41d1802fd746ddc51dea2a35d635d3762d4c62c8
[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     return pow(coef * Q, 0.75) + 0.4054;
65 }
66
67 static void quantize_bands(int (*out)[2], const float *in, const float *scaled, int size, float Q34, int is_signed, int maxval)
68 {
69     int i;
70     double qc;
71     for (i = 0; i < size; i++) {
72         qc = scaled[i] * Q34;
73         out[i][0] = (int)FFMIN((int)qc, maxval);
74         out[i][1] = (int)FFMIN((int)(qc + 0.4054), maxval);
75         if (is_signed && in[i] < 0.0f) {
76             out[i][0] = -out[i][0];
77             out[i][1] = -out[i][1];
78         }
79     }
80 }
81
82 static void abs_pow34_v(float *out, const float* in, const int size)
83 {
84 #ifndef USE_REALLY_FULL_SEARCH
85     int i;
86     for (i = 0; i < size; i++) {
87         out[i] = pow(fabsf(in[i]), 0.75);
88     }
89 #endif /* USE_REALLY_FULL_SEARCH */
90 }
91
92 static av_always_inline int quant2(float coef, const float Q)
93 {
94     return pow(coef * Q, 0.75);
95 }
96
97 static const uint8_t aac_cb_range [12] = {0, 3, 3, 3, 3, 9, 9, 8, 8, 13, 13, 17};
98 static const uint8_t aac_cb_maxval[12] = {0, 1, 1, 2, 2, 4, 4, 7, 7, 12, 12, 16};
99
100 /**
101  * Calculate rate distortion cost for quantizing with given codebook
102  *
103  * @return quantization distortion
104  */
105 static float quantize_band_cost(struct AACEncContext *s, const float *in, const float *scaled, int size, int scale_idx, int cb,
106                                  const float lambda, const float uplim, 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 = pow(Q, 0.75);
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]*lambda;
125         return cost;
126     }
127 #ifndef USE_REALLY_FULL_SEARCH
128     offs[0] = 1;
129     for (i = 1; i < dim; i++)
130         offs[i] = offs[i-1]*range;
131     quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
132 #endif /* USE_REALLY_FULL_SEARCH */
133     for (i = 0; i < size; i += dim) {
134         float mincost;
135         int minidx = 0;
136         int minbits = 0;
137         const float *vec;
138 #ifndef USE_REALLY_FULL_SEARCH
139         int (*quants)[2] = &s->qcoefs[i];
140         mincost = 0.0f;
141         for (j = 0; j < dim; j++) {
142             mincost += in[i+j]*in[i+j]*lambda;
143         }
144         minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
145         minbits = ff_aac_spectral_bits[cb-1][minidx];
146         mincost += minbits;
147         for (j = 0; j < (1<<dim); j++) {
148             float rd = 0.0f;
149             int curbits;
150             int curidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
151             int same = 0;
152             for (k = 0; k < dim; k++) {
153                 if ((j & (1 << k)) && quants[k][0] == quants[k][1]) {
154                     same = 1;
155                     break;
156                 }
157             }
158             if (same)
159                 continue;
160             for (k = 0; k < dim; k++)
161                 curidx += quants[k][!!(j & (1 << k))] * offs[dim - 1 - k];
162             curbits = ff_aac_spectral_bits[cb-1][curidx];
163             vec = &ff_aac_codebook_vectors[cb-1][curidx*dim];
164 #else
165         mincost = INFINITY;
166         vec = ff_aac_codebook_vectors[cb-1];
167         for (j = 0; j < ff_aac_spectral_sizes[cb-1]; j++, vec += dim) {
168             float rd = 0.0f;
169             int curbits = ff_aac_spectral_bits[cb-1][j];
170 #endif /* USE_REALLY_FULL_SEARCH */
171             if (IS_CODEBOOK_UNSIGNED(cb)) {
172                 for (k = 0; k < dim; k++) {
173                     float t = fabsf(in[i+k]);
174                     float di;
175                     //do not code with escape sequence small values
176                     if (vec[k] == 64.0f && t < 39.0f*IQ) {
177                         rd = INFINITY;
178                         break;
179                     }
180                     if (vec[k] == 64.0f) {//FIXME: slow
181                         if (t >= CLIPPED_ESCAPE) {
182                             di = t - CLIPPED_ESCAPE;
183                             curbits += 21;
184                         } else {
185                             int c = av_clip(quant(t, Q), 0, 8191);
186                             di = t - c*cbrt(c)*IQ;
187                             curbits += av_log2(c)*2 - 4 + 1;
188                         }
189                     } else {
190                         di = t - vec[k]*IQ;
191                     }
192                     if (vec[k] != 0.0f)
193                         curbits++;
194                     rd += di*di*lambda;
195                 }
196             } else {
197                 for (k = 0; k < dim; k++) {
198                     float di = in[i+k] - vec[k]*IQ;
199                     rd += di*di*lambda;
200                 }
201             }
202             rd += curbits;
203             if (rd < mincost) {
204                 mincost = rd;
205                 minidx = j;
206                 minbits = curbits;
