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 0.5073f
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 = coef2minsf(q0f);
265 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
266 q1 = coef2maxsf(q1f);
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) {
283 for (i = 0; i < TRELLIS_STATES; i++) {
284 paths[0][i].cost = 0.0f;
285 paths[0][i].prev = -1;
287 for (j = 1; j < TRELLIS_STAGES; j++) {
288 for (i = 0; i < TRELLIS_STATES; i++) {
289 paths[j][i].cost = INFINITY;
290 paths[j][i].prev = -2;
294 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
295 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
297 for (g = 0; g < sce->ics.num_swb; g++) {
298 const float *coefs = &sce->coeffs[start];
302 bandaddr[idx] = w * 16 + g;
305 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
306 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
307 if (band->energy <= band->threshold || band->threshold == 0.0f) {
308 sce->zeroes[(w+w2)*16+g] = 1;
311 sce->zeroes[(w+w2)*16+g] = 0;
313 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
314 float t = fabsf(coefs[w2*128+i]);
316 qmin = FFMIN(qmin, t);
317 qmax = FFMAX(qmax, t);
321 int minscale, maxscale;
322 float minrd = INFINITY;
324 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
325 minscale = coef2minsf(qmin);
326 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
327 maxscale = coef2maxsf(qmax);
328 minscale = av_clip(minscale - q0, 0, TRELLIS_STATES - 1);
329 maxscale = av_clip(maxscale - q0, 0, TRELLIS_STATES);
330 maxval = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], s->scoefs+start);
331 for (q = minscale; q < maxscale; q++) {
333 int cb = find_min_book(maxval, sce->sf_idx[w*16+g]);
334 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
335 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
336 dist += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
337 q + q0, cb, lambda / band->threshold, INFINITY, NULL, NULL, 0);
339 minrd = FFMIN(minrd, dist);
341 for (i = 0; i < q1 - q0; i++) {
343 cost = paths[idx - 1][i].cost + dist
344 + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
345 if (cost < paths[idx][q].cost) {
346 paths[idx][q].cost = cost;
347 paths[idx][q].prev = i;
352 for (q = 0; q < q1 - q0; q++) {
353 paths[idx][q].cost = paths[idx - 1][q].cost + 1;
354 paths[idx][q].prev = q;
357 sce->zeroes[w*16+g] = !nz;
358 start += sce->ics.swb_sizes[g];
363 mincost = paths[idx][0].cost;
365 for (i = 1; i < TRELLIS_STATES; i++) {
366 if (paths[idx][i].cost < mincost) {
367 mincost = paths[idx][i].cost;
372 sce->sf_idx[bandaddr[idx]] = minq + q0;
373 minq = paths[idx][minq].prev;
376 //set the same quantizers inside window groups
377 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
378 for (g = 0; g < sce->ics.num_swb; g++)
379 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
380 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
384 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
385 SingleChannelElement *sce,
388 int start = 0, i, w, w2, g;
389 float uplim[128], maxq[128];
391 float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
392 int last = 0, lastband = 0, curband = 0;
393 float avg_energy = 0.0;
394 if (sce->ics.num_windows == 1) {
396 for (i = 0; i < 1024; i++) {
397 if (i - start >= sce->ics.swb_sizes[curband]) {
398 start += sce->ics.swb_sizes[curband];
401 if (sce->coeffs[i]) {
402 avg_energy += sce->coeffs[i] * sce->coeffs[i];
408 for (w = 0; w < 8; w++) {
409 const float *coeffs = &sce->coeffs[w*128];
411 for (i = 0; i < 128; i++) {
