]> git.sesse.net Git - ffmpeg/blob - libavcodec/aaccoder.c
Merge commit '4012fe1ee819edc7689e182189e66c5401fb4b41'
[ffmpeg] / libavcodec / aaccoder.c
1 /*
2  * AAC coefficients encoder
3  * Copyright (C) 2008-2009 Konstantin Shishkov
4  *
5  * This file is part of FFmpeg.
6  *
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.
11  *
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.
16  *
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
20  */
21
22 /**
23  * @file
24  * AAC coefficients encoder
25  */
26
27 /***********************************
28  *              TODOs:
29  * speedup quantizer selection
30  * add sane pulse detection
31  ***********************************/
32
33 #include "libavutil/libm.h" // brought forward to work around cygwin header breakage
34
35 #include <float.h>
36
37 #include "libavutil/mathematics.h"
38 #include "mathops.h"
39 #include "avcodec.h"
40 #include "put_bits.h"
41 #include "aac.h"
42 #include "aacenc.h"
43 #include "aactab.h"
44 #include "aacenctab.h"
45 #include "aacenc_utils.h"
46 #include "aacenc_quantization.h"
47
48 #include "aacenc_is.h"
49 #include "aacenc_tns.h"
50 #include "aacenc_ltp.h"
51 #include "aacenc_pred.h"
52
53 #include "libavcodec/aaccoder_twoloop.h"
54
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
58
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
62
63 #include "libavcodec/aaccoder_trellis.h"
64
65 /**
66  * structure used in optimal codebook search
67  */
68 typedef struct BandCodingPath {
69     int prev_idx; ///< pointer to the previous path point
70     float cost;   ///< path cost
71     int run;
72 } BandCodingPath;
73
74 /**
75  * Encode band info for single window group bands.
76  */
77 static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
78                                      int win, int group_len, const float lambda)
79 {
80     BandCodingPath path[120][CB_TOT_ALL];
81     int w, swb, cb, start, size;
82     int i, j;
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;
86     int idx, ppos, count;
87     int stackrun[120], stackcb[120], stack_len;
88     float next_minrd = INFINITY;
89     int next_mincb = 0;
90
91     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
92     start = win*128;
93     for (cb = 0; cb < CB_TOT_ALL; cb++) {
94         path[0][cb].cost     = 0.0f;
95         path[0][cb].prev_idx = -1;
96         path[0][cb].run      = 0;
97     }
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;
105             }
106         } else {
107             float minrd = next_minrd;
108             int mincb = next_mincb;
109             next_minrd = INFINITY;
110             next_mincb = 0;
111             for (cb = 0; cb < CB_TOT_ALL; cb++) {
112                 float cost_stay_here, cost_get_here;
113                 float rd = 0.0f;
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;
119                     continue;
120                 }
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);
127                 }
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;
137                 } else {
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;
141                 }
142                 if (path[swb+1][cb].cost < next_minrd) {
143                     next_minrd = path[swb+1][cb].cost;
144                     next_mincb = cb;
145                 }
146             }
147         }
148         start += sce->ics.swb_sizes[swb];
149     }
150
151     //convert resulting path from backward-linked list
152     stack_len = 0;
153     idx       = 0;
154     for (cb = 1; cb < CB_TOT_ALL; cb++)
155         if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
156             idx = cb;
157     ppos = max_sfb;
158     while (ppos > 0) {
159         av_assert1(idx >= 0);
160         cb = idx;
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;
165         stack_len++;
166     }
167     //perform actual band info encoding
168     start = 0;
169     for (i = stack_len - 1; i >= 0; i--) {
170         cb = aac_cb_out_map[stackcb[i]];
171         put_bits(&s->pb, 4, cb);
172         count = stackrun[i];
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;
177             start++;
178         }
179         while (count >= run_esc) {
180             put_bits(&s->pb, run_bits, run_esc);
181             count -= run_esc;
182         }
183         put_bits(&s->pb, run_bits, count);
184     }
185 }
186
187
188 typedef struct TrellisPath {
189     float cost;
190     int prev;
191 } TrellisPath;
192
193 #define TRELLIS_STAGES 121
194 #define TRELLIS_STATES (SCALE_MAX_DIFF+1)
195
196 static void set_special_band_scalefactors(AACEncContext *s, SingleChannelElement *sce)
