aacenc: Split paths in the scalefactor selection trellis into a 2-D array.
[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 libavcodec/aaccoder.c
24  * AAC coefficients encoder
25  */
26
27 /***********************************
28  *              TODOs:
29  * speedup quantizer selection
30  * add sane pulse detection
31  ***********************************/
32
33 #include "avcodec.h"
34 #include "put_bits.h"
35 #include "aac.h"
36 #include "aacenc.h"
37 #include "aactab.h"
38
39 /** bits needed to code codebook run value for long windows */
40 static const uint8_t run_value_bits_long[64] = {
41      5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,
42      5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5, 10,
43     10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
44     10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 15
45 };
46
47 /** bits needed to code codebook run value for short windows */
48 static const uint8_t run_value_bits_short[16] = {
49     3, 3, 3, 3, 3, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 9
50 };
51
52 static const uint8_t *run_value_bits[2] = {
53     run_value_bits_long, run_value_bits_short
54 };
55
56
57 /**
58  * Quantize one coefficient.
59  * @return absolute value of the quantized coefficient
60  * @see 3GPP TS26.403 5.6.2 "Scalefactor determination"
61  */
62 static av_always_inline int quant(float coef, const float Q)
63 {
64     float a = coef * Q;
65     return sqrtf(a * sqrtf(a)) + 0.4054;
66 }
67
68 static void quantize_bands(int (*out)[2], const float *in, const float *scaled,
69                            int size, float Q34, int is_signed, int maxval)
70 {
71     int i;
72     double qc;
73     for (i = 0; i < size; i++) {
74         qc = scaled[i] * Q34;
75         out[i][0] = (int)FFMIN(qc,          (double)maxval);
76         out[i][1] = (int)FFMIN(qc + 0.4054, (double)maxval);
77         if (is_signed && in[i] < 0.0f) {
78             out[i][0] = -out[i][0];
79             out[i][1] = -out[i][1];
80         }
81     }
82 }
83
84 static void abs_pow34_v(float *out, const float *in, const int size)
85 {
86 #ifndef USE_REALLY_FULL_SEARCH
87     int i;
88     for (i = 0; i < size; i++) {
89         float a = fabsf(in[i]);
90         out[i] = sqrtf(a * sqrtf(a));
91     }
92 #endif /* USE_REALLY_FULL_SEARCH */
93 }
94
95 static const uint8_t aac_cb_range [12] = {0, 3, 3, 3, 3, 9, 9, 8, 8, 13, 13, 17};
96 static const uint8_t aac_cb_maxval[12] = {0, 1, 1, 2, 2, 4, 4, 7, 7, 12, 12, 16};
97
98 /**
99  * Calculate rate distortion cost for quantizing with given codebook
100  *
101  * @return quantization distortion
102  */
103 static float quantize_band_cost(struct AACEncContext *s, const float *in,
104                                 const float *scaled, int size, int scale_idx,
105                                 int cb, const float lambda, const float uplim,
106                                 int *bits)
107 {
108     const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
109     const float  Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
110     const float CLIPPED_ESCAPE = 165140.0f*IQ;
111     int i, j, k;
112     float cost = 0;
113     const int dim = cb < FIRST_PAIR_BT ? 4 : 2;
114     int resbits = 0;
115 #ifndef USE_REALLY_FULL_SEARCH
116     const float  Q34 = sqrtf(Q * sqrtf(Q));
117     const int range  = aac_cb_range[cb];
118     const int maxval = aac_cb_maxval[cb];
119     int offs[4];
120 #endif /* USE_REALLY_FULL_SEARCH */
121
122     if (!cb) {
123         for (i = 0; i < size; i++)
124             cost += in[i]*in[i];
125         if (bits)
126             *bits = 0;
127         return cost * lambda;
128     }
129 #ifndef USE_REALLY_FULL_SEARCH
130     offs[0] = 1;
131     for (i = 1; i < dim; i++)
132         offs[i] = offs[i-1]*range;
133     quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
134 #endif /* USE_REALLY_FULL_SEARCH */
135     for (i = 0; i < size; i += dim) {
136         float mincost;
137         int minidx  = 0;
138         int minbits = 0;
139         const float *vec;
140 #ifndef USE_REALLY_FULL_SEARCH
141         int (*quants)[2] = &s->qcoefs[i];
142         mincost = 0.0f;
143         for (j = 0; j < dim; j++)
144             mincost += in[i+j]*in[i+j];
