bf515c21c19924f5efe0a603a66e5d680231ec43
[ffmpeg.git] / 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, start = 0;
199     int minscaler_n = sce->sf_idx[0], minscaler_i = sce->sf_idx[0];
200     int bands = 0;
201
202     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
203         start = 0;
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]);
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                 minscaler_n = FFMIN(minscaler_n, sce->sf_idx[w*16+g]);
212                 bands++;
213             }
214             start += sce->ics.swb_sizes[g];
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->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);
228             }
229         }
230     }
231 }
232
233 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
234                                        SingleChannelElement *sce,
235                                        const float lambda)
236 {
237     int q, w, w2, g, start = 0;
238     int i, j;
239     int idx;
240     TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
241     int bandaddr[TRELLIS_STAGES];
242     int minq;
243     float mincost;
244     float q0f = FLT_MAX, q1f = 0.0f, qnrgf = 0.0f;
245     int q0, q1, qcnt = 0;
246
247     for (i = 0; i < 1024; i++) {
248         float t = fabsf(sce->coeffs[i]);
249         if (t > 0.0f) {
250             q0f = FFMIN(q0f, t);
251             q1f = FFMAX(q1f, t);
252             qnrgf += t*t;
253             qcnt++;
254         }
255     }
256
257     if (!qcnt) {
258         memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
259         memset(sce->zeroes, 1, sizeof(sce->zeroes));
260         return;
261     }
262
263     //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
264     q0 = av_clip(coef2minsf(q0f), 0, SCALE_MAX_POS-1);
265     //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
266     q1 = av_clip(coef2maxsf(q1f), 1, SCALE_MAX_POS);
267     if (q1 - q0 > 60) {
268         int q0low  = q0;
269         int q1high = q1;
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);
272         q1 = qnrg + 30;
273         q0 = qnrg - 30;
274         if (q0 < q0low) {
275             q1 += q0low - q0;
276             q0  = q0low;
277         } else if (q1 > q1high) {
278             q0 -= q1 - q1high;
279             q1  = q1high;
280         }
281     }
282     // q0 == q1 isn't really a legal situation
283     if (q0 == q1) {
284         // the following is indirect but guarantees q1 != q0 && q1 near q0
285         q1 = av_clip(q0+1, 1, SCALE_MAX_POS);
286         q0 = av_clip(q1-1, 0, SCALE_MAX_POS - 1);
287     }
288
289     for (i = 0; i < TRELLIS_STATES; i++) {
290         paths[0][i].cost    = 0.0f;
291         paths[0][i].prev    = -1;
292     }
293     for (j = 1; j < TRELLIS_STAGES; j++) {
294         for (i = 0; i < TRELLIS_STATES; i++) {
295             paths[j][i].cost    = INFINITY;
296             paths[j][i].prev    = -2;
297         }
298     }
299     idx = 1;
300     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
301     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
302         start = w*128;
303         for (g = 0; g < sce->ics.num_swb; g++) {
304             const float *coefs = &sce->coeffs[start];
305             float qmin, qmax;
306             int nz = 0;
307
308             bandaddr[idx] = w * 16 + g;
309             qmin = INT_MAX;
310             qmax = 0.0f;
311             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
312                 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
313                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
314                     sce->zeroes[(w+w2)*16+g] = 1;
315                     continue;
316                 }
317                 sce->zeroes[(w+w2)*16+g] = 0;
318                 nz = 1;
319                 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
320                     float t = fabsf(coefs[w2*128+i]);
321                     if (t > 0.0f)
322                         qmin = FFMIN(qmin, t);
323                     qmax = FFMAX(qmax, t);
324                 }
325             }
326             if (nz) {
327                 int minscale, maxscale;
328                 float minrd = INFINITY;
329                 float maxval;
330                 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
331                 minscale = coef2minsf(qmin);
332                 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
333                 maxscale = coef2maxsf(qmax);
334                 minscale = av_clip(minscale - q0, 0, TRELLIS_STATES - 1);
335                 maxscale = av_clip(maxscale - q0, 0, TRELLIS_STATES);
336                 if (minscale == maxscale) {
337                     maxscale = av_clip(minscale+1, 1, TRELLIS_STATES);