197 {
198     int w, g;
199     int prevscaler_n = -255, prevscaler_i = 0;
200     int bands = 0;
201
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])
205                 continue;
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);
208                 bands++;
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];
213                 bands++;
214             }
215         }
216     }
217
218     if (!bands)
219         return;
220
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])
225                 continue;
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);
230             }
231         }
232     }
233 }
234
235 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
236                                        SingleChannelElement *sce,
237                                        const float lambda)
238 {
239     int q, w, w2, g, start = 0;
240     int i, j;
241     int idx;
242     TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
243     int bandaddr[TRELLIS_STAGES];
244     int minq;
245     float mincost;
246     float q0f = FLT_MAX, q1f = 0.0f, qnrgf = 0.0f;
247     int q0, q1, qcnt = 0;
248
249     for (i = 0; i < 1024; i++) {
250         float t = fabsf(sce->coeffs[i]);
251         if (t > 0.0f) {
252             q0f = FFMIN(q0f, t);
253             q1f = FFMAX(q1f, t);
254             qnrgf += t*t;
255             qcnt++;
256         }
257     }
258
259     if (!qcnt) {
260         memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
261         memset(sce->zeroes, 1, sizeof(sce->zeroes));
262         return;
263     }
264
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);
269     if (q1 - q0 > 60) {
270         int q0low  = q0;
271         int q1high = q1;
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);
274         q1 = qnrg + 30;
275         q0 = qnrg - 30;
276         if (q0 < q0low) {
277             q1 += q0low - q0;
278             q0  = q0low;
279         } else if (q1 > q1high) {
280             q0 -= q1 - q1high;
281             q1  = q1high;
282         }
283     }
284     // q0 == q1 isn't really a legal situation
285     if (q0 == q1) {
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);
289     }
290
291     for (i = 0; i < TRELLIS_STATES; i++) {
292         paths[0][i].cost    = 0.0f;
293         paths[0][i].prev    = -1;
294     }
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;
299         }
300     }
301     idx = 1;
302     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
303     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
304         start = w*128;
305         for (g = 0; g < sce->ics.num_swb; g++) {
306             const float *coefs = &sce->coeffs[start];
307             float qmin, qmax;
308             int nz = 0;
309
310             bandaddr[idx] = w * 16 + g;
311             qmin = INT_MAX;
312             qmax = 0.0f;
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;
317                     continue;
318                 }
319                 sce->zeroes[(w+w2)*16+g] = 0;
320                 nz = 1;
321                 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
322                     float t = fabsf(coefs[w2*128+i]);
323                     if (t > 0.0f)
324                         qmin = FFMIN(qmin, t);
325                     qmax = FFMAX(qmax, t);
326                 }
327             }
328             if (nz) {
329                 int minscale, maxscale;
330                 float minrd = INFINITY;
331                 float maxval;
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);
341                 }
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++) {
344                     float dist = 0;
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);
350                     }
351                     minrd = FFMIN(minrd, dist);
352
353                     for (i = 0; i < q1 - q0; i++) {
354                         float cost;
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;
360                         }
361                     }
362                 }
363             } else {
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;
367                 }
368             }
369             sce->zeroes[w*16+g] = !nz;
370             start += sce->ics.swb_sizes[g];
371             idx++;
372         }
373     }
374     idx--;
375     mincost = paths[idx][0].cost;
376     minq    = 0;
377     for (i = 1; i < TRELLIS_STATES; i++) {
378         if (paths[idx][i].cost < mincost) {
379             mincost = paths[idx][i].cost;
380             minq = i;
381         }
382     }
383     while (idx) {
384         sce->sf_idx[bandaddr[idx]] = minq + q0;