145         minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
146         minbits = ff_aac_spectral_bits[cb-1][minidx];
147         mincost = mincost * lambda + minbits;
148         for (j = 0; j < (1<<dim); j++) {
149             float rd = 0.0f;
150             int curbits;
151             int curidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
152             int same   = 0;
153             for (k = 0; k < dim; k++) {
154                 if ((j & (1 << k)) && quants[k][0] == quants[k][1]) {
155                     same = 1;
156                     break;
157                 }
158             }
159             if (same)
160                 continue;
161             for (k = 0; k < dim; k++)
162                 curidx += quants[k][!!(j & (1 << k))] * offs[dim - 1 - k];
163             curbits =  ff_aac_spectral_bits[cb-1][curidx];
164             vec     = &ff_aac_codebook_vectors[cb-1][curidx*dim];
165 #else
166         mincost = INFINITY;
167         vec = ff_aac_codebook_vectors[cb-1];
168         for (j = 0; j < ff_aac_spectral_sizes[cb-1]; j++, vec += dim) {
169             float rd = 0.0f;
170             int curbits = ff_aac_spectral_bits[cb-1][j];
171 #endif /* USE_REALLY_FULL_SEARCH */
172             if (IS_CODEBOOK_UNSIGNED(cb)) {
173                 for (k = 0; k < dim; k++) {
174                     float t = fabsf(in[i+k]);
175                     float di;
176                     if (vec[k] == 64.0f) { //FIXME: slow
177                         //do not code with escape sequence small values
178                         if (t < 39.0f*IQ) {
179                             rd = INFINITY;
180                             break;
181                         }
182                         if (t >= CLIPPED_ESCAPE) {
183                             di = t - CLIPPED_ESCAPE;
184                             curbits += 21;
185                         } else {
186                             int c = av_clip(quant(t, Q), 0, 8191);
187                             di = t - c*cbrtf(c)*IQ;
188                             curbits += av_log2(c)*2 - 4 + 1;
189                         }
190                     } else {
191                         di = t - vec[k]*IQ;
192                     }
193                     if (vec[k] != 0.0f)
194                         curbits++;
195                     rd += di*di;
196                 }
197             } else {
198                 for (k = 0; k < dim; k++) {
199                     float di = in[i+k] - vec[k]*IQ;
200                     rd += di*di;
201                 }
202             }
203             rd = rd * lambda + curbits;
204             if (rd < mincost) {
205                 mincost = rd;
206                 minidx  = j;
207                 minbits = curbits;
208             }
209         }
210         cost    += mincost;
211         resbits += minbits;
212         if (cost >= uplim)
213             return uplim;
214     }
215
216     if (bits)
217         *bits = resbits;
218     return cost;
219 }
220
221 static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
222                                      const float *in, int size, int scale_idx,
223                                      int cb, const float lambda)
224 {
225     const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
226     const float  Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
227     const float CLIPPED_ESCAPE = 165140.0f*IQ;
228     const int dim = (cb < FIRST_PAIR_BT) ? 4 : 2;
229     int i, j, k;
230 #ifndef USE_REALLY_FULL_SEARCH
231     const float  Q34 = sqrtf(Q * sqrtf(Q));
232     const int range  = aac_cb_range[cb];
233     const int maxval = aac_cb_maxval[cb];
234     int offs[4];
235     float *scaled = s->scoefs;
236 #endif /* USE_REALLY_FULL_SEARCH */
237
238 //START_TIMER
239     if (!cb)
240         return;
241
242 #ifndef USE_REALLY_FULL_SEARCH
243     offs[0] = 1;
244     for (i = 1; i < dim; i++)
245         offs[i] = offs[i-1]*range;
246     abs_pow34_v(scaled, in, size);
247     quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
248 #endif /* USE_REALLY_FULL_SEARCH */
249     for (i = 0; i < size; i += dim) {
250         float mincost;
251         int minidx  = 0;
252         int minbits = 0;
