_rate = (avctx->flags & CODEC_FLAG_QSCALE)
622 ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
623 : (avctx->bit_rate / avctx->channels);
625 frame_bit_rate *= 1.15f;
627 if (avctx->cutoff > 0) {
628 bandwidth = avctx->cutoff;
630 bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
633 cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
635 memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type));
636 ff_init_nextband_map(sce, nextband);
637 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
639 for (g = 0; g < sce->ics.num_swb; g++) {
641 float dist1 = 0.0f, dist2 = 0.0f, noise_amp;
642 float pns_energy = 0.0f, pns_tgt_energy, energy_ratio, dist_thresh;
643 float sfb_energy = 0.0f, threshold = 0.0f, spread = 2.0f;
644 float min_energy = -1.0f, max_energy = 0.0f;
645 const int start = wstart+sce->ics.swb_offset[g];
646 const float freq = (start-wstart)*freq_mult;
647 const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
648 if (freq < NOISE_LOW_LIMIT || (start-wstart) >= cutoff) {
649 if (!sce->zeroes[w*16+g])
650 prev_sf = sce->sf_idx[w*16+g];
653 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
654 band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
655 sfb_energy += band->energy;
656 spread = FFMIN(spread, band->spread);
657 threshold += band->threshold;
659 min_energy = max_energy = band->energy;
661 min_energy = FFMIN(min_energy, band->energy);
662 max_energy = FFMAX(max_energy, band->energy);
666 /* Ramps down at ~8000Hz and loosens the dist threshold */
667 dist_thresh = av_clipf(2.5f*NOISE_LOW_LIMIT/freq, 0.5f, 2.5f) * dist_bias;
669 /* PNS is acceptable when all of these are true:
670 * 1. high spread energy (noise-like band)
671 * 2. near-threshold energy (high PE means the random nature of PNS content will be noticed)
672 * 3. on short window groups, all windows have similar energy (variations in energy would be destroyed by PNS)
674 * At this stage, point 2 is relaxed for zeroed bands near the noise threshold (hole avoidance is more important)
676 if ((!sce->zeroes[w*16+g] && !ff_sfdelta_can_remove_band(sce, nextband, prev_sf, w*16+g)) ||
677 ((sce->zeroes[w*16+g] || !sce->band_alt[w*16+g]) && sfb_energy < threshold*sqrtf(1.0f/freq_boost)) || spread < spread_threshold ||
678 (!sce->zeroes[w*16+g] && sce->band_alt[w*16+g] && sfb_energy > threshold*thr_mult*freq_boost) ||
679 min_energy < pns_transient_energy_r * max_energy ) {
680 sce->pns_ener[w*16+g] = sfb_energy;
681 if (!sce->zeroes[w*16+g])
682 prev_sf = sce->sf_idx[w*16+g];
686 pns_tgt_energy = sfb_energy*FFMIN(1.0f, spread*spread);
687 noise_sfi = av_clip(roundf(log2f(pns_tgt_energy)*2), -100, 155); /* Quantize */
688 noise_amp = -ff_aac_pow2sf_tab[noise_sfi + POW_SF2_ZERO]; /* Dequantize */
690 int noise_sfdiff = noise_sfi - prev + SCALE_DIFF_ZERO;
691 if (noise_sfdiff < 0 || noise_sfdiff > 2*SCALE_MAX_DIFF) {
692 if (!sce->zeroes[w*16+g])
693 prev_sf = sce->sf_idx[w*16+g];
697 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
698 float band_energy, scale, pns_senergy;
699 const int start_c = (w+w2)*128+sce->ics.swb_offset[g];
700 band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
701 for (i = 0; i < sce->ics.swb_sizes[g]; i+=2) {
703 av_bmg_get(&s->lfg, rnd);
704 PNS[i+0] = (float)rnd[0];
705 PNS[i+1] = (float)rnd[1];
707 band_energy = s->fdsp->scalarproduct_float(PNS, PNS, sce->ics.swb_sizes[g]);