412 if (i - start >= sce->ics.swb_sizes[curband]) {
413 start += sce->ics.swb_sizes[curband];
417 avg_energy += coeffs[i] * coeffs[i];
418 last = FFMAX(last, i);
419 lastband = FFMAX(lastband, curband);
426 if (avg_energy == 0.0f) {
427 for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
428 sce->sf_idx[i] = SCALE_ONE_POS;
431 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
433 for (g = 0; g < sce->ics.num_swb; g++) {
434 float *coefs = &sce->coeffs[start];
435 const int size = sce->ics.swb_sizes[g];
436 int start2 = start, end2 = start + size, peakpos = start;
437 float maxval = -1, thr = 0.0f, t;
442 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
443 memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
446 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
447 for (i = 0; i < size; i++) {
448 float t = coefs[w2*128+i]*coefs[w2*128+i];
449 maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
451 if (sce->ics.num_windows == 1 && maxval < t) {
457 if (sce->ics.num_windows == 1) {
458 start2 = FFMAX(peakpos - 2, start2);
459 end2 = FFMIN(peakpos + 3, end2);
465 thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
466 t = 1.0 - (1.0 * start2 / last);
467 uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
470 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
471 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
472 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
474 for (g = 0; g < sce->ics.num_swb; g++) {
475 const float *coefs = &sce->coeffs[start];
476 const float *scaled = &s->scoefs[start];
477 const int size = sce->ics.swb_sizes[g];
478 int scf, prev_scf, step;
479 int min_scf = -1, max_scf = 256;
481 if (maxq[w*16+g] < 21.544) {
482 sce->zeroes[w*16+g] = 1;
486 sce->zeroes[w*16+g] = 0;
487 scf = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2f(1/maxq[w*16+g])*16/3, 60, 218);
492 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
494 dist += quantize_band_cost(s, coefs + w2*128,
496 sce->ics.swb_sizes[g],
505 dist *= 1.0f / 512.0f / lambda;
506 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[POW_SF2_ZERO - scf + SCALE_ONE_POS - SCALE_DIV_512], ROUND_STANDARD);
507 if (quant_max >= 8191) { // too much, return to the previous quantizer
508 sce->sf_idx[w*16+g] = prev_scf;
512 curdiff = fabsf(dist - uplim[w*16+g]);
516 step = log2f(curdiff);
517 if (dist > uplim[w*16+g])
520 scf = av_clip_uint8(scf);
521 step = scf - prev_scf;
522 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
523 sce->sf_idx[w*16+g] = av_clip(scf, min_scf, max_scf);
534 minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
535 for (i = 1; i < 128; i++) {
537 sce->sf_idx[i] = sce->sf_idx[i-1];
539 minq = FFMIN(minq, sce->sf_idx[i]);
543 minq = FFMIN(minq, SCALE_MAX_POS);
544 maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
545 for (i = 126; i >= 0; i--) {
547 sce->sf_idx[i] = sce->sf_idx[i+1];
548 sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
552 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
553 SingleChannelElement *sce,
559 memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
560 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
561 for (g = 0; g < sce->ics.num_swb; g++) {
562 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
563 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
564 if (band->energy <= band->threshold) {
565 sce->sf_idx[(w+w2)*16+g] = 218;
566 sce->zeroes[(w+w2)*16+g] = 1;
568 sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2f(band->threshold), 80, 218);