338                     minscale = av_clip(maxscale-1, 0, TRELLIS_STATES - 1);
339                 }
340                 maxval = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], s->scoefs+start);
341                 for (q = minscale; q < maxscale; q++) {
342                     float dist = 0;
343                     int cb = find_min_book(maxval, sce->sf_idx[w*16+g]);
344                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
345                         FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
346                         dist += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
347                                                    q + q0, cb, lambda / band->threshold, INFINITY, NULL, NULL, 0);
348                     }
349                     minrd = FFMIN(minrd, dist);
350
351                     for (i = 0; i < q1 - q0; i++) {
352                         float cost;
353                         cost = paths[idx - 1][i].cost + dist
354                                + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
355                         if (cost < paths[idx][q].cost) {
356                             paths[idx][q].cost    = cost;
357                             paths[idx][q].prev    = i;
358                         }
359                     }
360                 }
361             } else {
362                 for (q = 0; q < q1 - q0; q++) {
363                     paths[idx][q].cost = paths[idx - 1][q].cost + 1;
364                     paths[idx][q].prev = q;
365                 }
366             }
367             sce->zeroes[w*16+g] = !nz;
368             start += sce->ics.swb_sizes[g];
369             idx++;
370         }
371     }
372     idx--;
373     mincost = paths[idx][0].cost;
374     minq    = 0;
375     for (i = 1; i < TRELLIS_STATES; i++) {
376         if (paths[idx][i].cost < mincost) {
377             mincost = paths[idx][i].cost;
378             minq = i;
379         }
380     }
381     while (idx) {
382         sce->sf_idx[bandaddr[idx]] = minq + q0;
383         minq = FFMAX(paths[idx][minq].prev, 0);
384         idx--;
385     }
386     //set the same quantizers inside window groups
387     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
388         for (g = 0;  g < sce->ics.num_swb; g++)
389             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
390                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
391 }
392
393
394 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
395                                        SingleChannelElement *sce,
396                                        const float lambda)
397 {
398     int start = 0, i, w, w2, g;
399     float uplim[128], maxq[128];
400     int minq, maxsf;
401     float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
402     int last = 0, lastband = 0, curband = 0;
403     float avg_energy = 0.0;
404     if (sce->ics.num_windows == 1) {
405         start = 0;
406         for (i = 0; i < 1024; i++) {
407             if (i - start >= sce->ics.swb_sizes[curband]) {
408                 start += sce->ics.swb_sizes[curband];
409                 curband++;
410             }
411             if (sce->coeffs[i]) {
412                 avg_energy += sce->coeffs[i] * sce->coeffs[i];
413                 last = i;
414                 lastband = curband;
415             }
416         }
417     } else {
418         for (w = 0; w < 8; w++) {
419             const float *coeffs = &sce->coeffs[w*128];
420             curband = start = 0;
421             for (i = 0; i < 128; i++) {
422                 if (i - start >= sce->ics.swb_sizes[curband]) {
423                     start += sce->ics.swb_sizes[curband];
424                     curband++;
425                 }
426                 if (coeffs[i]) {
427                     avg_energy += coeffs[i] * coeffs[i];
428                     last = FFMAX(last, i);
429                     lastband = FFMAX(lastband, curband);
430                 }
431             }
432         }
433     }
434     last++;
435     avg_energy /= last;
436     if (avg_energy == 0.0f) {
437         for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
438             sce->sf_idx[i] = SCALE_ONE_POS;
439         return;
440     }
441     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
442         start = w*128;
443         for (g = 0; g < sce->ics.num_swb; g++) {
444             float *coefs   = &sce->coeffs[start];
445             const int size = sce->ics.swb_sizes[g];
446             int start2 = start, end2 = start + size, peakpos = start;
447             float maxval = -1, thr = 0.0f, t;
448             maxq[w*16+g] = 0.0f;
449             if (g > lastband) {
450                 maxq[w*16+g] = 0.0f;
451                 start += size;