385         minq = FFMAX(paths[idx][minq].prev, 0);
386         idx--;
387     }
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];
393 }
394
395 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
396                                        SingleChannelElement *sce,
397                                        const float lambda)
398 {
399     int i, w, w2, g;
400     int minq = 255;
401
402     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
403     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
404         for (g = 0; g < sce->ics.num_swb; g++) {
405             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
406                 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
407                 if (band->energy <= band->threshold) {
408                     sce->sf_idx[(w+w2)*16+g] = 218;
409                     sce->zeroes[(w+w2)*16+g] = 1;
410                 } else {
411                     sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2f(band->threshold), 80, 218);
412                     sce->zeroes[(w+w2)*16+g] = 0;
413                 }
414                 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
415             }
416         }
417     }
418     for (i = 0; i < 128; i++) {
419         sce->sf_idx[i] = 140;
420         //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
421     }
422     //set the same quantizers inside window groups
423     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
424         for (g = 0;  g < sce->ics.num_swb; g++)
425             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
426                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
427 }
428
429 static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChannelElement *sce)
430 {
431     FFPsyBand *band;
432     int w, g, w2, i;
433     int wlen = 1024 / sce->ics.num_windows;
434     int bandwidth, cutoff;
435     float *PNS = &s->scoefs[0*128], *PNS34 = &s->scoefs[1*128];
436     float *NOR34 = &s->scoefs[3*128];
437     uint8_t nextband[128];
438     const float lambda = s->lambda;
439     const float freq_mult = avctx->sample_rate*0.5f/wlen;
440     const float thr_mult = NOISE_LAMBDA_REPLACE*(100.0f/lambda);
441     const float spread_threshold = FFMIN(0.75f, NOISE_SPREAD_THRESHOLD*FFMAX(0.5f, lambda/100.f));
442     const float dist_bias = av_clipf(4.f * 120 / lambda, 0.25f, 4.0f);
443     const float pns_transient_energy_r = FFMIN(0.7f, lambda / 140.f);
444
445     int refbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
446         / ((avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : avctx->channels)
447         * (lambda / 120.f);
448
449     /** Keep this in sync with twoloop's cutoff selection */
450     float rate_bandwidth_multiplier = 1.5f;
451     int prev = -1000, prev_sf = -1;
452     int frame_bit_rate = (avctx->flags & CODEC_FLAG_QSCALE)
453         ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
454         : (avctx->bit_rate / avctx->channels);
455
456     frame_bit_rate *= 1.15f;
457
458     if (avctx->cutoff > 0) {
459         bandwidth = avctx->cutoff;
460     } else {
461         bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
462     }
463
464     cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
465
466     memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type));
467     ff_init_nextband_map(sce, nextband);
468     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
469         int wstart = w*128;
470         for (g = 0;  g < sce->ics.num_swb; g++) {
471             int noise_sfi;
472             float dist1 = 0.0f, dist2 = 0.0f, noise_amp;
473             float pns_energy = 0.0f, pns_tgt_energy, energy_ratio, dist_thresh;
474             float sfb_energy = 0.0f, threshold = 0.0f, spread = 2.0f;
475             float min_energy = -1.0f, max_energy = 0.0f;
476             const int start = wstart+sce->ics.swb_offset[g];
477             const float freq = (start-wstart)*freq_mult;
478             const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
479             if (freq < NOISE_LOW_LIMIT || (start-wstart) >= cutoff) {
480                 if (!sce->zeroes[w*16+g])
481                     prev_sf = sce->sf_idx[w*16+g];
482                 continue;
483             }
484             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
485                 band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
486                 sfb_energy += band->energy;
487                 spread     = FFMIN(spread, band->spread);
488                 threshold  += band->threshold;
489                 if (!w2) {