207             }
208         }
209         cost += mincost;
210         resbits += minbits;
211         if (cost >= uplim)
212             return uplim;
213     }
214
215     if (bits)
216         *bits = resbits;
217     return cost;
218 }
219
220 static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb, const float *in, int size,
221                                      int scale_idx, int cb, const float lambda)
222 {
223     const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
224     const float  Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
225     const float CLIPPED_ESCAPE = 165140.0f*IQ;
226     const int dim = (cb < FIRST_PAIR_BT) ? 4 : 2;
227     int i, j, k;
228 #ifndef USE_REALLY_FULL_SEARCH
229     const float  Q34 = pow(Q, 0.75);
230     const int range = aac_cb_range[cb];
231     const int maxval = aac_cb_maxval[cb];
232     int offs[4];
233     float *scaled = s->scoefs;
234 #endif /* USE_REALLY_FULL_SEARCH */
235
236 //START_TIMER
237     if (!cb)
238         return;
239
240 #ifndef USE_REALLY_FULL_SEARCH
241     offs[0] = 1;
242     for (i = 1; i < dim; i++)
243         offs[i] = offs[i-1]*range;
244     abs_pow34_v(scaled, in, size);
245     quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
246 #endif /* USE_REALLY_FULL_SEARCH */
247     for (i = 0; i < size; i += dim) {
248         float mincost;
249         int minidx = 0;
250         int minbits = 0;
251         const float *vec;
252 #ifndef USE_REALLY_FULL_SEARCH
253         int (*quants)[2] = &s->qcoefs[i];
254         mincost = 0.0f;
255         for (j = 0; j < dim; j++) {
256             mincost += in[i+j]*in[i+j]*lambda;
257         }
258         minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
259         minbits = ff_aac_spectral_bits[cb-1][minidx];
260         mincost += minbits;
261         for (j = 0; j < (1<<dim); j++) {
262             float rd = 0.0f;
263             int curbits;
264             int curidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
265             int same = 0;
266             for (k = 0; k < dim; k++) {
267                 if ((j & (1 << k)) && quants[k][0] == quants[k][1]) {
268                     same = 1;
269                     break;
270                 }
271             }
272             if (same)
273                 continue;
274             for (k = 0; k < dim; k++)
275                 curidx += quants[k][!!(j & (1 << k))] * offs[dim - 1 - k];
276             curbits = ff_aac_spectral_bits[cb-1][curidx];
277             vec = &ff_aac_codebook_vectors[cb-1][curidx*dim];
278 #else
279         vec = ff_aac_codebook_vectors[cb-1];
280         mincost = INFINITY;
281         for (j = 0; j < ff_aac_spectral_sizes[cb-1]; j++, vec += dim) {
282             float rd = 0.0f;
283             int curbits = ff_aac_spectral_bits[cb-1][j];
284             int curidx = j;
285 #endif /* USE_REALLY_FULL_SEARCH */
286             if (IS_CODEBOOK_UNSIGNED(cb)) {
287                 for (k = 0; k < dim; k++) {
288                     float t = fabsf(in[i+k]);
289                     float di;
290                     //do not code with escape sequence small values
291                     if (vec[k] == 64.0f && t < 39.0f*IQ) {
292                         rd = INFINITY;
293                         break;
294                     }
295                     if (vec[k] == 64.0f) {//FIXME: slow
296                         if (t >= CLIPPED_ESCAPE) {
297                             di = t - CLIPPED_ESCAPE;
298                             curbits += 21;
299                         } else {
300                             int c = av_clip(quant(t, Q), 0, 8191);
301                             di = t - c*cbrt(c)*IQ;
302                             curbits += av_log2(c)*2 - 4 + 1;
303                         }
304                     } else {
305                         di = t - vec[k]*IQ;
306                     }
307                     if (vec[k] != 0.0f)
308                         curbits++;
309                     rd += di*di*lambda;
310                 }
311             } else {
312                 for (k = 0; k < dim; k++) {
313                     float di = in[i+k] - vec[k]*IQ;
314                     rd += di*di*lambda;
315                 }
316             }
317             rd += curbits;
318             if (rd < mincost) {
319                 mincost = rd;
320                 minidx = curidx;
321                 minbits = curbits;
322             }
323         }
324         put_bits(pb, ff_aac_spectral_bits[cb-1][minidx], ff_aac_spectral_codes[cb-1][minidx]);
325         if (IS_CODEBOOK_UNSIGNED(cb))
326             for (j = 0; j < dim; j++)
327                 if (ff_aac_codebook_vectors[cb-1][minidx*dim+j] != 0.0f)
328                     put_bits(pb, 1, in[i+j] < 0.0f);
329         if (cb == ESC_BT) {
330             for (j = 0; j < 2; j++) {
331                 if (ff_aac_codebook_vectors[cb-1][minidx*2+j] == 64.0f) {
332                     int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