452                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
453                     memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
454                 continue;
455             }
456             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
457                 for (i = 0; i < size; i++) {
458                     float t = coefs[w2*128+i]*coefs[w2*128+i];
459                     maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
460                     thr += t;
461                     if (sce->ics.num_windows == 1 && maxval < t) {
462                         maxval  = t;
463                         peakpos = start+i;
464                     }
465                 }
466             }
467             if (sce->ics.num_windows == 1) {
468                 start2 = FFMAX(peakpos - 2, start2);
469                 end2   = FFMIN(peakpos + 3, end2);
470             } else {
471                 start2 -= start;
472                 end2   -= start;
473             }
474             start += size;
475             thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
476             t   = 1.0 - (1.0 * start2 / last);
477             uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
478         }
479     }
480     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
481     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
482     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
483         start = w*128;
484         for (g = 0;  g < sce->ics.num_swb; g++) {
485             const float *coefs  = &sce->coeffs[start];
486             const float *scaled = &s->scoefs[start];
487             const int size      = sce->ics.swb_sizes[g];
488             int scf, prev_scf, step;
489             int min_scf = -1, max_scf = 256;
490             float curdiff;
491             if (maxq[w*16+g] < 21.544) {
492                 sce->zeroes[w*16+g] = 1;
493                 start += size;
494                 continue;
495             }
496             sce->zeroes[w*16+g] = 0;
497             scf  = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2f(1/maxq[w*16+g])*16/3, 60, 218);
498             for (;;) {
499                 float dist = 0.0f;
500                 int quant_max;
501
502                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
503                     int b;
504                     dist += quantize_band_cost(s, coefs + w2*128,
505                                                scaled + w2*128,
506                                                sce->ics.swb_sizes[g],
507                                                scf,
508                                                ESC_BT,
509                                                lambda,
510                                                INFINITY,
511                                                &b, NULL,
512                                                0);
513                     dist -= b;
514                 }
515                 dist *= 1.0f / 512.0f / lambda;
516                 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[POW_SF2_ZERO - scf + SCALE_ONE_POS - SCALE_DIV_512], ROUND_STANDARD);
517                 if (quant_max >= 8191) { // too much, return to the previous quantizer
518                     sce->sf_idx[w*16+g] = prev_scf;
519                     break;
520                 }
521                 prev_scf = scf;
522                 curdiff = fabsf(dist - uplim[w*16+g]);
523                 if (curdiff <= 1.0f)
524                     step = 0;
525                 else
526                     step = log2f(curdiff);
527                 if (dist > uplim[w*16+g])
528                     step = -step;
529                 scf += step;
530                 scf = av_clip_uint8(scf);
531                 step = scf - prev_scf;
532                 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
533                     sce->sf_idx[w*16+g] = av_clip(scf, min_scf, max_scf);
534                     break;
535                 }
536                 if (step > 0)
537                     min_scf = prev_scf;
538                 else
539                     max_scf = prev_scf;
540             }
541             start += size;
542         }
543     }
544     minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
545     for (i = 1; i < 128; i++) {
546         if (!sce->sf_idx[i])
547             sce->sf_idx[i] = sce->sf_idx[i-1];
548         else
549             minq = FFMIN(minq, sce->sf_idx[i]);
550     }
551     if (minq == INT_MAX)
552         minq = 0;
553     minq = FFMIN(minq, SCALE_MAX_POS);
554     maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
555     for (i = 126; i >= 0; i--) {
556         if (!sce->sf_idx[i])
557             sce->sf_idx[i] = sce->sf_idx[i+1];