708 scale = noise_amp/sqrtf(band_energy);
709 s->fdsp->vector_fmul_scalar(PNS, PNS, scale, sce->ics.swb_sizes[g]);
710 pns_senergy = s->fdsp->scalarproduct_float(PNS, PNS, sce->ics.swb_sizes[g]);
711 pns_energy += pns_senergy;
712 abs_pow34_v(NOR34, &sce->coeffs[start_c], sce->ics.swb_sizes[g]);
713 abs_pow34_v(PNS34, PNS, sce->ics.swb_sizes[g]);
714 dist1 += quantize_band_cost(s, &sce->coeffs[start_c],
716 sce->ics.swb_sizes[g],
717 sce->sf_idx[(w+w2)*16+g],
718 sce->band_alt[(w+w2)*16+g],
719 lambda/band->threshold, INFINITY, NULL, NULL, 0);
720 /* Estimate rd on average as 5 bits for SF, 4 for the CB, plus spread energy * lambda/thr */
721 dist2 += band->energy/(band->spread*band->spread)*lambda*dist_thresh/band->threshold;
723 if (g && sce->band_type[w*16+g-1] == NOISE_BT) {
728 energy_ratio = pns_tgt_energy/pns_energy; /* Compensates for quantization error */
729 sce->pns_ener[w*16+g] = energy_ratio*pns_tgt_energy;
730 if (sce->zeroes[w*16+g] || !sce->band_alt[w*16+g] || (energy_ratio > 0.85f && energy_ratio < 1.25f && dist2 < dist1)) {
731 sce->band_type[w*16+g] = NOISE_BT;
732 sce->zeroes[w*16+g] = 0;
735 if (!sce->zeroes[w*16+g])
736 prev_sf = sce->sf_idx[w*16+g];
742 static void mark_pns(AACEncContext *s, AVCodecContext *avctx, SingleChannelElement *sce)
746 int wlen = 1024 / sce->ics.num_windows;
747 int bandwidth, cutoff;
748 const float lambda = s->lambda;
749 const float freq_mult = avctx->sample_rate*0.5f/wlen;
750 const float spread_threshold = FFMIN(0.75f, NOISE_SPREAD_THRESHOLD*FFMAX(0.5f, lambda/100.f));
751 const float pns_transient_energy_r = FFMIN(0.7f, lambda / 140.f);
753 int refbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
754 / ((avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : avctx->channels)
757 /** Keep this in sync with twoloop's cutoff selection */
758 float rate_bandwidth_multiplier = 1.5f;
759 int frame_bit_rate = (avctx->flags & CODEC_FLAG_QSCALE)
760 ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
761 : (avctx->bit_rate / avctx->channels);
763 frame_bit_rate *= 1.15f;
765 if (avctx->cutoff > 0) {
766 bandwidth = avctx->cutoff;
768 bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
771 cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
773 memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type));
774 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
775 for (g = 0; g < sce->ics.num_swb; g++) {
776 float sfb_energy = 0.0f, threshold = 0.0f, spread = 2.0f;
777 float min_energy = -1.0f, max_energy = 0.0f;
778 const int start = sce->ics.swb_offset[g];
779 const float freq = start*freq_mult;
780 const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
781 if (freq < NOISE_LOW_LIMIT || start >= cutoff) {
782 sce->can_pns[w*16+g] = 0;
785 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
786 band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
787 sfb_energy += band->energy;
788 spread = FFMIN(spread, band->spread);
789 threshold += band->threshold;
791 min_energy = max_energy = band->energy;
793 min_energy = FFMIN(min_energy, band->energy);
794 max_energy = FFMAX(max_energy, band->energy);
798 /* PNS is acceptable when all of these are true:
799 * 1. high spread energy (noise-like band)
800 * 2. near-threshold energy (high PE means the random nature of PNS content will be noticed)
801 * 3. on short window groups, all windows have similar energy (variations in energy would be destroyed by PNS)