253         const float *vec;
254 #ifndef USE_REALLY_FULL_SEARCH
255         int (*quants)[2] = &s->qcoefs[i];
256         mincost = 0.0f;
257         for (j = 0; j < dim; j++)
258             mincost += in[i+j]*in[i+j];
259         minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
260         minbits = ff_aac_spectral_bits[cb-1][minidx];
261         mincost = mincost * lambda + minbits;
262         for (j = 0; j < (1<<dim); j++) {
263             float rd = 0.0f;
264             int curbits;
265             int curidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
266             int same   = 0;
267             for (k = 0; k < dim; k++) {
268                 if ((j & (1 << k)) && quants[k][0] == quants[k][1]) {
269                     same = 1;
270                     break;
271                 }
272             }
273             if (same)
274                 continue;
275             for (k = 0; k < dim; k++)
276                 curidx += quants[k][!!(j & (1 << k))] * offs[dim - 1 - k];
277             curbits =  ff_aac_spectral_bits[cb-1][curidx];
278             vec     = &ff_aac_codebook_vectors[cb-1][curidx*dim];
279 #else
280         vec = ff_aac_codebook_vectors[cb-1];
281         mincost = INFINITY;
282         for (j = 0; j < ff_aac_spectral_sizes[cb-1]; j++, vec += dim) {
283             float rd = 0.0f;
284             int curbits = ff_aac_spectral_bits[cb-1][j];
285             int curidx  = j;
286 #endif /* USE_REALLY_FULL_SEARCH */
287             if (IS_CODEBOOK_UNSIGNED(cb)) {
288                 for (k = 0; k < dim; k++) {
289                     float t = fabsf(in[i+k]);
290                     float di;
291                     if (vec[k] == 64.0f) { //FIXME: slow
292                         //do not code with escape sequence small values
293                         if (t < 39.0f*IQ) {
294                             rd = INFINITY;
295                             break;
296                         }
297                         if (t >= CLIPPED_ESCAPE) {
298                             di = t - CLIPPED_ESCAPE;
299                             curbits += 21;
300                         } else {
301                             int c = av_clip(quant(t, Q), 0, 8191);
302                             di = t - c*cbrtf(c)*IQ;
303                             curbits += av_log2(c)*2 - 4 + 1;
304                         }
305                     } else {
306                         di = t - vec[k]*IQ;
307                     }
308                     if (vec[k] != 0.0f)
309                         curbits++;
310                     rd += di*di;
311                 }
312             } else {
313                 for (k = 0; k < dim; k++) {
314                     float di = in[i+k] - vec[k]*IQ;
315                     rd += di*di;
316                 }
317             }
318             rd = rd * lambda + curbits;
319             if (rd < mincost) {
320                 mincost = rd;
321                 minidx  = curidx;
322                 minbits = curbits;
323             }
324         }
325         put_bits(pb, ff_aac_spectral_bits[cb-1][minidx], ff_aac_spectral_codes[cb-1][minidx]);
326         if (IS_CODEBOOK_UNSIGNED(cb))
327             for (j = 0; j < dim; j++)
328                 if (ff_aac_codebook_vectors[cb-1][minidx*dim+j] != 0.0f)
329                     put_bits(pb, 1, in[i+j] < 0.0f);
330         if (cb == ESC_BT) {
331             for (j = 0; j < 2; j++) {
332                 if (ff_aac_codebook_vectors[cb-1][minidx*2+j] == 64.0f) {
333                     int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
334                     int len = av_log2(coef);
335
336                     put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
337                     put_bits(pb, len, coef & ((1 << len) - 1));
338                 }
339             }
340         }
341     }
342 //STOP_TIMER("quantize_and_encode")
343 }
344
345 /**
346  * structure used in optimal codebook search
347  */
348 typedef struct BandCodingPath {
349     int prev_idx; ///< pointer to the previous path point
350     float cost;   ///< path cost
351     int run;
352 } BandCodingPath;
353
354 /**
355  * Encode band info for single window group bands.