333                     int len = av_log2(coef);
334
335                     put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
336                     put_bits(pb, len, coef & ((1 << len) - 1));
337                 }
338             }
339         }
340     }
341 //STOP_TIMER("quantize_and_encode")
342 }
343
344 /**
345  * structure used in optimal codebook search
346  */
347 typedef struct BandCodingPath {
348     int prev_idx; ///< pointer to the previous path point
349     int codebook; ///< codebook for coding band run
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     }
432     ppos = max_sfb;
433     while(ppos > 0) {
434         cb = idx;
435         stackrun[stack_len] = path[ppos][cb].run;
436         stackcb [stack_len] = cb;
437         idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
438         ppos -= path[ppos][cb].run;
439         stack_len++;
440     }
441     //perform actual band info encoding
442     start = 0;
443     for (i = stack_len - 1; i >= 0; i--) {
444         put_bits(&s->pb, 4, stackcb[i]);
445         count = stackrun[i];
446         memset(sce->zeroes + win*16 + start, !stackcb[i], count);
447         //XXX: memset when band_type is also uint8_t
448         for (j = 0; j < count; j++) {
449             sce->band_type[win*16 + start] =  stackcb[i];
450             start++;
451         }
452         while(count >= run_esc) {
453             put_bits(&s->pb, run_bits, run_esc);
454             count -= run_esc;
455         }
456         put_bits(&s->pb, run_bits, count);
457     }
458 }
459
460 static void encode_window_bands_info_fixed(AACEncContext *s, SingleChannelElement *sce,
461                                            int win, int group_len, const float lambda)
462 {
463     encode_window_bands_info(s, sce, win, group_len, 1.0f);
464 }
465
466
467 typedef struct TrellisPath {
468     float cost;
469     int prev;
470     int min_val;
471     int max_val;
472 } TrellisPath;
473
474 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
475                                        SingleChannelElement *sce, const float lambda)
476 {
477     int q, w, w2, g, start = 0;
478     int i;
479     int idx;
480     TrellisPath paths[256*121];
481     int bandaddr[121];
482     int minq;
483     float mincost;
484
485     for (i = 0; i < 256; i++) {
486         paths[i].cost = 0.0f;
487         paths[i].prev = -1;
488         paths[i].min_val = i;
489         paths[i].max_val = i;
490     }
491     for (i = 256; i < 256*121; i++) {
492         paths[i].cost = INFINITY;
493         paths[i].prev = -2;
494         paths[i].min_val = INT_MAX;
495         paths[i].max_val = 0;
496     }
497     idx = 256;
498     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
499     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
500         start = w*128;
501         for (g = 0; g < sce->ics.num_swb; g++) {
502             const float *coefs = sce->coeffs + start;
503             float qmin, qmax;
504             int nz = 0;
505
506             bandaddr[idx >> 8] = w*16+g;
507             qmin = INT_MAX;
508             qmax = 0.0f;
509             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
510                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
511                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
512                     sce->zeroes[(w+w2)*16+g] = 1;
513                     continue;
514                 }
515                 sce->zeroes[(w+w2)*16+g] = 0;
516                 nz = 1;
517                 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
518                     float t = fabsf(coefs[w2*128+i]);
519                     if (t > 0.0f) qmin = fminf(qmin, t);
520                     qmax = fmaxf(qmax, t);
521                 }
522             }
523             if (nz) {
524                 int minscale, maxscale;
525                 float minrd = INFINITY;
526                 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
527                 minscale = av_clip_uint8(log2(qmin)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
528                 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
529                 maxscale = av_clip_uint8(log2(qmax)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
530                 for (q = minscale; q < maxscale; q++) {
531                     float dists[12], dist;
532                     memset(dists, 0, sizeof(dists));
533                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
534                         FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
535                         int cb;
536                         for (cb = 0; cb <= ESC_BT; cb++) {
537                             dists[cb] += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
538                                                             q, cb, lambda / band->threshold, INFINITY, NULL);
539                         }
540                     }
541                     dist = dists[0];
542                     for (i = 1; i <= ESC_BT; i++)
