Update to MPlayer SVN rev 29473 and FFmpeg SVN rev 19572.
[vaapi:athaifas-mplayer.git] / libavcodec / .svn / text-base / aaccoder.c.svn-base
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]*lambda;
125         if (bits)
126             *bits = 0;
127         return cost;
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]*lambda;
145         minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
146         minbits = ff_aac_spectral_bits[cb-1][minidx];
147         mincost += 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                     //do not code with escape sequence small values
177                     if (vec[k] == 64.0f && t < 39.0f*IQ) {
178                         rd = INFINITY;
179                         break;
180                     }
181                     if (vec[k] == 64.0f) { //FIXME: slow
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*cbrt(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*lambda;
196                 }
197             } else {
198                 for (k = 0; k < dim; k++) {
199                     float di = in[i+k] - vec[k]*IQ;
200                     rd += di*di*lambda;
201                 }
202             }
203             rd += 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]*lambda;
259         minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
260         minbits = ff_aac_spectral_bits[cb-1][minidx];
261         mincost += 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                     //do not code with escape sequence small values
292                     if (vec[k] == 64.0f && t < 39.0f*IQ) {
293                         rd = INFINITY;
294                         break;
295                     }
296                     if (vec[k] == 64.0f) { //FIXME: slow
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*cbrt(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*lambda;
311                 }
312             } else {
313                 for (k = 0; k < dim; k++) {
314                     float di = in[i+k] - vec[k]*IQ;
315                     rd += di*di*lambda;
316                 }
317             }
318             rd += 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;
472     int idx;
473     TrellisPath paths[256*121];
474     int bandaddr[121];
475     int minq;
476     float mincost;
477
478     for (i = 0; i < 256; i++) {
479         paths[i].cost    = 0.0f;
480         paths[i].prev    = -1;
481         paths[i].min_val = i;
482         paths[i].max_val = i;
483     }
484     for (i = 256; i < 256*121; i++) {
485         paths[i].cost    = INFINITY;
486         paths[i].prev    = -2;
487         paths[i].min_val = INT_MAX;
488         paths[i].max_val = 0;
489     }
490     idx = 256;
491     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
492     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
493         start = w*128;
494         for (g = 0; g < sce->ics.num_swb; g++) {
495             const float *coefs = sce->coeffs + start;
496             float qmin, qmax;
497             int nz = 0;
498
499             bandaddr[idx >> 8] = w * 16 + g;
500             qmin = INT_MAX;
501             qmax = 0.0f;
502             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
503                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
504                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
505                     sce->zeroes[(w+w2)*16+g] = 1;
506                     continue;
507                 }
508                 sce->zeroes[(w+w2)*16+g] = 0;
509                 nz = 1;
510                 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
511                     float t = fabsf(coefs[w2*128+i]);
512                     if (t > 0.0f)
513                         qmin = FFMIN(qmin, t);
514                     qmax = FFMAX(qmax, t);
515                 }
516             }
517             if (nz) {
518                 int minscale, maxscale;
519                 float minrd = INFINITY;
520                 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
521                 minscale = av_clip_uint8(log2(qmin)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
522                 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
523                 maxscale = av_clip_uint8(log2(qmax)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
524                 for (q = minscale; q < maxscale; q++) {
525                     float dists[12], dist;
526                     memset(dists, 0, sizeof(dists));
527                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
528                         FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
529                         int cb;
530                         for (cb = 0; cb <= ESC_BT; cb++)
531                             dists[cb] += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
532                                                             q, cb, lambda / band->threshold, INFINITY, NULL);
533                     }
534                     dist = dists[0];
535                     for (i = 1; i <= ESC_BT; i++)
536                         dist = FFMIN(dist, dists[i]);
537                     minrd = FFMIN(minrd, dist);
538
539                     for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, 256); i++) {
540                         float cost;
541                         int minv, maxv;
542                         if (isinf(paths[idx - 256 + i].cost))
543                             continue;
544                         cost = paths[idx - 256 + i].cost + dist
545                                + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
546                         minv = FFMIN(paths[idx - 256 + i].min_val, q);
547                         maxv = FFMAX(paths[idx - 256 + i].max_val, q);
548                         if (cost < paths[idx + q].cost && maxv-minv < SCALE_MAX_DIFF) {
549                             paths[idx + q].cost    = cost;
550                             paths[idx + q].prev    = idx - 256 + i;
551                             paths[idx + q].min_val = minv;
552                             paths[idx + q].max_val = maxv;