356  */
357 static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
358                                      int win, int group_len, const float lambda)
359 {
360     BandCodingPath path[120][12];
361     int w, swb, cb, start, start2, size;
362     int i, j;
363     const int max_sfb  = sce->ics.max_sfb;
364     const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
365     const int run_esc  = (1 << run_bits) - 1;
366     int idx, ppos, count;
367     int stackrun[120], stackcb[120], stack_len;
368     float next_minrd = INFINITY;
369     int next_mincb = 0;
370
371     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
372     start = win*128;
373     for (cb = 0; cb < 12; cb++) {
374         path[0][cb].cost     = 0.0f;
375         path[0][cb].prev_idx = -1;
376         path[0][cb].run      = 0;
377     }
378     for (swb = 0; swb < max_sfb; swb++) {
379         start2 = start;
380         size = sce->ics.swb_sizes[swb];
381         if (sce->zeroes[win*16 + swb]) {
382             for (cb = 0; cb < 12; cb++) {
383                 path[swb+1][cb].prev_idx = cb;
384                 path[swb+1][cb].cost     = path[swb][cb].cost;
385                 path[swb+1][cb].run      = path[swb][cb].run + 1;
386             }
387         } else {
388             float minrd = next_minrd;
389             int mincb = next_mincb;
390             next_minrd = INFINITY;
391             next_mincb = 0;
392             for (cb = 0; cb < 12; cb++) {
393                 float cost_stay_here, cost_get_here;
394                 float rd = 0.0f;
395                 for (w = 0; w < group_len; w++) {
396                     FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(win+w)*16+swb];
397                     rd += quantize_band_cost(s, sce->coeffs + start + w*128,
398                                              s->scoefs + start + w*128, size,
399                                              sce->sf_idx[(win+w)*16+swb], cb,
400                                              lambda / band->threshold, INFINITY, NULL);
401                 }
402                 cost_stay_here = path[swb][cb].cost + rd;
403                 cost_get_here  = minrd              + rd + run_bits + 4;
404                 if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
405                     != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
406                     cost_stay_here += run_bits;
407                 if (cost_get_here < cost_stay_here) {
408                     path[swb+1][cb].prev_idx = mincb;
409                     path[swb+1][cb].cost     = cost_get_here;
410                     path[swb+1][cb].run      = 1;
411                 } else {
412                     path[swb+1][cb].prev_idx = cb;
413                     path[swb+1][cb].cost     = cost_stay_here;
414                     path[swb+1][cb].run      = path[swb][cb].run + 1;
415                 }
416                 if (path[swb+1][cb].cost < next_minrd) {
417                     next_minrd = path[swb+1][cb].cost;
418                     next_mincb = cb;
419                 }
420             }
421         }
422         start += sce->ics.swb_sizes[swb];
423     }
424
425     //convert resulting path from backward-linked list
426     stack_len = 0;
427     idx       = 0;
428     for (cb = 1; cb < 12; cb++)
429         if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
430             idx = cb;
431     ppos = max_sfb;
432     while (ppos > 0) {
433         cb = idx;
434         stackrun[stack_len] = path[ppos][cb].run;
435         stackcb [stack_len] = cb;
436         idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
437         ppos -= path[ppos][cb].run;
438         stack_len++;
439     }
440     //perform actual band info encoding
441     start = 0;
442     for (i = stack_len - 1; i >= 0; i--) {
443         put_bits(&s->pb, 4, stackcb[i]);
444         count = stackrun[i];
445         memset(sce->zeroes + win*16 + start, !stackcb[i], count);
446         //XXX: memset when band_type is also uint8_t
447         for (j = 0; j < count; j++) {
448             sce->band_type[win*16 + start] =  stackcb[i];
449             start++;
450         }
451         while (count >= run_esc) {
452             put_bits(&s->pb, run_bits, run_esc);
453             count -= run_esc;
454         }
455         put_bits(&s->pb, run_bits, count);
456     }
457 }
458
459 typedef struct TrellisPath {
460     float cost;
461     int prev;
462     int min_val;
463     int max_val;
464 } TrellisPath;
465
466 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
467                                        SingleChannelElement *sce,
468                                        const float lambda)
469 {
470     int q, w, w2, g, start = 0;
471     int i, j;
472     int idx;
473     TrellisPath paths[121][256];
474     int bandaddr[121];
475     int minq;
476     float mincost;
477
478     for (i = 0; i < 256; i++) {
479         paths[0][i].cost    = 0.0f;
480         paths[0][i].prev    = -1;
481         paths[0][i].min_val = i;
482         paths[0][i].max_val = i;
483     }
484     for (j = 1; j < 121; j++) {
485         for (i = 0; i < 256; i++) {
486             paths[j][i].cost    = INFINITY;
487             paths[j][i].prev    = -2;
488             paths[j][i].min_val = INT_MAX;
489             paths[j][i].max_val = 0;
490         }
491     }
492     idx = 1;
493     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
494     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
495         start = w*128;
496         for (g = 0; g < sce->ics.num_swb; g++) {
497             const float *coefs = sce->coeffs + start;
498             float qmin, qmax;
499             int nz = 0;
500
501             bandaddr[idx] = w * 16 + g;
502             qmin = INT_MAX;
503             qmax = 0.0f;
504             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
505                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
506                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
507                     sce->zeroes[(w+w2)*16+g] = 1;
508                     continue;