490                     min_energy = max_energy = band->energy;
491                 } else {
492                     min_energy = FFMIN(min_energy, band->energy);
493                     max_energy = FFMAX(max_energy, band->energy);
494                 }
495             }
496
497             /* Ramps down at ~8000Hz and loosens the dist threshold */
498             dist_thresh = av_clipf(2.5f*NOISE_LOW_LIMIT/freq, 0.5f, 2.5f) * dist_bias;
499
500             /* PNS is acceptable when all of these are true:
501              * 1. high spread energy (noise-like band)
502              * 2. near-threshold energy (high PE means the random nature of PNS content will be noticed)
503              * 3. on short window groups, all windows have similar energy (variations in energy would be destroyed by PNS)
504              *
505              * At this stage, point 2 is relaxed for zeroed bands near the noise threshold (hole avoidance is more important)
506              */
507             if ((!sce->zeroes[w*16+g] && !ff_sfdelta_can_remove_band(sce, nextband, prev_sf, w*16+g)) ||
508                 ((sce->zeroes[w*16+g] || !sce->band_alt[w*16+g]) && sfb_energy < threshold*sqrtf(1.0f/freq_boost)) || spread < spread_threshold ||
509                 (!sce->zeroes[w*16+g] && sce->band_alt[w*16+g] && sfb_energy > threshold*thr_mult*freq_boost) ||
510                 min_energy < pns_transient_energy_r * max_energy ) {
511                 sce->pns_ener[w*16+g] = sfb_energy;
512                 if (!sce->zeroes[w*16+g])
513                     prev_sf = sce->sf_idx[w*16+g];
514                 continue;
515             }
516
517             pns_tgt_energy = sfb_energy*FFMIN(1.0f, spread*spread);
518             noise_sfi = av_clip(roundf(log2f(pns_tgt_energy)*2), -100, 155); /* Quantize */
519             noise_amp = -ff_aac_pow2sf_tab[noise_sfi + POW_SF2_ZERO];    /* Dequantize */
520             if (prev != -1000) {
521                 int noise_sfdiff = noise_sfi - prev + SCALE_DIFF_ZERO;
522                 if (noise_sfdiff < 0 || noise_sfdiff > 2*SCALE_MAX_DIFF) {
523                     if (!sce->zeroes[w*16+g])
524                         prev_sf = sce->sf_idx[w*16+g];
525                     continue;
526                 }
527             }
528             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
529                 float band_energy, scale, pns_senergy;
530                 const int start_c = (w+w2)*128+sce->ics.swb_offset[g];
531                 band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
532                 for (i = 0; i < sce->ics.swb_sizes[g]; i+=2) {
533                     double rnd[2];
534                     av_bmg_get(&s->lfg, rnd);
535                     PNS[i+0] = (float)rnd[0];
536                     PNS[i+1] = (float)rnd[1];
537                 }
538                 band_energy = s->fdsp->scalarproduct_float(PNS, PNS, sce->ics.swb_sizes[g]);
539                 scale = noise_amp/sqrtf(band_energy);
540                 s->fdsp->vector_fmul_scalar(PNS, PNS, scale, sce->ics.swb_sizes[g]);
541                 pns_senergy = s->fdsp->scalarproduct_float(PNS, PNS, sce->ics.swb_sizes[g]);
542                 pns_energy += pns_senergy;
543                 abs_pow34_v(NOR34, &sce->coeffs[start_c], sce->ics.swb_sizes[g]);
544                 abs_pow34_v(PNS34, PNS, sce->ics.swb_sizes[g]);
545                 dist1 += quantize_band_cost(s, &sce->coeffs[start_c],
546                                             NOR34,
547                                             sce->ics.swb_sizes[g],
548                                             sce->sf_idx[(w+w2)*16+g],
549                                             sce->band_alt[(w+w2)*16+g],
550                                             lambda/band->threshold, INFINITY, NULL, NULL, 0);
551                 /* Estimate rd on average as 5 bits for SF, 4 for the CB, plus spread energy * lambda/thr */
552                 dist2 += band->energy/(band->spread*band->spread)*lambda*dist_thresh/band->threshold;
553             }
554             if (g && sce->band_type[w*16+g-1] == NOISE_BT) {
555                 dist2 += 5;
556             } else {
557                 dist2 += 9;
558             }
559             energy_ratio = pns_tgt_energy/pns_energy; /* Compensates for quantization error */
560             sce->pns_ener[w*16+g] = energy_ratio*pns_tgt_energy;
561             if (sce->zeroes[w*16+g] || !sce->band_alt[w*16+g] || (energy_ratio > 0.85f && energy_ratio < 1.25f && dist2 < dist1)) {
562                 sce->band_type[w*16+g] = NOISE_BT;
563                 sce->zeroes[w*16+g] = 0;
564                 prev = noise_sfi;