558         sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
559     }
560 }
561
562 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
563                                        SingleChannelElement *sce,
564                                        const float lambda)
565 {
566     int i, w, w2, g;
567     int minq = 255;
568
569     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
570     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
571         for (g = 0; g < sce->ics.num_swb; g++) {
572             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
573                 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
574                 if (band->energy <= band->threshold) {
575                     sce->sf_idx[(w+w2)*16+g] = 218;
576                     sce->zeroes[(w+w2)*16+g] = 1;
577                 } else {
578                     sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2f(band->threshold), 80, 218);
579                     sce->zeroes[(w+w2)*16+g] = 0;
580                 }
581                 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
582             }
583         }
584     }
585     for (i = 0; i < 128; i++) {
586         sce->sf_idx[i] = 140;
587         //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
588     }
589     //set the same quantizers inside window groups
590     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
591         for (g = 0;  g < sce->ics.num_swb; g++)
592             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
593                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
594 }
595
596 static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChannelElement *sce)
597 {
598     FFPsyBand *band;
599     int w, g, w2, i;
600     int wlen = 1024 / sce->ics.num_windows;
601     int bandwidth, cutoff;
602     float *PNS = &s->scoefs[0*128], *PNS34 = &s->scoefs[1*128];
603     float *NOR34 = &s->scoefs[3*128];
604     uint8_t nextband[128];
605     const float lambda = s->lambda;
606     const float freq_mult = avctx->sample_rate*0.5f/wlen;
607     const float thr_mult = NOISE_LAMBDA_REPLACE*(100.0f/lambda);
608     const float spread_threshold = FFMIN(0.75f, NOISE_SPREAD_THRESHOLD*FFMAX(0.5f, lambda/100.f));
609     const float dist_bias = av_clipf(4.f * 120 / lambda, 0.25f, 4.0f);
610     const float pns_transient_energy_r = FFMIN(0.7f, lambda / 140.f);
611
612     int refbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
613         / ((avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : avctx->channels)
614         * (lambda / 120.f);
615
616     /** Keep this in sync with twoloop's cutoff selection */
617     float rate_bandwidth_multiplier = 1.5f;
618     int prev = -1000, prev_sf = -1;
619     int frame_bit_rate = (avctx->flags & CODEC_FLAG_QSCALE)
620         ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
621         : (avctx->bit_rate / avctx->channels);
622
623     frame_bit_rate *= 1.15f;
624
625     if (avctx->cutoff > 0) {
626         bandwidth = avctx->cutoff;
627     } else {
628         bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
629     }
630
631     cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
632
633     memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type));
634     ff_init_nextband_map(sce, nextband);
635     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
636         int wstart = w*128;
637         for (g = 0;  g < sce->ics.num_swb; g++) {
638             int noise_sfi;
639             float dist1 = 0.0f, dist2 = 0.0f, noise_amp;
640             float pns_energy = 0.0f, pns_tgt_energy, energy_ratio, dist_thresh;
641             float sfb_energy = 0.0f, threshold = 0.0f, spread = 2.0f;
642             float min_energy = -1.0f, max_energy = 0.0f;
643             const int start = wstart+sce->ics.swb_offset[g];
644             const float freq = (start-wstart)*freq_mult;
645             const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
646             if (freq < NOISE_LOW_LIMIT || (start-wstart) >= cutoff) {
647                 if (!sce->zeroes[w*16+g])
648                     prev_sf = sce->sf_idx[w*16+g];
649                 continue;
650             }
651             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
652                 band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
653                 sfb_energy += band->energy;
654                 spread     = FFMIN(spread, band->spread);
655                 threshold  += band->threshold;
656                 if (!w2) {
657                     min_energy = max_energy = band->energy;
658                 } else {
659                     min_energy = FFMIN(min_energy, band->energy);
660                     max_energy = FFMAX(max_energy, band->energy);