543                         dist = fminf(dist, dists[i]);
544                     minrd = fminf(minrd, dist);
545
546                     for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, 256); i++) {
547                         float cost;
548                         int minv, maxv;
549                         if (isinf(paths[idx - 256 + i].cost))
550                             continue;
551                         cost = paths[idx - 256 + i].cost + dist
552                                + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
553                         minv = FFMIN(paths[idx - 256 + i].min_val, q);
554                         maxv = FFMAX(paths[idx - 256 + i].max_val, q);
555                         if (cost < paths[idx + q].cost && maxv-minv < SCALE_MAX_DIFF) {
556                             paths[idx + q].cost = cost;
557                             paths[idx + q].prev = idx - 256 + i;
558                             paths[idx + q].min_val = minv;
559                             paths[idx + q].max_val = maxv;
560                         }
561                     }
562                 }
563             } else {
564                 for (q = 0; q < 256; q++) {
565                     if (!isinf(paths[idx - 256 + q].cost)) {
566                         paths[idx + q].cost = paths[idx - 256 + q].cost + 1;
567                         paths[idx + q].prev = idx - 256 + q;
568                         paths[idx + q].min_val = FFMIN(paths[idx - 256 + q].min_val, q);
569                         paths[idx + q].max_val = FFMAX(paths[idx - 256 + q].max_val, q);
570                         continue;
571                     }
572                     for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, 256); i++) {
573                         float cost;
574                         int minv, maxv;
575                         if (isinf(paths[idx - 256 + i].cost))
576                             continue;
577                         cost = paths[idx - 256 + i].cost + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
578                         minv = FFMIN(paths[idx - 256 + i].min_val, q);
579                         maxv = FFMAX(paths[idx - 256 + i].max_val, q);
580                         if (cost < paths[idx + q].cost && maxv-minv < SCALE_MAX_DIFF) {
581                             paths[idx + q].cost = cost;
582                             paths[idx + q].prev = idx - 256 + i;
583                             paths[idx + q].min_val = minv;
584                             paths[idx + q].max_val = maxv;
585                         }
586                     }
587                 }
588             }
589             sce->zeroes[w*16+g] = !nz;
590             start += sce->ics.swb_sizes[g];
591             idx += 256;
592         }
593     }
594     idx -= 256;
595     mincost = paths[idx].cost;
596     minq = idx;
597     for (i = 1; i < 256; i++) {
598         if (paths[idx + i].cost < mincost) {
599             mincost = paths[idx + i].cost;
600             minq = idx + i;
601         }
602     }
603     while(minq >= 256) {
604         sce->sf_idx[bandaddr[minq>>8]] = minq & 0xFF;
605         minq = paths[minq].prev;
606     }
607     //set the same quantizers inside window groups
608     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
609         for (g = 0;  g < sce->ics.num_swb; g++)
610             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
611                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
612 }
613
614 /**
615  * two-loop quantizers search taken from ISO 13818-7 Appendix C
616  */
617 static void search_for_quantizers_twoloop(AVCodecContext *avctx, AACEncContext *s,
618                                           SingleChannelElement *sce, const float lambda)
619 {
620     int start = 0, i, w, w2, g;
621     int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels;
622     float dists[128], uplims[128];
623     int fflag, minscaler;
624     int its = 0;
625     int allz = 0;
626     float minthr = INFINITY;
627
628     //XXX: some heuristic to determine initial quantizers will reduce search time
629     memset(dists, 0, sizeof(dists));
630     //determine zero bands and upper limits
631     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
632         for (g = 0;  g < sce->ics.num_swb; g++) {
633             int nz = 0;
634             float uplim = 0.0f;
635             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
636                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
637                 uplim += band->threshold;
638                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
639                     sce->zeroes[(w+w2)*16+g] = 1;
640                     continue;
641                 }
642                 nz = 1;
643             }
644             uplims[w*16+g] = uplim *512;
645             sce->zeroes[w*16+g] = !nz;
646             if (nz)
647                 minthr = fminf(minthr, uplim);
648             allz = FFMAX(allz, nz);
649         }