553                         }
554                     }
555                 }
556             } else {
557                 for (q = 0; q < 256; q++) {
558                     if (!isinf(paths[idx - 256 + q].cost)) {
559                         paths[idx + q].cost = paths[idx - 256 + q].cost + 1;
560                         paths[idx + q].prev = idx - 256 + q;
561                         paths[idx + q].min_val = FFMIN(paths[idx - 256 + q].min_val, q);
562                         paths[idx + q].max_val = FFMAX(paths[idx - 256 + q].max_val, q);
563                         continue;
564                     }
565                     for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, 256); i++) {
566                         float cost;
567                         int minv, maxv;
568                         if (isinf(paths[idx - 256 + i].cost))
569                             continue;
570                         cost = paths[idx - 256 + i].cost + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
571                         minv = FFMIN(paths[idx - 256 + i].min_val, q);
572                         maxv = FFMAX(paths[idx - 256 + i].max_val, q);
573                         if (cost < paths[idx + q].cost && maxv-minv < SCALE_MAX_DIFF) {
574                             paths[idx + q].cost    = cost;
575                             paths[idx + q].prev    = idx - 256 + i;
576                             paths[idx + q].min_val = minv;
577                             paths[idx + q].max_val = maxv;
578                         }
579                     }
580                 }
581             }
582             sce->zeroes[w*16+g] = !nz;
583             start += sce->ics.swb_sizes[g];
584             idx   += 256;
585         }
586     }
587     idx -= 256;
588     mincost = paths[idx].cost;
589     minq    = idx;
590     for (i = 1; i < 256; i++) {
591         if (paths[idx + i].cost < mincost) {
592             mincost = paths[idx + i].cost;
593             minq = idx + i;
594         }
595     }
596     while (minq >= 256) {
597         sce->sf_idx[bandaddr[minq>>8]] = minq & 0xFF;
598         minq = paths[minq].prev;
599     }
600     //set the same quantizers inside window groups
601     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
602         for (g = 0;  g < sce->ics.num_swb; g++)
603             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
604                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
605 }
606
607 /**
608  * two-loop quantizers search taken from ISO 13818-7 Appendix C
609  */
610 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
611                                           AACEncContext *s,
612                                           SingleChannelElement *sce,
613                                           const float lambda)
614 {
615     int start = 0, i, w, w2, g;
616     int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels;
617     float dists[128], uplims[128];
618     int fflag, minscaler;
619     int its  = 0;
620     int allz = 0;
621     float minthr = INFINITY;
622
623     //XXX: some heuristic to determine initial quantizers will reduce search time
624     memset(dists, 0, sizeof(dists));
625     //determine zero bands and upper limits
626     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
627         for (g = 0;  g < sce->ics.num_swb; g++) {
628             int nz = 0;
629             float uplim = 0.0f;
630             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
631                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
632                 uplim += band->threshold;
633                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
634                     sce->zeroes[(w+w2)*16+g] = 1;
635                     continue;
636                 }
637                 nz = 1;
638             }
639             uplims[w*16+g] = uplim *512;
640             sce->zeroes[w*16+g] = !nz;
641             if (nz)
642                 minthr = FFMIN(minthr, uplim);
643             allz = FFMAX(allz, nz);
644         }
645     }
646     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
647         for (g = 0;  g < sce->ics.num_swb; g++) {
648             if (sce->zeroes[w*16+g]) {
649                 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
650                 continue;
651             }
652             sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2(uplims[w*16+g]/minthr)*4,59);
653         }
654     }
655
656     if (!allz)
657         return;
658     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
659     //perform two-loop search
660     //outer loop - improve quality
661     do {
662         int tbits, qstep;
663         minscaler = sce->sf_idx[0];
664         //inner loop - quantize spectrum to fit into given number of bits
665         qstep = its ? 1 : 32;
666         do {
667             int prev = -1;
668             tbits = 0;
669             fflag = 0;
670             for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
671                 start = w*128;
672                 for (g = 0;  g < sce->ics.num_swb; g++) {
673                     const float *coefs = sce->coeffs + start;
674                     const float *scaled = s->scoefs + start;
675                     int bits = 0;
676                     int cb;
677                     float mindist = INFINITY;
678                     int minbits = 0;
679
680                     if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
681                         start += sce->ics.swb_sizes[g];
682                         continue;
683                     }
684                     minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
685                     for (cb = 0; cb <= ESC_BT; cb++) {
686                         float dist = 0.0f;
687                         int bb = 0;
688                         for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
689                             int b;
690                             dist += quantize_band_cost(s, coefs + w2*128,
691                                                        scaled + w2*128,