569 sce->zeroes[(w+w2)*16+g] = 0;
571 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
575 for (i = 0; i < 128; i++) {
576 sce->sf_idx[i] = 140;
577 //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
579 //set the same quantizers inside window groups
580 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
581 for (g = 0; g < sce->ics.num_swb; g++)
582 for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
583 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
586 static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChannelElement *sce)
590 int wlen = 1024 / sce->ics.num_windows;
591 int bandwidth, cutoff;
592 float *PNS = &s->scoefs[0*128], *PNS34 = &s->scoefs[1*128];
593 float *NOR34 = &s->scoefs[3*128];
594 const float lambda = s->lambda;
595 const float freq_mult = avctx->sample_rate*0.5f/wlen;
596 const float thr_mult = NOISE_LAMBDA_REPLACE*(100.0f/lambda);
597 const float spread_threshold = FFMIN(0.75f, NOISE_SPREAD_THRESHOLD*FFMAX(0.5f, lambda/100.f));
598 const float dist_bias = av_clipf(4.f * 120 / lambda, 0.25f, 4.0f);
599 const float pns_transient_energy_r = FFMIN(0.7f, lambda / 140.f);
601 int refbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
602 / ((avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : avctx->channels)
605 /** Keep this in sync with twoloop's cutoff selection */
606 float rate_bandwidth_multiplier = 1.5f;
607 int frame_bit_rate = (avctx->flags & CODEC_FLAG_QSCALE)
608 ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
609 : (avctx->bit_rate / avctx->channels);
611 frame_bit_rate *= 1.15f;
613 if (avctx->cutoff > 0) {
614 bandwidth = avctx->cutoff;
616 bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
619 cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
621 memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type));
622 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
624 for (g = 0; g < sce->ics.num_swb; g++) {
626 float dist1 = 0.0f, dist2 = 0.0f, noise_amp;
627 float pns_energy = 0.0f, pns_tgt_energy, energy_ratio, dist_thresh;
628 float sfb_energy = 0.0f, threshold = 0.0f, spread = 2.0f;
629 float min_energy = -1.0f, max_energy = 0.0f;
630 const int start = wstart+sce->ics.swb_offset[g];
631 const float freq = (start-wstart)*freq_mult;
632 const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
633 if (freq < NOISE_LOW_LIMIT || (start-wstart) >= cutoff)
635 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
636 band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
637 sfb_energy += band->energy;
638 spread = FFMIN(spread, band->spread);
639 threshold += band->threshold;
641 min_energy = max_energy = band->energy;
643 min_energy = FFMIN(min_energy, band->energy);
644 max_energy = FFMAX(max_energy, band->energy);
648 /* Ramps down at ~8000Hz and loosens the dist threshold */
649 dist_thresh = av_clipf(2.5f*NOISE_LOW_LIMIT/freq, 0.5f, 2.5f) * dist_bias;
651 /* PNS is acceptable when all of these are true:
652 * 1. high spread energy (noise-like band)
653 * 2. near-threshold energy (high PE means the random nature of PNS content will be noticed)
654 * 3. on short window groups, all windows have similar energy (variations in energy would be destroyed by PNS)
656 * At this stage, point 2 is relaxed for zeroed bands near the noise threshold (hole avoidance is more important)
658 if (((sce->zeroes[w*16+g] || !sce->band_alt[w*16+g]) && sfb_energy < threshold*sqrtf(1.5f/freq_boost)) || spread < spread_threshold ||
659 (!sce->zeroes[w*16+g] && sce->band_alt[w*16+g] && sfb_energy > threshold*thr_mult*freq_boost) ||