650     }
651     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
652         for (g = 0;  g < sce->ics.num_swb; g++) {
653             if (sce->zeroes[w*16+g]) {
654                 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
655                 continue;
656             }
657             sce->sf_idx[w*16+g] = SCALE_ONE_POS + fminf(log2(uplims[w*16+g]/minthr)*4,59);
658         }
659     }
660
661     if (!allz)
662         return;
663     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
664     //perform two-loop search
665     //outer loop - improve quality
666     do{
667         int tbits, qstep;
668         minscaler = sce->sf_idx[0];
669         //inner loop - quantize spectrum to fit into given number of bits
670         qstep = its ? 1 : 32;
671         do{
672             int prev = -1;
673             tbits = 0;
674             fflag = 0;
675             for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
676                 start = w*128;
677                 for (g = 0;  g < sce->ics.num_swb; g++) {
678                     const float *coefs = sce->coeffs + start;
679                     const float *scaled = s->scoefs + start;
680                     int bits = 0;
681                     int cb;
682                     float mindist = INFINITY;
683                     int minbits = 0;
684
685                     if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218)
686                         continue;
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                                                        ESC_BT,
698                                                        1.0,
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;
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                     }
723                 }
724             } else {
725                 for (i = 0; i < 128; i++) {
726                     if (sce->sf_idx[i] > 60 - qstep) {
727                         sce->sf_idx[i] -= qstep;
728                     }
729                 }
730             }
731             qstep >>= 1;
732             if (!qstep && tbits > destbits*1.02)
733                 qstep = 1;
734             if (sce->sf_idx[0] >= 217)break;
735         }while(qstep);
736
737         fflag = 0;
738         minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
739         for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
740             start = w*128;
741             for (g = 0; g < sce->ics.num_swb; g++) {
742                 int prevsc = sce->sf_idx[w*16+g];
743                 if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60)
744                     sce->sf_idx[w*16+g]--;
745                 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
746                 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
747                 if (sce->sf_idx[w*16+g] != prevsc)
748                     fflag = 1;
749             }
750         }
751         its++;
752     }while(fflag && its < 10);
753 }
754
755 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
756                                        SingleChannelElement *sce, const float lambda)
757 {
758     int start = 0, i, w, w2, g;
759     float uplim[128], maxq[128];
760     int minq, maxsf;
761     float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
762     int last = 0, lastband = 0, curband = 0;
763     float avg_energy = 0.0;
764     if (sce->ics.num_windows == 1) {
765         start = 0;
766         for (i = 0; i < 1024; i++) {
767             if (i - start >= sce->ics.swb_sizes[curband]) {
768                 start += sce->ics.swb_sizes[curband];
769                 curband++;
770             }
771             if (sce->coeffs[i]) {
772                 avg_energy += sce->coeffs[i] * sce->coeffs[i];
773                 last = i;
774                 lastband = curband;
775             }
776         }
777     } else {
778         for (w = 0; w < 8; w++) {
779             const float *coeffs = sce->coeffs + w*128;
780             start = 0;
781             for (i = 0; i < 128; i++) {
782                 if (i - start >= sce->ics.swb_sizes[curband]) {
783                     start += sce->ics.swb_sizes[curband];
784                     curband++;
785                 }
786                 if (coeffs[i]) {
787                     avg_energy += coeffs[i] * coeffs[i];
788                     last = FFMAX(last, i);
789                     lastband = FFMAX(lastband, curband);
790                 }
791             }
792         }
793     }
794     last++;
795     avg_energy /= last;
796     if (avg_energy == 0.0f) {
797         for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
798             sce->sf_idx[i] = SCALE_ONE_POS;
799         return;
800     }
801     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
802         start = w*128;
803         for (g = 0; g < sce->ics.num_swb; g++) {
804             float *coefs = sce->coeffs + start;
805             const int size = sce->ics.swb_sizes[g];