803 sce->pns_ener[w*16+g] = sfb_energy;
804 if (sfb_energy < threshold*sqrtf(1.5f/freq_boost) || spread < spread_threshold || min_energy < pns_transient_energy_r * max_energy) {
805 sce->can_pns[w*16+g] = 0;
807 sce->can_pns[w*16+g] = 1;
813 static void search_for_ms(AACEncContext *s, ChannelElement *cpe)
815 int start = 0, i, w, w2, g, sid_sf_boost, prev_mid, prev_side;
816 uint8_t nextband0[128], nextband1[128];
817 float M[128], S[128];
818 float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
819 const float lambda = s->lambda;
820 const float mslambda = FFMIN(1.0f, lambda / 120.f);
821 SingleChannelElement *sce0 = &cpe->ch[0];
822 SingleChannelElement *sce1 = &cpe->ch[1];
823 if (!cpe->common_window)
826 /** Scout out next nonzero bands */
827 ff_init_nextband_map(sce0, nextband0);
828 ff_init_nextband_map(sce1, nextband1);
830 prev_mid = sce0->sf_idx[0];
831 prev_side = sce1->sf_idx[0];
832 for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
834 for (g = 0; g < sce0->ics.num_swb; g++) {
835 float bmax = bval2bmax(g * 17.0f / sce0->ics.num_swb) / 0.0045f;
836 if (!cpe->is_mask[w*16+g])
837 cpe->ms_mask[w*16+g] = 0;
838 if (!sce0->zeroes[w*16+g] && !sce1->zeroes[w*16+g] && !cpe->is_mask[w*16+g]) {
839 float Mmax = 0.0f, Smax = 0.0f;
841 /* Must compute mid/side SF and book for the whole window group */
842 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
843 for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
844 M[i] = (sce0->coeffs[start+(w+w2)*128+i]
845 + sce1->coeffs[start+(w+w2)*128+i]) * 0.5;
847 - sce1->coeffs[start+(w+w2)*128+i];
849 abs_pow34_v(M34, M, sce0->ics.swb_sizes[g]);
850 abs_pow34_v(S34, S, sce0->ics.swb_sizes[g]);
851 for (i = 0; i < sce0->ics.swb_sizes[g]; i++ ) {
852 Mmax = FFMAX(Mmax, M34[i]);
853 Smax = FFMAX(Smax, S34[i]);
857 for (sid_sf_boost = 0; sid_sf_boost < 4; sid_sf_boost++) {
858 float dist1 = 0.0f, dist2 = 0.0f;
864 minidx = FFMIN(sce0->sf_idx[w*16+g], sce1->sf_idx[w*16+g]);
865 mididx = av_clip(minidx, 0, SCALE_MAX_POS - SCALE_DIV_512);
866 sididx = av_clip(minidx - sid_sf_boost * 3, 0, SCALE_MAX_POS - SCALE_DIV_512);
867 if (sce0->band_type[w*16+g] != NOISE_BT && sce1->band_type[w*16+g] != NOISE_BT
868 && ( !ff_sfdelta_can_replace(sce0, nextband0, prev_mid, mididx, w*16+g)
869 || !ff_sfdelta_can_replace(sce1, nextband1, prev_side, sididx, w*16+g))) {
870 /* scalefactor range violation, bad stuff, will decrease quality unacceptably */
874 midcb = find_min_book(Mmax, mididx);
875 sidcb = find_min_book(Smax, sididx);
877 /* No CB can be zero */
878 midcb = FFMAX(1,midcb);
879 sidcb = FFMAX(1,sidcb);
881 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
882 FFPsyBand *band0 = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g];
883 FFPsyBand *band1 = &s->psy.ch[s->cur_channel+1].psy_bands[(w+w2)*16+g];
884 float minthr = FFMIN(band0->threshold, band1->threshold);
886 for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
887 M[i] = (sce0->coeffs[start+(w+w2)*128+i]
888 + sce1->coeffs[start+(w+w2)*128+i]) * 0.5;
890 - sce1->coeffs[start+(w+w2)*128+i];
893 abs_pow34_v(L34, sce0->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
894 abs_pow34_v(R34, sce1->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
895 abs_pow34_v(M34, M, sce0->ics.swb_sizes[g]);
896 abs_pow34_v(S34, S, sce0->ics.swb_sizes[g]);