509                 }
510                 sce->zeroes[(w+w2)*16+g] = 0;
511                 nz = 1;
512                 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
513                     float t = fabsf(coefs[w2*128+i]);
514                     if (t > 0.0f)
515                         qmin = FFMIN(qmin, t);
516                     qmax = FFMAX(qmax, t);
517                 }
518             }
519             if (nz) {
520                 int minscale, maxscale;
521                 float minrd = INFINITY;
522                 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
523                 minscale = av_clip_uint8(log2(qmin)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
524                 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
525                 maxscale = av_clip_uint8(log2(qmax)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
526                 for (q = minscale; q < maxscale; q++) {
527                     float dists[12], dist;
528                     memset(dists, 0, sizeof(dists));
529                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
530                         FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
531                         int cb;
532                         for (cb = 0; cb <= ESC_BT; cb++)
533                             dists[cb] += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
534                                                             q, cb, lambda / band->threshold, INFINITY, NULL);
535                     }
536                     dist = dists[0];
537                     for (i = 1; i <= ESC_BT; i++)
538                         dist = FFMIN(dist, dists[i]);
539                     minrd = FFMIN(minrd, dist);
540
541                     for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, 256); i++) {
542                         float cost;
543                         int minv, maxv;
544                         if (isinf(paths[idx - 1][i].cost))
545                             continue;
546                         cost = paths[idx - 1][i].cost + dist
547                                + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
548                         minv = FFMIN(paths[idx - 1][i].min_val, q);
549                         maxv = FFMAX(paths[idx - 1][i].max_val, q);
550                         if (cost < paths[idx][q].cost && maxv-minv < SCALE_MAX_DIFF) {
551                             paths[idx][q].cost    = cost;
552                             paths[idx][q].prev    = i;
553                             paths[idx][q].min_val = minv;
554                             paths[idx][q].max_val = maxv;
555                         }
556                     }
557                 }
558             } else {
559                 for (q = 0; q < 256; q++) {
560                     if (!isinf(paths[idx - 1][q].cost)) {
561                         paths[idx][q].cost = paths[idx - 1][q].cost + 1;
562                         paths[idx][q].prev = q;
563                         paths[idx][q].min_val = FFMIN(paths[idx - 1][q].min_val, q);
564                         paths[idx][q].max_val = FFMAX(paths[idx - 1][q].max_val, q);
565                         continue;
566                     }
567                     for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, 256); i++) {
568                         float cost;
569                         int minv, maxv;
570                         if (isinf(paths[idx - 1][i].cost))
571                             continue;
572                         cost = paths[idx - 1][i].cost + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
573                         minv = FFMIN(paths[idx - 1][i].min_val, q);
574                         maxv = FFMAX(paths[idx - 1][i].max_val, q);
575                         if (cost < paths[idx][q].cost && maxv-minv < SCALE_MAX_DIFF) {
576                             paths[idx][q].cost    = cost;
577                             paths[idx][q].prev    = i;
578                             paths[idx][q].min_val = minv;
579                             paths[idx][q].max_val = maxv;
580                         }
581                     }
582                 }
583             }
584             sce->zeroes[w*16+g] = !nz;
585             start += sce->ics.swb_sizes[g];
586             idx++;
587         }
588     }
589     idx--;
590     mincost = paths[idx][0].cost;
591     minq    = 0;
592     for (i = 1; i < 256; i++) {
593         if (paths[idx][i].cost < mincost) {
594             mincost = paths[idx][i].cost;
595             minq = i;
596         }
597     }
598     while (idx) {
599         sce->sf_idx[bandaddr[idx]] = minq;
600         minq = paths[idx][minq].prev;
601         idx--;
602     }
603     //set the same quantizers inside window groups
604     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
605         for (g = 0;  g < sce->ics.num_swb; g++)
606             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
607                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
608 }
609
610 /**
611  * two-loop quantizers search taken from ISO 13818-7 Appendix C
612  */
613 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
614                                           AACEncContext *s,
615                                           SingleChannelElement *sce,
616                                           const float lambda)
617 {
618     int start = 0, i, w, w2, g;
619     int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels;
620     float dists[128], uplims[128];
621     int fflag, minscaler;
622     int its  = 0;
623     int allz = 0;
624     float minthr = INFINITY;
625
626     //XXX: some heuristic to determine initial quantizers will reduce search time
627     memset(dists, 0, sizeof(dists));
628     //determine zero bands and upper limits
629     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
630         for (g = 0;  g < sce->ics.num_swb; g++) {
631             int nz = 0;
632             float uplim = 0.0f;
633             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