565             } else {
566                 if (!sce->zeroes[w*16+g])
567                     prev_sf = sce->sf_idx[w*16+g];
568             }
569         }
570     }
571 }
572
573 static void mark_pns(AACEncContext *s, AVCodecContext *avctx, SingleChannelElement *sce)
574 {
575     FFPsyBand *band;
576     int w, g, w2;
577     int wlen = 1024 / sce->ics.num_windows;
578     int bandwidth, cutoff;
579     const float lambda = s->lambda;
580     const float freq_mult = avctx->sample_rate*0.5f/wlen;
581     const float spread_threshold = FFMIN(0.75f, NOISE_SPREAD_THRESHOLD*FFMAX(0.5f, lambda/100.f));
582     const float pns_transient_energy_r = FFMIN(0.7f, lambda / 140.f);
583
584     int refbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
585         / ((avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : avctx->channels)
586         * (lambda / 120.f);
587
588     /** Keep this in sync with twoloop's cutoff selection */
589     float rate_bandwidth_multiplier = 1.5f;
590     int frame_bit_rate = (avctx->flags & CODEC_FLAG_QSCALE)
591         ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
592         : (avctx->bit_rate / avctx->channels);
593
594     frame_bit_rate *= 1.15f;
595
596     if (avctx->cutoff > 0) {
597         bandwidth = avctx->cutoff;
598     } else {
599         bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
600     }
601
602     cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
603
604     memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type));
605     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
606         for (g = 0;  g < sce->ics.num_swb; g++) {
607             float sfb_energy = 0.0f, threshold = 0.0f, spread = 2.0f;
608             float min_energy = -1.0f, max_energy = 0.0f;
609             const int start = sce->ics.swb_offset[g];
610             const float freq = start*freq_mult;
611             const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
612             if (freq < NOISE_LOW_LIMIT || start >= cutoff) {
613                 sce->can_pns[w*16+g] = 0;
614                 continue;
615             }
616             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
617                 band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
618                 sfb_energy += band->energy;
619                 spread     = FFMIN(spread, band->spread);
620                 threshold  += band->threshold;
621                 if (!w2) {
622                     min_energy = max_energy = band->energy;
623                 } else {
624                     min_energy = FFMIN(min_energy, band->energy);
625                     max_energy = FFMAX(max_energy, band->energy);
626                 }
627             }
628
629             /* PNS is acceptable when all of these are true:
630              * 1. high spread energy (noise-like band)
631              * 2. near-threshold energy (high PE means the random nature of PNS content will be noticed)
632              * 3. on short window groups, all windows have similar energy (variations in energy would be destroyed by PNS)
633              */
634             sce->pns_ener[w*16+g] = sfb_energy;
635             if (sfb_energy < threshold*sqrtf(1.5f/freq_boost) || spread < spread_threshold || min_energy < pns_transient_energy_r * max_energy) {
636                 sce->can_pns[w*16+g] = 0;
637             } else {
638                 sce->can_pns[w*16+g] = 1;
639             }
640         }
641     }
642 }
643
644 static void search_for_ms(AACEncContext *s, ChannelElement *cpe)
645 {
646     int start = 0, i, w, w2, g, sid_sf_boost, prev_mid, prev_side;
647     uint8_t nextband0[128], nextband1[128];
648     float M[128], S[128];
649     float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
650     const float lambda = s->lambda;
651     const float mslambda = FFMIN(1.0f, lambda / 120.f);
652     SingleChannelElement *sce0 = &cpe->ch[0];
653     SingleChannelElement *sce1 = &cpe->ch[1];
654     if (!cpe->common_window)
655         return;
656
657     /** Scout out next nonzero bands */
658     ff_init_nextband_map(sce0, nextband0);
659     ff_init_nextband_map(sce1, nextband1);
660
661     prev_mid = sce0->sf_idx[0];
662     prev_side = sce1->sf_idx[0];
663     for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
664         start = 0;
665         for (g = 0;  g < sce0->ics.num_swb; g++) {
666             float bmax = bval2bmax(g * 17.0f / sce0->ics.num_swb) / 0.0045f;
667             if (!cpe->is_mask[w*16+g])
668                 cpe->ms_mask[w*16+g] = 0;
669             if (!sce0->zeroes[w*16+g] && !sce1->zeroes[w*16+g] && !cpe->is_mask[w*16+g]) {