660 min_energy < pns_transient_energy_r * max_energy ) {
661 sce->pns_ener[w*16+g] = sfb_energy;
665 pns_tgt_energy = sfb_energy*FFMIN(1.0f, spread*spread);
666 noise_sfi = av_clip(roundf(log2f(pns_tgt_energy)*2), -100, 155); /* Quantize */
667 noise_amp = -ff_aac_pow2sf_tab[noise_sfi + POW_SF2_ZERO]; /* Dequantize */
668 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
669 float band_energy, scale, pns_senergy;
670 const int start_c = (w+w2)*128+sce->ics.swb_offset[g];
671 band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
672 for (i = 0; i < sce->ics.swb_sizes[g]; i++)
673 PNS[i] = s->random_state = lcg_random(s->random_state);
674 band_energy = s->fdsp->scalarproduct_float(PNS, PNS, sce->ics.swb_sizes[g]);
675 scale = noise_amp/sqrtf(band_energy);
676 s->fdsp->vector_fmul_scalar(PNS, PNS, scale, sce->ics.swb_sizes[g]);
677 pns_senergy = s->fdsp->scalarproduct_float(PNS, PNS, sce->ics.swb_sizes[g]);
678 pns_energy += pns_senergy;
679 abs_pow34_v(NOR34, &sce->coeffs[start_c], sce->ics.swb_sizes[g]);
680 abs_pow34_v(PNS34, PNS, sce->ics.swb_sizes[g]);
681 dist1 += quantize_band_cost(s, &sce->coeffs[start_c],
683 sce->ics.swb_sizes[g],
684 sce->sf_idx[(w+w2)*16+g],
685 sce->band_alt[(w+w2)*16+g],
686 lambda/band->threshold, INFINITY, NULL, NULL, 0);
687 /* Estimate rd on average as 5 bits for SF, 4 for the CB, plus spread energy * lambda/thr */
688 dist2 += band->energy/(band->spread*band->spread)*lambda*dist_thresh/band->threshold;
690 if (g && sce->sf_idx[(w+w2)*16+g-1] == NOISE_BT) {
695 energy_ratio = pns_tgt_energy/pns_energy; /* Compensates for quantization error */
696 sce->pns_ener[w*16+g] = energy_ratio*pns_tgt_energy;
697 if (sce->zeroes[w*16+g] || !sce->band_alt[w*16+g] || (energy_ratio > 0.85f && energy_ratio < 1.25f && dist2 < dist1)) {
698 sce->band_type[w*16+g] = NOISE_BT;
699 sce->zeroes[w*16+g] = 0;
705 static void mark_pns(AACEncContext *s, AVCodecContext *avctx, SingleChannelElement *sce)
709 int wlen = 1024 / sce->ics.num_windows;
710 int bandwidth, cutoff;
711 const float lambda = s->lambda;
712 const float freq_mult = avctx->sample_rate*0.5f/wlen;
713 const float spread_threshold = FFMIN(0.75f, NOISE_SPREAD_THRESHOLD*FFMAX(0.5f, lambda/100.f));
714 const float pns_transient_energy_r = FFMIN(0.7f, lambda / 140.f);
716 int refbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
717 / ((avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : avctx->channels)
720 /** Keep this in sync with twoloop's cutoff selection */
721 float rate_bandwidth_multiplier = 1.5f;
722 int frame_bit_rate = (avctx->flags & CODEC_FLAG_QSCALE)
723 ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
724 : (avctx->bit_rate / avctx->channels);
726 frame_bit_rate *= 1.15f;
728 if (avctx->cutoff > 0) {
729 bandwidth = avctx->cutoff;
731 bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
734 cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
736 memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type));
737 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
738 for (g = 0; g < sce->ics.num_swb; g++) {
739 float sfb_energy = 0.0f, threshold = 0.0f, spread = 2.0f;
740 float min_energy = -1.0f, max_energy = 0.0f;
741 const int start = sce->ics.swb_offset[g];
742 const float freq = start*freq_mult;
743 const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
744 if (freq < NOISE_LOW_LIMIT || start >= cutoff) {
745 sce->can_pns[w*16+g] = 0;
748 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