634                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
635                 uplim += band->threshold;
636                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
637                     sce->zeroes[(w+w2)*16+g] = 1;
638                     continue;
639                 }
640                 nz = 1;
641             }
642             uplims[w*16+g] = uplim *512;
643             sce->zeroes[w*16+g] = !nz;
644             if (nz)
645                 minthr = FFMIN(minthr, uplim);
646             allz = FFMAX(allz, nz);
647         }
648     }
649     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
650         for (g = 0;  g < sce->ics.num_swb; g++) {
651             if (sce->zeroes[w*16+g]) {
652                 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
653                 continue;
654             }
655             sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2(uplims[w*16+g]/minthr)*4,59);
656         }
657     }
658
659     if (!allz)
660         return;
661     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
662     //perform two-loop search
663     //outer loop - improve quality
664     do {
665         int tbits, qstep;
666         minscaler = sce->sf_idx[0];
667         //inner loop - quantize spectrum to fit into given number of bits
668         qstep = its ? 1 : 32;
669         do {
670             int prev = -1;
671             tbits = 0;
672             fflag = 0;
673             for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
674                 start = w*128;
675                 for (g = 0;  g < sce->ics.num_swb; g++) {
676                     const float *coefs = sce->coeffs + start;
677                     const float *scaled = s->scoefs + start;
678                     int bits = 0;
679                     int cb;
680                     float mindist = INFINITY;
681                     int minbits = 0;
682
683                     if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
684                         start += sce->ics.swb_sizes[g];
685                         continue;
686                     }
687                     minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
688                     for (cb = 0; cb <= ESC_BT; cb++) {
689                         float dist = 0.0f;
690                         int bb = 0;
691                         for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
692                             int b;
693                             dist += quantize_band_cost(s, coefs + w2*128,
694                                                        scaled + w2*128,
695                                                        sce->ics.swb_sizes[g],
696                                                        sce->sf_idx[w*16+g],
697                                                        cb,
698                                                        lambda,
699                                                        INFINITY,
700                                                        &b);
701                             bb += b;
702                         }
703                         if (dist < mindist) {
704                             mindist = dist;
705                             minbits = bb;
706                         }
707                     }
708                     dists[w*16+g] = (mindist - minbits) / lambda;
709                     bits = minbits;
710                     if (prev != -1) {
711                         bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
712                     }
713                     tbits += bits;
714                     start += sce->ics.swb_sizes[g];
715                     prev = sce->sf_idx[w*16+g];
716                 }
717             }
718             if (tbits > destbits) {
719                 for (i = 0; i < 128; i++)
720                     if (sce->sf_idx[i] < 218 - qstep)
721                         sce->sf_idx[i] += qstep;
722             } else {
723                 for (i = 0; i < 128; i++)
724                     if (sce->sf_idx[i] > 60 - qstep)
725                         sce->sf_idx[i] -= qstep;
726             }
727             qstep >>= 1;
728             if (!qstep && tbits > destbits*1.02)
729                 qstep = 1;
730             if (sce->sf_idx[0] >= 217)
731                 break;
732         } while (qstep);
733
734         fflag = 0;
735         minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
736         for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
737             start = w*128;
738             for (g = 0; g < sce->ics.num_swb; g++) {
739                 int prevsc = sce->sf_idx[w*16+g];
740                 if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60)
741                     sce->sf_idx[w*16+g]--;
742                 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
743                 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
744                 if (sce->sf_idx[w*16+g] != prevsc)
745                     fflag = 1;
746             }
747         }
748         its++;
749     } while (fflag && its < 10);
750 }
751
752 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
753                                        SingleChannelElement *sce,
754                                        const float lambda)
755 {
756     int start = 0, i, w, w2, g;
757     float uplim[128], maxq[128];
758     int minq, maxsf;
759     float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
760     int last = 0, lastband = 0, curband = 0;
761     float avg_energy = 0.0;
762     if (sce->ics.num_windows == 1) {
763         start = 0;
764         for (i = 0; i < 1024; i++) {
765             if (i - start >= sce->ics.swb_sizes[curband]) {
766                 start += sce->ics.swb_sizes[curband];
767                 curband++;
768             }
769             if (sce->coeffs[i]) {
770                 avg_energy += sce->coeffs[i] * sce->coeffs[i];
771                 last = i;
772                 lastband = curband;
773             }
774         }
775     } else {
776         for (w = 0; w < 8; w++) {