806             int start2 = start, end2 = start + size, peakpos = start;
807             float maxval = -1, thr = 0.0f, t;
808             maxq[w*16+g] = 0.0f;
809             if (g > lastband) {
810                 maxq[w*16+g] = 0.0f;
811                 start += size;
812                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
813                     memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
814                 continue;
815             }
816             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
817                 for (i = 0; i < size; i++) {
818                     float t = coefs[w2*128+i]*coefs[w2*128+i];
819                     maxq[w*16+g] = fmaxf(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
820                     thr += t;
821                     if (sce->ics.num_windows == 1 && maxval < t) {
822                         maxval = t;
823                         peakpos = start+i;
824                     }
825                 }
826             }
827             if (sce->ics.num_windows == 1) {
828                 start2 = FFMAX(peakpos - 2, start2);
829                 end2   = FFMIN(peakpos + 3, end2);
830             } else {
831                 start2 -= start;
832                 end2   -= start;
833             }
834             start += size;
835             thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
836             t = 1.0 - (1.0 * start2 / last);
837             uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
838         }
839     }
840     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
841     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
842     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
843         start = w*128;
844         for (g = 0;  g < sce->ics.num_swb; g++) {
845             const float *coefs = sce->coeffs + start;
846             const float *scaled = s->scoefs + start;
847             const int size = sce->ics.swb_sizes[g];
848             int scf, prev_scf, step;
849             int min_scf = 0, max_scf = 255;
850             float curdiff;
851             if (maxq[w*16+g] < 21.544) {
852                 sce->zeroes[w*16+g] = 1;
853                 start += size;
854                 continue;
855             }
856             sce->zeroes[w*16+g] = 0;
857             scf = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2(1/maxq[w*16+g])*16/3, 60, 218);
858             step = 16;
859             for (;;) {
860                 float dist = 0.0f;
861                 int quant_max;
862
863                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
864                     int b;
865                     dist += quantize_band_cost(s, coefs + w2*128,
866                                                scaled + w2*128,
867                                                sce->ics.swb_sizes[g],
868                                                scf,
869                                                ESC_BT,
870                                                1.0,
871                                                INFINITY,
872                                                &b);
873                     dist -= b;
874                 }
875                 dist *= 1.0f/512.0f;
876                 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[200 - scf + SCALE_ONE_POS - SCALE_DIV_512]);
877                 if (quant_max >= 8191) { // too much, return to the previous quantizer
878                     sce->sf_idx[w*16+g] = prev_scf;
879                     break;
880                 }
881                 prev_scf = scf;
882                 curdiff = fabsf(dist - uplim[w*16+g]);
883                 if (curdiff == 0.0f)
884                     step = 0;
885                 else
886                     step = fabsf(log2(curdiff));
887                 if (dist > uplim[w*16+g])
888                     step = -step;
889                 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
890                     sce->sf_idx[w*16+g] = scf;
891                     break;
892                 }
893                 scf += step;
894                 if (step > 0)
895                     min_scf = scf;
896                 else
897                     max_scf = scf;
898             }
899             start += size;
900         }
901     }
902     minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
903     for (i = 1; i < 128; i++) {
904         if (!sce->sf_idx[i])
905             sce->sf_idx[i] = sce->sf_idx[i-1];
906         else
907             minq = FFMIN(minq, sce->sf_idx[i]);
908     }
909     if (minq == INT_MAX) minq = 0;
910     minq = FFMIN(minq, SCALE_MAX_POS);
911     maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
912     for (i = 126; i >= 0; i--) {
913         if (!sce->sf_idx[i])
914             sce->sf_idx[i] = sce->sf_idx[i+1];
915         sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
916     }
917 }
918
919 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