661                 }
662             }
663
664             /* Ramps down at ~8000Hz and loosens the dist threshold */
665             dist_thresh = av_clipf(2.5f*NOISE_LOW_LIMIT/freq, 0.5f, 2.5f) * dist_bias;
666
667             /* PNS is acceptable when all of these are true:
668              * 1. high spread energy (noise-like band)
669              * 2. near-threshold energy (high PE means the random nature of PNS content will be noticed)
670              * 3. on short window groups, all windows have similar energy (variations in energy would be destroyed by PNS)
671              *
672              * At this stage, point 2 is relaxed for zeroed bands near the noise threshold (hole avoidance is more important)
673              */
674             if ((!sce->zeroes[w*16+g] && !ff_sfdelta_can_remove_band(sce, nextband, prev_sf, w*16+g)) ||
675                 ((sce->zeroes[w*16+g] || !sce->band_alt[w*16+g]) && sfb_energy < threshold*sqrtf(1.0f/freq_boost)) || spread < spread_threshold ||
676                 (!sce->zeroes[w*16+g] && sce->band_alt[w*16+g] && sfb_energy > threshold*thr_mult*freq_boost) ||
677                 min_energy < pns_transient_energy_r * max_energy ) {
678                 sce->pns_ener[w*16+g] = sfb_energy;
679                 if (!sce->zeroes[w*16+g])
680                     prev_sf = sce->sf_idx[w*16+g];
681                 continue;
682             }
683
684             pns_tgt_energy = sfb_energy*FFMIN(1.0f, spread*spread);
685             noise_sfi = av_clip(roundf(log2f(pns_tgt_energy)*2), -100, 155); /* Quantize */
686             noise_amp = -ff_aac_pow2sf_tab[noise_sfi + POW_SF2_ZERO];    /* Dequantize */
687             if (prev != -1000) {
688                 int noise_sfdiff = noise_sfi - prev + SCALE_DIFF_ZERO;
689                 if (noise_sfdiff < 0 || noise_sfdiff > 2*SCALE_MAX_DIFF) {
690                     if (!sce->zeroes[w*16+g])
691                         prev_sf = sce->sf_idx[w*16+g];
692                     continue;
693                 }
694             }
695             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
696                 float band_energy, scale, pns_senergy;
697                 const int start_c = (w+w2)*128+sce->ics.swb_offset[g];
698                 band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
699                 for (i = 0; i < sce->ics.swb_sizes[g]; i+=2) {
700                     double rnd[2];
701                     av_bmg_get(&s->lfg, rnd);
702                     PNS[i+0] = (float)rnd[0];
703                     PNS[i+1] = (float)rnd[1];
704                 }
705                 band_energy = s->fdsp->scalarproduct_float(PNS, PNS, sce->ics.swb_sizes[g]);
706                 scale = noise_amp/sqrtf(band_energy);
707                 s->fdsp->vector_fmul_scalar(PNS, PNS, scale, sce->ics.swb_sizes[g]);
708                 pns_senergy = s->fdsp->scalarproduct_float(PNS, PNS, sce->ics.swb_sizes[g]);
709                 pns_energy += pns_senergy;
710                 abs_pow34_v(NOR34, &sce->coeffs[start_c], sce->ics.swb_sizes[g]);
711                 abs_pow34_v(PNS34, PNS, sce->ics.swb_sizes[g]);
712                 dist1 += quantize_band_cost(s, &sce->coeffs[start_c],
713                                             NOR34,
714                                             sce->ics.swb_sizes[g],
715                                             sce->sf_idx[(w+w2)*16+g],
716                                             sce->band_alt[(w+w2)*16+g],
717                                             lambda/band->threshold, INFINITY, NULL, NULL, 0);
718                 /* Estimate rd on average as 5 bits for SF, 4 for the CB, plus spread energy * lambda/thr */
719                 dist2 += band->energy/(band->spread*band->spread)*lambda*dist_thresh/band->threshold;
720             }
721             if (g && sce->band_type[w*16+g-1] == NOISE_BT) {
722                 dist2 += 5;
723             } else {
724                 dist2 += 9;
725             }
726             energy_ratio = pns_tgt_energy/pns_energy; /* Compensates for quantization error */
727             sce->pns_ener[w*16+g] = energy_ratio*pns_tgt_energy;
728             if (sce->zeroes[w*16+g] || !sce->band_alt[w*16+g] || (energy_ratio > 0.85f && energy_ratio < 1.25f && dist2 < dist1)) {
729                 sce->band_type[w*16+g] = NOISE_BT;
730                 sce->zeroes[w*16+g] = 0;
731                 prev = noise_sfi;
732             } else {
733                 if (!sce->zeroes[w*16+g])
734                     prev_sf = sce->sf_idx[w*16+g];
735             }
736         }
737     }
738 }
739
740 static void mark_pns(AACEncContext *s, AVCodecContext *avctx, SingleChannelElement *sce)
741 {
742     FFPsyBand *band;
743     int w, g, w2;