777             const float *coeffs = sce->coeffs + w*128;
778             start = 0;
779             for (i = 0; i < 128; i++) {
780                 if (i - start >= sce->ics.swb_sizes[curband]) {
781                     start += sce->ics.swb_sizes[curband];
782                     curband++;
783                 }
784                 if (coeffs[i]) {
785                     avg_energy += coeffs[i] * coeffs[i];
786                     last = FFMAX(last, i);
787                     lastband = FFMAX(lastband, curband);
788                 }
789             }
790         }
791     }
792     last++;
793     avg_energy /= last;
794     if (avg_energy == 0.0f) {
795         for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
796             sce->sf_idx[i] = SCALE_ONE_POS;
797         return;
798     }
799     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
800         start = w*128;
801         for (g = 0; g < sce->ics.num_swb; g++) {
802             float *coefs   = sce->coeffs + start;
803             const int size = sce->ics.swb_sizes[g];
804             int start2 = start, end2 = start + size, peakpos = start;
805             float maxval = -1, thr = 0.0f, t;
806             maxq[w*16+g] = 0.0f;
807             if (g > lastband) {
808                 maxq[w*16+g] = 0.0f;
809                 start += size;
810                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
811                     memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
812                 continue;
813             }
814             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
815                 for (i = 0; i < size; i++) {
816                     float t = coefs[w2*128+i]*coefs[w2*128+i];
817                     maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
818                     thr += t;
819                     if (sce->ics.num_windows == 1 && maxval < t) {
820                         maxval  = t;
821                         peakpos = start+i;
822                     }
823                 }
824             }
825             if (sce->ics.num_windows == 1) {
826                 start2 = FFMAX(peakpos - 2, start2);
827                 end2   = FFMIN(peakpos + 3, end2);
828             } else {
829                 start2 -= start;
830                 end2   -= start;
831             }
832             start += size;
833             thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
834             t   = 1.0 - (1.0 * start2 / last);
835             uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
836         }
837     }
838     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
839     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
840     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
841         start = w*128;
842         for (g = 0;  g < sce->ics.num_swb; g++) {
843             const float *coefs  = sce->coeffs + start;
844             const float *scaled = s->scoefs   + start;
845             const int size      = sce->ics.swb_sizes[g];
846             int scf, prev_scf, step;
847             int min_scf = 0, max_scf = 255;
848             float curdiff;
849             if (maxq[w*16+g] < 21.544) {
850                 sce->zeroes[w*16+g] = 1;
851                 start += size;
852                 continue;
853             }
854             sce->zeroes[w*16+g] = 0;
855             scf  = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2(1/maxq[w*16+g])*16/3, 60, 218);
856             step = 16;
857             for (;;) {
858                 float dist = 0.0f;
859                 int quant_max;
860
861                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
862                     int b;
863                     dist += quantize_band_cost(s, coefs + w2*128,
864                                                scaled + w2*128,
865                                                sce->ics.swb_sizes[g],
866                                                scf,
867                                                ESC_BT,
868                                                lambda,
869                                                INFINITY,
870                                                &b);
871                     dist -= b;
872                 }
873                 dist *= 1.0f / 512.0f / lambda;
874                 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[200 - scf + SCALE_ONE_POS - SCALE_DIV_512]);
875                 if (quant_max >= 8191) { // too much, return to the previous quantizer
876                     sce->sf_idx[w*16+g] = prev_scf;
877                     break;
878                 }
879                 prev_scf = scf;
880                 curdiff = fabsf(dist - uplim[w*16+g]);
881                 if (curdiff == 0.0f)
882                     step = 0;
883                 else
884                     step = fabsf(log2(curdiff));
885                 if (dist > uplim[w*16+g])
886                     step = -step;
887                 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
888                     sce->sf_idx[w*16+g] = scf;
889                     break;
890                 }
891                 scf += step;
892                 if (step > 0)
893                     min_scf = scf;
894                 else
895                     max_scf = scf;
896             }
897             start += size;
898         }
899     }
900     minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
901     for (i = 1; i < 128; i++) {
902         if (!sce->sf_idx[i])
903             sce->sf_idx[i] = sce->sf_idx[i-1];
904         else
905             minq = FFMIN(minq, sce->sf_idx[i]);
906     }
907     if (minq == INT_MAX)
908         minq = 0;
909     minq = FFMIN(minq, SCALE_MAX_POS);
910     maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
911     for (i = 126; i >= 0; i--) {
912         if (!sce->sf_idx[i])
913             sce->sf_idx[i] = sce->sf_idx[i+1];
914         sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
915     }
916 }
917
918 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