920                                        SingleChannelElement *sce, 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;//av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
944     }
945     //set the same quantizers inside window groups
946     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
947         for (g = 0;  g < sce->ics.num_swb; g++)
948             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
949                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
950 }
951
952 static void search_for_ms(AACEncContext *s, ChannelElement *cpe, const float lambda)
953 {
954     int start = 0, i, w, w2, g;
955     float M[128], S[128];
956     float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
957     SingleChannelElement *sce0 = &cpe->ch[0];
958     SingleChannelElement *sce1 = &cpe->ch[1];
959     if (!cpe->common_window)
960         return;
961     for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
962         for (g = 0;  g < sce0->ics.num_swb; g++) {
963             if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
964                 float dist1 = 0.0f, dist2 = 0.0f;
965                 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
966                     FFPsyBand *band0 = &s->psy.psy_bands[(s->cur_channel+0)*PSY_MAX_BANDS+(w+w2)*16+g];
967                     FFPsyBand *band1 = &s->psy.psy_bands[(s->cur_channel+1)*PSY_MAX_BANDS+(w+w2)*16+g];
968                     float minthr = fminf(band0->threshold, band1->threshold);
969                     float maxthr = fmaxf(band0->threshold, band1->threshold);
970                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
971                         M[i] = (sce0->coeffs[start+w2*128+i]
972                               + sce1->coeffs[start+w2*128+i])*0.5;
973                         S[i] =  sce0->coeffs[start+w2*128+i]
974                               - sce1->coeffs[start+w2*128+i];
975                     }
976                     abs_pow34_v(L34, sce0->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
977                     abs_pow34_v(R34, sce1->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
978                     abs_pow34_v(M34, M,                         sce0->ics.swb_sizes[g]);
979                     abs_pow34_v(S34, S,                         sce0->ics.swb_sizes[g]);
980                     dist1 += quantize_band_cost(s, sce0->coeffs + start + w2*128,
981                                                 L34,
982                                                 sce0->ics.swb_sizes[g],
983                                                 sce0->sf_idx[(w+w2)*16+g],
984                                                 sce0->band_type[(w+w2)*16+g],
985                                                 lambda / band0->threshold, INFINITY, NULL);
986                     dist1 += quantize_band_cost(s, sce1->coeffs + start + w2*128,
987                                                 R34,
988                                                 sce1->ics.swb_sizes[g],
989                                                 sce1->sf_idx[(w+w2)*16+g],
990                                                 sce1->band_type[(w+w2)*16+g],
991                                                 lambda / band1->threshold, INFINITY, NULL);
992                     dist2 += quantize_band_cost(s, M,
993                                                 M34,
994                                                 sce0->ics.swb_sizes[g],
995                                                 sce0->sf_idx[(w+w2)*16+g],
996                                                 sce0->band_type[(w+w2)*16+g],
997                                                 lambda / maxthr, INFINITY, NULL);
998                     dist2 += quantize_band_cost(s, S,
999                                                 S34,
1000                                                 sce1->ics.swb_sizes[g],
1001                                                 sce1->sf_idx[(w+w2)*16+g],
1002                                                 sce1->band_type[(w+w2)*16+g],
1003                                                 lambda / minthr, INFINITY, NULL);
1004                 }
1005                 cpe->ms_mask[w*16+g] = dist2 < dist1;
1006             }
1007             start += sce0->ics.swb_sizes[g];
1008         }
1009     }
1010 }
1011
1012 AACCoefficientsEncoder ff_aac_coders[] = {
1013     {
1014         search_for_quantizers_faac,
1015         encode_window_bands_info_fixed,
1016         quantize_and_encode_band,
1017 //        search_for_ms,
1018     },
1019     {
1020         search_for_quantizers_anmr,
1021         encode_window_bands_info,
1022         quantize_and_encode_band,
1023 //        search_for_ms,
1024     },
1025     {
1026         search_for_quantizers_twoloop,
1027         encode_window_bands_info,
1028         quantize_and_encode_band,
1029 //        search_for_ms,
1030     },
1031     {
1032         search_for_quantizers_fast,
1033         encode_window_bands_info,
1034         quantize_and_encode_band,
1035 //        search_for_ms,
1036     },
1037 };