897 dist1 += quantize_band_cost(s, &sce0->coeffs[start + (w+w2)*128],
899 sce0->ics.swb_sizes[g],
900 sce0->sf_idx[w*16+g],
901 sce0->band_type[w*16+g],
902 lambda / band0->threshold, INFINITY, &b1, NULL, 0);
903 dist1 += quantize_band_cost(s, &sce1->coeffs[start + (w+w2)*128],
905 sce1->ics.swb_sizes[g],
906 sce1->sf_idx[w*16+g],
907 sce1->band_type[w*16+g],
908 lambda / band1->threshold, INFINITY, &b2, NULL, 0);
909 dist2 += quantize_band_cost(s, M,
911 sce0->ics.swb_sizes[g],
914 lambda / minthr, INFINITY, &b3, NULL, 0);
915 dist2 += quantize_band_cost(s, S,
917 sce1->ics.swb_sizes[g],
920 mslambda / (minthr * bmax), INFINITY, &b4, NULL, 0);
926 cpe->ms_mask[w*16+g] = dist2 <= dist1 && B1 < B0;
927 if (cpe->ms_mask[w*16+g]) {
928 if (sce0->band_type[w*16+g] != NOISE_BT && sce1->band_type[w*16+g] != NOISE_BT) {
929 sce0->sf_idx[w*16+g] = mididx;
930 sce1->sf_idx[w*16+g] = sididx;
931 sce0->band_type[w*16+g] = midcb;
932 sce1->band_type[w*16+g] = sidcb;
933 } else if ((sce0->band_type[w*16+g] != NOISE_BT) ^ (sce1->band_type[w*16+g] != NOISE_BT)) {
934 /* ms_mask unneeded, and it confuses some decoders */
935 cpe->ms_mask[w*16+g] = 0;
938 } else if (B1 > B0) {
939 /* More boost won't fix this */
944 if (!sce0->zeroes[w*16+g] && sce0->band_type[w*16+g] < RESERVED_BT)
945 prev_mid = sce0->sf_idx[w*16+g];
946 if (!sce1->zeroes[w*16+g] && !cpe->is_mask[w*16+g] && sce1->band_type[w*16+g] < RESERVED_BT)
947 prev_side = sce1->sf_idx[w*16+g];
948 start += sce0->ics.swb_sizes[g];
953 AACCoefficientsEncoder ff_aac_coders[AAC_CODER_NB] = {
955 search_for_quantizers_faac,
956 encode_window_bands_info,
957 quantize_and_encode_band,
958 ff_aac_encode_tns_info,
959 ff_aac_encode_ltp_info,
960 ff_aac_encode_main_pred,
961 ff_aac_adjust_common_pred,
962 ff_aac_adjust_common_ltp,
963 ff_aac_apply_main_pred,
966 ff_aac_ltp_insert_new_frame,
967 set_special_band_scalefactors,
970 ff_aac_search_for_tns,
971 ff_aac_search_for_ltp,
973 ff_aac_search_for_is,
974 ff_aac_search_for_pred,
977 search_for_quantizers_anmr,
978 encode_window_bands_info,
979 quantize_and_encode_band,
980 ff_aac_encode_tns_info,
981 ff_aac_encode_ltp_info,
982 ff_aac_encode_main_pred,
983 ff_aac_adjust_common_pred,
984 ff_aac_adjust_common_ltp,
985 ff_aac_apply_main_pred,
988 ff_aac_ltp_insert_new_frame,
989 set_special_band_scalefactors,
992 ff_aac_search_for_tns,
993 ff_aac_search_for_ltp,
995 ff_aac_search_for_is,
996 ff_aac_search_for_pred,
998 [AAC_CODER_TWOLOOP] = {
999 search_for_quantizers_twoloop,
1000 codebook_trellis_rate,
1001 quantize_and_encode_band,
1002 ff_aac_encode_tns_info,
1003 ff_aac_encode_ltp_info,
1004 ff_aac_encode_main_pred,
1005 ff_aac_adjust_common_pred,
1006 ff_aac_adjust_common_ltp,
1007 ff_aac_apply_main_pred,
1010 ff_aac_ltp_insert_new_frame,
1011 set_special_band_scalefactors,
1014 ff_aac_search_for_tns,
1015 ff_aac_search_for_ltp,
1017 ff_aac_search_for_is,
1018 ff_aac_search_for_pred,
1020 [AAC_CODER_FAST] = {
1021 search_for_quantizers_fast,
1022 encode_window_bands_info,
1023 quantize_and_encode_band,
1024 ff_aac_encode_tns_info,
1025 ff_aac_encode_ltp_info,
1026 ff_aac_encode_main_pred,
1027 ff_aac_adjust_common_pred,
1028 ff_aac_adjust_common_ltp,
1029 ff_aac_apply_main_pred,
1032 ff_aac_ltp_insert_new_frame,
1033 set_special_band_scalefactors,
1036 ff_aac_search_for_tns,
1037 ff_aac_search_for_ltp,
1039 ff_aac_search_for_is,
1040 ff_aac_search_for_pred,