670                 float Mmax = 0.0f, Smax = 0.0f;
671
672                 /* Must compute mid/side SF and book for the whole window group */
673                 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
674                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
675                         M[i] = (sce0->coeffs[start+(w+w2)*128+i]
676                               + sce1->coeffs[start+(w+w2)*128+i]) * 0.5;
677                         S[i] =  M[i]
678                               - sce1->coeffs[start+(w+w2)*128+i];
679                     }
680                     abs_pow34_v(M34, M, sce0->ics.swb_sizes[g]);
681                     abs_pow34_v(S34, S, sce0->ics.swb_sizes[g]);
682                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++ ) {
683                         Mmax = FFMAX(Mmax, M34[i]);
684                         Smax = FFMAX(Smax, S34[i]);
685                     }
686                 }
687
688                 for (sid_sf_boost = 0; sid_sf_boost < 4; sid_sf_boost++) {
689                     float dist1 = 0.0f, dist2 = 0.0f;
690                     int B0 = 0, B1 = 0;
691                     int minidx;
692                     int mididx, sididx;
693                     int midcb, sidcb;
694
695                     minidx = FFMIN(sce0->sf_idx[w*16+g], sce1->sf_idx[w*16+g]);
696                     mididx = av_clip(minidx, 0, SCALE_MAX_POS - SCALE_DIV_512);
697                     sididx = av_clip(minidx - sid_sf_boost * 3, 0, SCALE_MAX_POS - SCALE_DIV_512);
698                     if (sce0->band_type[w*16+g] != NOISE_BT && sce1->band_type[w*16+g] != NOISE_BT
699                         && (   !ff_sfdelta_can_replace(sce0, nextband0, prev_mid, mididx, w*16+g)
700                             || !ff_sfdelta_can_replace(sce1, nextband1, prev_side, sididx, w*16+g))) {
701                         /* scalefactor range violation, bad stuff, will decrease quality unacceptably */
702                         continue;
703                     }
704
705                     midcb = find_min_book(Mmax, mididx);
706                     sidcb = find_min_book(Smax, sididx);
707
708                     /* No CB can be zero */
709                     midcb = FFMAX(1,midcb);
710                     sidcb = FFMAX(1,sidcb);
711
712                     for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
713                         FFPsyBand *band0 = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g];
714                         FFPsyBand *band1 = &s->psy.ch[s->cur_channel+1].psy_bands[(w+w2)*16+g];
715                         float minthr = FFMIN(band0->threshold, band1->threshold);
716                         int b1,b2,b3,b4;
717                         for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
718                             M[i] = (sce0->coeffs[start+(w+w2)*128+i]
719                                   + sce1->coeffs[start+(w+w2)*128+i]) * 0.5;
720                             S[i] =  M[i]
721                                   - sce1->coeffs[start+(w+w2)*128+i];
722                         }
723
724                         abs_pow34_v(L34, sce0->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
725                         abs_pow34_v(R34, sce1->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
726                         abs_pow34_v(M34, M,                         sce0->ics.swb_sizes[g]);
727                         abs_pow34_v(S34, S,                         sce0->ics.swb_sizes[g]);
728                         dist1 += quantize_band_cost(s, &sce0->coeffs[start + (w+w2)*128],
729                                                     L34,
730                                                     sce0->ics.swb_sizes[g],
731                                                     sce0->sf_idx[w*16+g],
732                                                     sce0->band_type[w*16+g],
733                                                     lambda / band0->threshold, INFINITY, &b1, NULL, 0);
734                         dist1 += quantize_band_cost(s, &sce1->coeffs[start + (w+w2)*128],
735                                                     R34,
736                                                     sce1->ics.swb_sizes[g],
737                                                     sce1->sf_idx[w*16+g],
738                                                     sce1->band_type[w*16+g],
739                                                     lambda / band1->threshold, INFINITY, &b2, NULL, 0);
740                         dist2 += quantize_band_cost(s, M,
741                                                     M34,