744     int wlen = 1024 / sce->ics.num_windows;
745     int bandwidth, cutoff;
746     const float lambda = s->lambda;
747     const float freq_mult = avctx->sample_rate*0.5f/wlen;
748     const float spread_threshold = FFMIN(0.75f, NOISE_SPREAD_THRESHOLD*FFMAX(0.5f, lambda/100.f));
749     const float pns_transient_energy_r = FFMIN(0.7f, lambda / 140.f);
750
751     int refbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
752         / ((avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : avctx->channels)
753         * (lambda / 120.f);
754
755     /** Keep this in sync with twoloop's cutoff selection */
756     float rate_bandwidth_multiplier = 1.5f;
757     int frame_bit_rate = (avctx->flags & CODEC_FLAG_QSCALE)
758         ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
759         : (avctx->bit_rate / avctx->channels);
760
761     frame_bit_rate *= 1.15f;
762
763     if (avctx->cutoff > 0) {
764         bandwidth = avctx->cutoff;
765     } else {
766         bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
767     }
768
769     cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
770
771     memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type));
772     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
773         for (g = 0;  g < sce->ics.num_swb; g++) {
774             float sfb_energy = 0.0f, threshold = 0.0f, spread = 2.0f;
775             float min_energy = -1.0f, max_energy = 0.0f;
776             const int start = sce->ics.swb_offset[g];
777             const float freq = start*freq_mult;
778             const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
779             if (freq < NOISE_LOW_LIMIT || start >= cutoff) {
780                 sce->can_pns[w*16+g] = 0;
781                 continue;
782             }
783             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
784                 band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
785                 sfb_energy += band->energy;
786                 spread     = FFMIN(spread, band->spread);
787                 threshold  += band->threshold;
788                 if (!w2) {
789                     min_energy = max_energy = band->energy;
790                 } else {
791                     min_energy = FFMIN(min_energy, band->energy);
792                     max_energy = FFMAX(max_energy, band->energy);
793                 }
794             }
795
796             /* PNS is acceptable when all of these are true:
797              * 1. high spread energy (noise-like band)
798              * 2. near-threshold energy (high PE means the random nature of PNS content will be noticed)
799              * 3. on short window groups, all windows have similar energy (variations in energy would be destroyed by PNS)
800              */
801             sce->pns_ener[w*16+g] = sfb_energy;
802             if (sfb_energy < threshold*sqrtf(1.5f/freq_boost) || spread < spread_threshold || min_energy < pns_transient_energy_r * max_energy) {
803                 sce->can_pns[w*16+g] = 0;
804             } else {
805                 sce->can_pns[w*16+g] = 1;
806             }
807         }
808     }
809 }
810
811 static void search_for_ms(AACEncContext *s, ChannelElement *cpe)
812 {
813     int start = 0, i, w, w2, g, sid_sf_boost, prev_mid, prev_side;
814     uint8_t nextband0[128], nextband1[128];
815     float M[128], S[128];
816     float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
817     const float lambda = s->lambda;
818     const float mslambda = FFMIN(1.0f, lambda / 120.f);
819     SingleChannelElement *sce0 = &cpe->ch[0];
820     SingleChannelElement *sce1 = &cpe->ch[1];
821     if (!cpe->common_window)
822         return;
823
824     /** Scout out next nonzero bands */
825     ff_init_nextband_map(sce0, nextband0);
826     ff_init_nextband_map(sce1, nextband1);
827
828     prev_mid = sce0->sf_idx[0];
829     prev_side = sce1->sf_idx[0];
830     for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
831         start = 0;
832         for (g = 0;  g < sce0->ics.num_swb; g++) {
833             float bmax = bval2bmax(g * 17.0f / sce0->ics.num_swb) / 0.0045f;
834             if (!cpe->is_mask[w*16+g])
835                 cpe->ms_mask[w*16+g] = 0;
836             if (!sce0->zeroes[w*16+g] && !sce1->zeroes[w*16+g] && !cpe->is_mask[w*16+g]) {
837                 float Mmax = 0.0f, Smax = 0.0f;
838
839                 /* Must compute mid/side SF and book for the whole window group */
840                 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
841                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
842                         M[i] = (sce0->coeffs[start+(w+w2)*128+i]
843                               + sce1->coeffs[start+(w+w2)*128+i]) * 0.5;
844                         S[i] =  M[i]