919                                        SingleChannelElement *sce,
920                                        const float lambda)
921 {
922     int start = 0, i, w, w2, g;
923     int minq = 255;
924
925     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
926     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
927         start = w*128;
928         for (g = 0; g < sce->ics.num_swb; g++) {
929             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
930                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
931                 if (band->energy <= band->threshold) {
932                     sce->sf_idx[(w+w2)*16+g] = 218;
933                     sce->zeroes[(w+w2)*16+g] = 1;
934                 } else {
935                     sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2(band->threshold), 80, 218);
936                     sce->zeroes[(w+w2)*16+g] = 0;
937                 }
938                 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
939             }
940         }
941     }
942     for (i = 0; i < 128; i++) {
943         sce->sf_idx[i] = 140;
944         //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
945     }
946     //set the same quantizers inside window groups
947     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
948         for (g = 0;  g < sce->ics.num_swb; g++)
949             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
950                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
951 }
952
953 static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
954                           const float lambda)
955 {
956     int start = 0, i, w, w2, g;
957     float M[128], S[128];
958     float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
959     SingleChannelElement *sce0 = &cpe->ch[0];
960     SingleChannelElement *sce1 = &cpe->ch[1];
961     if (!cpe->common_window)
962         return;
963     for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
964         for (g = 0;  g < sce0->ics.num_swb; g++) {
965             if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
966                 float dist1 = 0.0f, dist2 = 0.0f;
967                 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
968                     FFPsyBand *band0 = &s->psy.psy_bands[(s->cur_channel+0)*PSY_MAX_BANDS+(w+w2)*16+g];
969                     FFPsyBand *band1 = &s->psy.psy_bands[(s->cur_channel+1)*PSY_MAX_BANDS+(w+w2)*16+g];
970                     float minthr = FFMIN(band0->threshold, band1->threshold);
971                     float maxthr = FFMAX(band0->threshold, band1->threshold);
972                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
973                         M[i] = (sce0->coeffs[start+w2*128+i]
974                               + sce1->coeffs[start+w2*128+i]) * 0.5;
975                         S[i] =  sce0->coeffs[start+w2*128+i]
976                               - sce1->coeffs[start+w2*128+i];
977                     }
978                     abs_pow34_v(L34, sce0->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
979                     abs_pow34_v(R34, sce1->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
980                     abs_pow34_v(M34, M,                         sce0->ics.swb_sizes[g]);
981                     abs_pow34_v(S34, S,                         sce0->ics.swb_sizes[g]);
982                     dist1 += quantize_band_cost(s, sce0->coeffs + start + w2*128,
983                                                 L34,
984                                                 sce0->ics.swb_sizes[g],
985                                                 sce0->sf_idx[(w+w2)*16+g],
986                                                 sce0->band_type[(w+w2)*16+g],
987                                                 lambda / band0->threshold, INFINITY, NULL);
988                     dist1 += quantize_band_cost(s, sce1->coeffs + start + w2*128,
989                                                 R34,
990                                                 sce1->ics.swb_sizes[g],
991                                                 sce1->sf_idx[(w+w2)*16+g],
992                                                 sce1->band_type[(w+w2)*16+g],
993                                                 lambda / band1->threshold, INFINITY, NULL);
994                     dist2 += quantize_band_cost(s, M,
995                                                 M34,
996                                                 sce0->ics.swb_sizes[g],
997                                                 sce0->sf_idx[(w+w2)*16+g],
998                                                 sce0->band_type[(w+w2)*16+g],
999                                                 lambda / maxthr, INFINITY, NULL);
1000                     dist2 += quantize_band_cost(s, S,
1001                                                 S34,
1002                                                 sce1->ics.swb_sizes[g],
1003                                                 sce1->sf_idx[(w+w2)*16+g],
1004                                                 sce1->band_type[(w+w2)*16+g],
1005                                                 lambda / minthr, INFINITY, NULL);
1006                 }
1007                 cpe->ms_mask[w*16+g] = dist2 < dist1;
1008             }
1009             start += sce0->ics.swb_sizes[g];
1010         }
1011     }
1012 }
1013
1014 AACCoefficientsEncoder ff_aac_coders[] = {
1015     {
1016         search_for_quantizers_faac,
1017         encode_window_bands_info,
1018         quantize_and_encode_band,
1019         search_for_ms,
1020     },
1021     {
1022         search_for_quantizers_anmr,
1023         encode_window_bands_info,
1024         quantize_and_encode_band,
1025         search_for_ms,
1026     },
1027     {
1028         search_for_quantizers_twoloop,
1029         encode_window_bands_info,
1030         quantize_and_encode_band,
1031         search_for_ms,
1032     },
1033     {
1034         search_for_quantizers_fast,
1035         encode_window_bands_info,
1036         quantize_and_encode_band,
1037         search_for_ms,
1038     },
1039 };