749 band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
750 sfb_energy += band->energy;
751 spread = FFMIN(spread, band->spread);
752 threshold += band->threshold;
754 min_energy = max_energy = band->energy;
756 min_energy = FFMIN(min_energy, band->energy);
757 max_energy = FFMAX(max_energy, band->energy);
761 /* PNS is acceptable when all of these are true:
762 * 1. high spread energy (noise-like band)
763 * 2. near-threshold energy (high PE means the random nature of PNS content will be noticed)
764 * 3. on short window groups, all windows have similar energy (variations in energy would be destroyed by PNS)
766 sce->pns_ener[w*16+g] = sfb_energy;
767 if (sfb_energy < threshold*sqrtf(1.5f/freq_boost) || spread < spread_threshold || min_energy < pns_transient_energy_r * max_energy) {
768 sce->can_pns[w*16+g] = 0;
770 sce->can_pns[w*16+g] = 1;
776 static void search_for_ms(AACEncContext *s, ChannelElement *cpe)
778 int start = 0, i, w, w2, g, sid_sf_boost;
779 float M[128], S[128];
780 float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
781 const float lambda = s->lambda;
782 const float mslambda = FFMIN(1.0f, lambda / 120.f);
783 SingleChannelElement *sce0 = &cpe->ch[0];
784 SingleChannelElement *sce1 = &cpe->ch[1];
785 if (!cpe->common_window)
787 for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
788 int min_sf_idx_mid = SCALE_MAX_POS;
789 int min_sf_idx_side = SCALE_MAX_POS;
790 for (g = 0; g < sce0->ics.num_swb; g++) {
791 if (!sce0->zeroes[w*16+g] && sce0->band_type[w*16+g] < RESERVED_BT)
792 min_sf_idx_mid = FFMIN(min_sf_idx_mid, sce0->sf_idx[w*16+g]);
793 if (!sce1->zeroes[w*16+g] && sce1->band_type[w*16+g] < RESERVED_BT)
794 min_sf_idx_side = FFMIN(min_sf_idx_side, sce1->sf_idx[w*16+g]);
798 for (g = 0; g < sce0->ics.num_swb; g++) {
799 float bmax = bval2bmax(g * 17.0f / sce0->ics.num_swb) / 0.0045f;
800 cpe->ms_mask[w*16+g] = 0;
801 if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
802 float Mmax = 0.0f, Smax = 0.0f;
804 /* Must compute mid/side SF and book for the whole window group */
805 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
806 for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
807 M[i] = (sce0->coeffs[start+(w+w2)*128+i]
808 + sce1->coeffs[start+(w+w2)*128+i]) * 0.5;
810 - sce1->coeffs[start+(w+w2)*128+i];
812 abs_pow34_v(M34, M, sce0->ics.swb_sizes[g]);
813 abs_pow34_v(S34, S, sce0->ics.swb_sizes[g]);
814 for (i = 0; i < sce0->ics.swb_sizes[g]; i++ ) {
815 Mmax = FFMAX(Mmax, M34[i]);
816 Smax = FFMAX(Smax, S34[i]);
820 for (sid_sf_boost = 0; sid_sf_boost < 4; sid_sf_boost++) {
821 float dist1 = 0.0f, dist2 = 0.0f;
827 minidx = FFMIN(sce0->sf_idx[w*16+g], sce1->sf_idx[w*16+g]);
828 mididx = av_clip(minidx, min_sf_idx_mid, min_sf_idx_mid + SCALE_MAX_DIFF);
829 sididx = av_clip(minidx - sid_sf_boost * 3, min_sf_idx_side, min_sf_idx_side + SCALE_MAX_DIFF);
830 midcb = find_min_book(Mmax, mididx);
831 sidcb = find_min_book(Smax, sididx);
833 if ((mididx > minidx) || (sididx > minidx)) {
834 /* scalefactor range violation, bad stuff, will decrease quality unacceptably */
838 /* No CB can be zero */
839 midcb = FFMAX(1,midcb);
840 sidcb = FFMAX(1,sidcb);
842 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
843 FFPsyBand *band0 = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g];
844 FFPsyBand *band1 = &s->psy.ch[s->cur_channel+1].psy_bands[(w+w2)*16+g];
845 float minthr = FFMIN(band0->threshold, band1->threshold);
847 for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
848 M[i] = (sce0->coeffs[start+(w+w2)*128+i]