845                               - sce1->coeffs[start+(w+w2)*128+i];
846                     }
847                     abs_pow34_v(M34, M, sce0->ics.swb_sizes[g]);
848                     abs_pow34_v(S34, S, sce0->ics.swb_sizes[g]);
849                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++ ) {
850                         Mmax = FFMAX(Mmax, M34[i]);
851                         Smax = FFMAX(Smax, S34[i]);
852                     }
853                 }
854
855                 for (sid_sf_boost = 0; sid_sf_boost < 4; sid_sf_boost++) {
856                     float dist1 = 0.0f, dist2 = 0.0f;
857                     int B0 = 0, B1 = 0;
858                     int minidx;
859                     int mididx, sididx;
860                     int midcb, sidcb;
861
862                     minidx = FFMIN(sce0->sf_idx[w*16+g], sce1->sf_idx[w*16+g]);
863                     mididx = av_clip(minidx, 0, SCALE_MAX_POS - SCALE_DIV_512);
864                     sididx = av_clip(minidx - sid_sf_boost * 3, 0, SCALE_MAX_POS - SCALE_DIV_512);
865                     if (sce0->band_type[w*16+g] != NOISE_BT && sce1->band_type[w*16+g] != NOISE_BT
866                         && (   !ff_sfdelta_can_replace(sce0, nextband0, prev_mid, mididx, w*16+g)
867                             || !ff_sfdelta_can_replace(sce1, nextband1, prev_side, sididx, w*16+g))) {
868                         /* scalefactor range violation, bad stuff, will decrease quality unacceptably */
869                         continue;
870                     }
871
872                     midcb = find_min_book(Mmax, mididx);
873                     sidcb = find_min_book(Smax, sididx);
874
875                     /* No CB can be zero */
876                     midcb = FFMAX(1,midcb);
877                     sidcb = FFMAX(1,sidcb);
878
879                     for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
880                         FFPsyBand *band0 = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g];
881                         FFPsyBand *band1 = &s->psy.ch[s->cur_channel+1].psy_bands[(w+w2)*16+g];
882                         float minthr = FFMIN(band0->threshold, band1->threshold);
883                         int b1,b2,b3,b4;
884                         for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
885                             M[i] = (sce0->coeffs[start+(w+w2)*128+i]
886                                   + sce1->coeffs[start+(w+w2)*128+i]) * 0.5;
887                             S[i] =  M[i]
888                                   - sce1->coeffs[start+(w+w2)*128+i];
889                         }
890
891                         abs_pow34_v(L34, sce0->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
892                         abs_pow34_v(R34, sce1->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
893                         abs_pow34_v(M34, M,                         sce0->ics.swb_sizes[g]);
894                         abs_pow34_v(S34, S,                         sce0->ics.swb_sizes[g]);
895                         dist1 += quantize_band_cost(s, &sce0->coeffs[start + (w+w2)*128],
896                                                     L34,
897                                                     sce0->ics.swb_sizes[g],
898                                                     sce0->sf_idx[w*16+g],
899                                                     sce0->band_type[w*16+g],
900                                                     lambda / band0->threshold, INFINITY, &b1, NULL, 0);
901                         dist1 += quantize_band_cost(s, &sce1->coeffs[start + (w+w2)*128],
902                                                     R34,
903                                                     sce1->ics.swb_sizes[g],
904                                                     sce1->sf_idx[w*16+g],
905                                                     sce1->band_type[w*16+g],
906                                                     lambda / band1->threshold, INFINITY, &b2, NULL, 0);
907                         dist2 += quantize_band_cost(s, M,
908                                                     M34,
909                                                     sce0->ics.swb_sizes[g],
910                                                     mididx,
911                                                     midcb,
912                                                     lambda / minthr, INFINITY, &b3, NULL, 0);
913                         dist2 += quantize_band_cost(s, S,
914                                                     S34,
915                                                     sce1->ics.swb_sizes[g],
916                                                     sididx,
917                                                     sidcb,