742                                                     sce0->ics.swb_sizes[g],
743                                                     mididx,
744                                                     midcb,
745                                                     lambda / minthr, INFINITY, &b3, NULL, 0);
746                         dist2 += quantize_band_cost(s, S,
747                                                     S34,
748                                                     sce1->ics.swb_sizes[g],
749                                                     sididx,
750                                                     sidcb,
751                                                     mslambda / (minthr * bmax), INFINITY, &b4, NULL, 0);
752                         B0 += b1+b2;
753                         B1 += b3+b4;
754                         dist1 -= b1+b2;
755                         dist2 -= b3+b4;
756                     }
757                     cpe->ms_mask[w*16+g] = dist2 <= dist1 && B1 < B0;
758                     if (cpe->ms_mask[w*16+g]) {
759                         if (sce0->band_type[w*16+g] != NOISE_BT && sce1->band_type[w*16+g] != NOISE_BT) {
760                             sce0->sf_idx[w*16+g] = mididx;
761                             sce1->sf_idx[w*16+g] = sididx;
762                             sce0->band_type[w*16+g] = midcb;
763                             sce1->band_type[w*16+g] = sidcb;
764                         } else if ((sce0->band_type[w*16+g] != NOISE_BT) ^ (sce1->band_type[w*16+g] != NOISE_BT)) {
765                             /* ms_mask unneeded, and it confuses some decoders */
766                             cpe->ms_mask[w*16+g] = 0;
767                         }
768                         break;
769                     } else if (B1 > B0) {
770                         /* More boost won't fix this */
771                         break;
772                     }
773                 }
774             }
775             if (!sce0->zeroes[w*16+g] && sce0->band_type[w*16+g] < RESERVED_BT)
776                 prev_mid = sce0->sf_idx[w*16+g];
777             if (!sce1->zeroes[w*16+g] && !cpe->is_mask[w*16+g] && sce1->band_type[w*16+g] < RESERVED_BT)
778                 prev_side = sce1->sf_idx[w*16+g];
779             start += sce0->ics.swb_sizes[g];
780         }
781     }
782 }
783
784 AACCoefficientsEncoder ff_aac_coders[AAC_CODER_NB] = {
785     [AAC_CODER_ANMR] = {
786         search_for_quantizers_anmr,
787         encode_window_bands_info,
788         quantize_and_encode_band,
789         ff_aac_encode_tns_info,
790         ff_aac_encode_ltp_info,
791         ff_aac_encode_main_pred,
792         ff_aac_adjust_common_pred,
793         ff_aac_adjust_common_ltp,
794         ff_aac_apply_main_pred,
795         ff_aac_apply_tns,
796         ff_aac_update_ltp,
797         ff_aac_ltp_insert_new_frame,
798         set_special_band_scalefactors,
799         search_for_pns,
800         mark_pns,
801         ff_aac_search_for_tns,
802         ff_aac_search_for_ltp,
803         search_for_ms,
804         ff_aac_search_for_is,
805         ff_aac_search_for_pred,
806     },
807     [AAC_CODER_TWOLOOP] = {
808         search_for_quantizers_twoloop,
809         codebook_trellis_rate,
810         quantize_and_encode_band,
811         ff_aac_encode_tns_info,
812         ff_aac_encode_ltp_info,
813         ff_aac_encode_main_pred,
814         ff_aac_adjust_common_pred,
815         ff_aac_adjust_common_ltp,
816         ff_aac_apply_main_pred,
817         ff_aac_apply_tns,
818         ff_aac_update_ltp,
819         ff_aac_ltp_insert_new_frame,
820         set_special_band_scalefactors,
821         search_for_pns,
822         mark_pns,
823         ff_aac_search_for_tns,
824         ff_aac_search_for_ltp,
825         search_for_ms,
826         ff_aac_search_for_is,
827         ff_aac_search_for_pred,
828     },
829     [AAC_CODER_FAST] = {
830         search_for_quantizers_fast,
831         encode_window_bands_info,
832         quantize_and_encode_band,
833         ff_aac_encode_tns_info,
834         ff_aac_encode_ltp_info,
835         ff_aac_encode_main_pred,
836         ff_aac_adjust_common_pred,
837         ff_aac_adjust_common_ltp,
838         ff_aac_apply_main_pred,
839         ff_aac_apply_tns,
840         ff_aac_update_ltp,
841         ff_aac_ltp_insert_new_frame,
842         set_special_band_scalefactors,
843         search_for_pns,
844         mark_pns,
845         ff_aac_search_for_tns,
846         ff_aac_search_for_ltp,
847         search_for_ms,
848         ff_aac_search_for_is,
849         ff_aac_search_for_pred,
850     },
851 };