849 + sce1->coeffs[start+(w+w2)*128+i]) * 0.5;
851 - sce1->coeffs[start+(w+w2)*128+i];
854 abs_pow34_v(L34, sce0->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
855 abs_pow34_v(R34, sce1->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
856 abs_pow34_v(M34, M, sce0->ics.swb_sizes[g]);
857 abs_pow34_v(S34, S, sce0->ics.swb_sizes[g]);
858 dist1 += quantize_band_cost(s, &sce0->coeffs[start + (w+w2)*128],
860 sce0->ics.swb_sizes[g],
861 sce0->sf_idx[(w+w2)*16+g],
862 sce0->band_type[(w+w2)*16+g],
863 lambda / band0->threshold, INFINITY, &b1, NULL, 0);
864 dist1 += quantize_band_cost(s, &sce1->coeffs[start + (w+w2)*128],
866 sce1->ics.swb_sizes[g],
867 sce1->sf_idx[(w+w2)*16+g],
868 sce1->band_type[(w+w2)*16+g],
869 lambda / band1->threshold, INFINITY, &b2, NULL, 0);
870 dist2 += quantize_band_cost(s, M,
872 sce0->ics.swb_sizes[g],
873 sce0->sf_idx[(w+w2)*16+g],
874 sce0->band_type[(w+w2)*16+g],
875 lambda / minthr, INFINITY, &b3, NULL, 0);
876 dist2 += quantize_band_cost(s, S,
878 sce1->ics.swb_sizes[g],
879 sce1->sf_idx[(w+w2)*16+g],
880 sce1->band_type[(w+w2)*16+g],
881 mslambda / (minthr * bmax), INFINITY, &b4, NULL, 0);
887 cpe->ms_mask[w*16+g] = dist2 <= dist1 && B1 < B0;
888 if (cpe->ms_mask[w*16+g]) {
889 /* Setting the M/S mask is useful with I/S or PNS, but only the flag */
890 if (!cpe->is_mask[w*16+g] && sce0->band_type[w*16+g] != NOISE_BT && sce1->band_type[w*16+g] != NOISE_BT) {
891 sce0->sf_idx[w*16+g] = mididx;
892 sce1->sf_idx[w*16+g] = sididx;
893 sce0->band_type[w*16+g] = midcb;
894 sce1->band_type[w*16+g] = sidcb;
897 } else if (B1 > B0) {
898 /* More boost won't fix this */
903 start += sce0->ics.swb_sizes[g];
908 AACCoefficientsEncoder ff_aac_coders[AAC_CODER_NB] = {
910 search_for_quantizers_faac,
911 encode_window_bands_info,
912 quantize_and_encode_band,
913 ff_aac_encode_tns_info,
914 ff_aac_encode_ltp_info,
915 ff_aac_encode_main_pred,
916 ff_aac_adjust_common_pred,
917 ff_aac_adjust_common_ltp,
918 ff_aac_apply_main_pred,
921 ff_aac_ltp_insert_new_frame,
922 set_special_band_scalefactors,
925 ff_aac_search_for_tns,
926 ff_aac_search_for_ltp,
928 ff_aac_search_for_is,
929 ff_aac_search_for_pred,
932 search_for_quantizers_anmr,
933 encode_window_bands_info,
934 quantize_and_encode_band,
935 ff_aac_encode_tns_info,
936 ff_aac_encode_ltp_info,
937 ff_aac_encode_main_pred,
938 ff_aac_adjust_common_pred,
939 ff_aac_adjust_common_ltp,
940 ff_aac_apply_main_pred,
943 ff_aac_ltp_insert_new_frame,
944 set_special_band_scalefactors,
947 ff_aac_search_for_tns,
948 ff_aac_search_for_ltp,
950 ff_aac_search_for_is,
951 ff_aac_search_for_pred,
953 [AAC_CODER_TWOLOOP] = {
954 search_for_quantizers_twoloop,
955 codebook_trellis_rate,
956 quantize_and_encode_band,
957 ff_aac_encode_tns_info,
958 ff_aac_encode_ltp_info,
959 ff_aac_encode_main_pred,
960 ff_aac_adjust_common_pred,
961 ff_aac_adjust_common_ltp,
962 ff_aac_apply_main_pred,
965 ff_aac_ltp_insert_new_frame,
966 set_special_band_scalefactors,
969 ff_aac_search_for_tns,
970 ff_aac_search_for_ltp,
972 ff_aac_search_for_is,
973 ff_aac_search_for_pred,
976 search_for_quantizers_fast,
977 encode_window_bands_info,
978 quantize_and_encode_band,
979 ff_aac_encode_tns_info,
980 ff_aac_encode_ltp_info,
981 ff_aac_encode_main_pred,
982 ff_aac_adjust_common_pred,
983 ff_aac_adjust_common_ltp,
984 ff_aac_apply_main_pred,
987 ff_aac_ltp_insert_new_frame,
988 set_special_band_scalefactors,
991 ff_aac_search_for_tns,
992 ff_aac_search_for_ltp,
994 ff_aac_search_for_is,
995 ff_aac_search_for_pred,