918                                                     mslambda / (minthr * bmax), INFINITY, &b4, NULL, 0);
919                         B0 += b1+b2;
920                         B1 += b3+b4;
921                         dist1 -= b1+b2;
922                         dist2 -= b3+b4;
923                     }
924                     cpe->ms_mask[w*16+g] = dist2 <= dist1 && B1 < B0;
925                     if (cpe->ms_mask[w*16+g]) {
926                         if (sce0->band_type[w*16+g] != NOISE_BT && sce1->band_type[w*16+g] != NOISE_BT) {
927                             sce0->sf_idx[w*16+g] = mididx;
928                             sce1->sf_idx[w*16+g] = sididx;
929                             sce0->band_type[w*16+g] = midcb;
930                             sce1->band_type[w*16+g] = sidcb;
931                         } else if ((sce0->band_type[w*16+g] != NOISE_BT) ^ (sce1->band_type[w*16+g] != NOISE_BT)) {
932                             /* ms_mask unneeded, and it confuses some decoders */
933                             cpe->ms_mask[w*16+g] = 0;
934                         }
935                         break;
936                     } else if (B1 > B0) {
937                         /* More boost won't fix this */
938                         break;
939                     }
940                 }
941             }
942             if (!sce0->zeroes[w*16+g] && sce0->band_type[w*16+g] < RESERVED_BT)
943                 prev_mid = sce0->sf_idx[w*16+g];
944             if (!sce1->zeroes[w*16+g] && !cpe->is_mask[w*16+g] && sce1->band_type[w*16+g] < RESERVED_BT)
945                 prev_side = sce1->sf_idx[w*16+g];
946             start += sce0->ics.swb_sizes[g];
947         }
948     }
949 }
950
951 AACCoefficientsEncoder ff_aac_coders[AAC_CODER_NB] = {
952     [AAC_CODER_FAAC] = {
953         search_for_quantizers_faac,
954         encode_window_bands_info,
955         quantize_and_encode_band,
956         ff_aac_encode_tns_info,
957         ff_aac_encode_ltp_info,
958         ff_aac_encode_main_pred,
959         ff_aac_adjust_common_pred,
960         ff_aac_adjust_common_ltp,
961         ff_aac_apply_main_pred,
962         ff_aac_apply_tns,
963         ff_aac_update_ltp,
964         ff_aac_ltp_insert_new_frame,
965         set_special_band_scalefactors,
966         search_for_pns,
967         mark_pns,
968         ff_aac_search_for_tns,
969         ff_aac_search_for_ltp,
970         search_for_ms,
971         ff_aac_search_for_is,
972         ff_aac_search_for_pred,
973     },
974     [AAC_CODER_ANMR] = {
975         search_for_quantizers_anmr,
976         encode_window_bands_info,
977         quantize_and_encode_band,
978         ff_aac_encode_tns_info,
979         ff_aac_encode_ltp_info,
980         ff_aac_encode_main_pred,
981         ff_aac_adjust_common_pred,
982         ff_aac_adjust_common_ltp,
983         ff_aac_apply_main_pred,
984         ff_aac_apply_tns,
985         ff_aac_update_ltp,
986         ff_aac_ltp_insert_new_frame,
987         set_special_band_scalefactors,
988         search_for_pns,
989         mark_pns,
990         ff_aac_search_for_tns,
991         ff_aac_search_for_ltp,
992         search_for_ms,
993         ff_aac_search_for_is,
994         ff_aac_search_for_pred,
995     },
996     [AAC_CODER_TWOLOOP] = {
997         search_for_quantizers_twoloop,
998         codebook_trellis_rate,
999         quantize_and_encode_band,
1000         ff_aac_encode_tns_info,
1001         ff_aac_encode_ltp_info,
1002         ff_aac_encode_main_pred,
1003         ff_aac_adjust_common_pred,
1004         ff_aac_adjust_common_ltp,
1005         ff_aac_apply_main_pred,
1006         ff_aac_apply_tns,
1007         ff_aac_update_ltp,
1008         ff_aac_ltp_insert_new_frame,
1009         set_special_band_scalefactors,
1010         search_for_pns,
1011         mark_pns,
1012         ff_aac_search_for_tns,
1013         ff_aac_search_for_ltp,
1014         search_for_ms,
1015         ff_aac_search_for_is,
1016         ff_aac_search_for_pred,
1017     },
1018     [AAC_CODER_FAST] = {
1019         search_for_quantizers_fast,
1020         encode_window_bands_info,
1021         quantize_and_encode_band,
1022         ff_aac_encode_tns_info,
1023         ff_aac_encode_ltp_info,
1024         ff_aac_encode_main_pred,
1025         ff_aac_adjust_common_pred,
1026         ff_aac_adjust_common_ltp,
1027         ff_aac_apply_main_pred,
1028         ff_aac_apply_tns,
1029         ff_aac_update_ltp,
1030         ff_aac_ltp_insert_new_frame,
1031         set_special_band_scalefactors,
1032         search_for_pns,
1033         mark_pns,
1034         ff_aac_search_for_tns,
1035         ff_aac_search_for_ltp,
1036         search_for_ms,
1037         ff_aac_search_for_is,
1038         ff_aac_search_for_pred,
1039     },
1040 };