692                                                        sce->ics.swb_sizes[g],
693                                                        sce->sf_idx[w*16+g],
694                                                        cb,
695                                                        lambda,
696                                                        INFINITY,
697                                                        &b);
698                             bb += b;
699                         }
700                         if (dist < mindist) {
701                             mindist = dist;
702                             minbits = bb;
703                         }
704                     }
705                     dists[w*16+g] = (mindist - minbits) / lambda;
706                     bits = minbits;
707                     if (prev != -1) {
708                         bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
709                     }
710                     tbits += bits;
711                     start += sce->ics.swb_sizes[g];
712                     prev = sce->sf_idx[w*16+g];
713                 }
714             }
715             if (tbits > destbits) {
716                 for (i = 0; i < 128; i++)
717                     if (sce->sf_idx[i] < 218 - qstep)
718                         sce->sf_idx[i] += qstep;
719             } else {
720                 for (i = 0; i < 128; i++)
721                     if (sce->sf_idx[i] > 60 - qstep)
722                         sce->sf_idx[i] -= qstep;
723             }
724             qstep >>= 1;
725             if (!qstep && tbits > destbits*1.02)
726                 qstep = 1;
727             if (sce->sf_idx[0] >= 217)
728                 break;
729         } while (qstep);
730
731         fflag = 0;
732         minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
733         for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
734             start = w*128;
735             for (g = 0; g < sce->ics.num_swb; g++) {
736                 int prevsc = sce->sf_idx[w*16+g];
737                 if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60)
738                     sce->sf_idx[w*16+g]--;
739                 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
740                 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
741                 if (sce->sf_idx[w*16+g] != prevsc)
742                     fflag = 1;
743             }
744         }
745         its++;
746     } while (fflag && its < 10);
747 }
748
749 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
750                                        SingleChannelElement *sce,
751                                        const float lambda)
752 {
753     int start = 0, i, w, w2, g;
754     float uplim[128], maxq[128];
755     int minq, maxsf;
756     float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
757     int last = 0, lastband = 0, curband = 0;
758     float avg_energy = 0.0;
759     if (sce->ics.num_windows == 1) {
760         start = 0;
761         for (i = 0; i < 1024; i++) {
762             if (i - start >= sce->ics.swb_sizes[curband]) {
763                 start += sce->ics.swb_sizes[curband];
764                 curband++;
765             }
766             if (sce->coeffs[i]) {
767                 avg_energy += sce->coeffs[i] * sce->coeffs[i];
768                 last = i;
769                 lastband = curband;
770             }
771         }
772     } else {
773         for (w = 0; w < 8; w++) {
774             const float *coeffs = sce->coeffs + w*128;
775             start = 0;
776             for (i = 0; i < 128; i++) {
777                 if (i - start >= sce->ics.swb_sizes[curband]) {
778                     start += sce->ics.swb_sizes[curband];
779                     curband++;
780                 }
781                 if (coeffs[i]) {
782                     avg_energy += coeffs[i] * coeffs[i];
783                     last = FFMAX(last, i);
784                     lastband = FFMAX(lastband, curband);
785                 }
786             }
787         }
788     }
789     last++;
790     avg_energy /= last;
791     if (avg_energy == 0.0f) {
792         for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
793             sce->sf_idx[i] = SCALE_ONE_POS;
794         return;
795     }
796     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
797         start = w*128;
798         for (g = 0; g < sce->ics.num_swb; g++) {
799             float *coefs   = sce->coeffs + start;
800             const int size = sce->ics.swb_sizes[g];
801             int start2 = start, end2 = start + size, peakpos = start;
802             float maxval = -1, thr = 0.0f, t;
803             maxq[w*16+g] = 0.0f;
804             if (g > lastband) {
805                 maxq[w*16+g] = 0.0f;
806                 start += size;
807                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
808                     memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
809                 continue;
810             }
811             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
812                 for (i = 0; i < size; i++) {
813                     float t = coefs[w2*128+i]*coefs[w2*128+i];
814                     maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
815                     thr += t;
816                     if (sce->ics.num_windows == 1 && maxval < t) {
817                         maxval  = t;
818                         peakpos = start+i;
819                     }
820                 }
821             }
822             if (sce->ics.num_windows == 1) {
823                 start2 = FFMAX(peakpos - 2, start2);
824                 end2   = FFMIN(peakpos + 3, end2);
825             } else {
826                 start2 -= start;
827                 end2   -= start;
828             }
829             start += size;
830             thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
831             t   = 1.0 - (1.0 * start2 / last);
832             uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
833         }
834     }
835     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
836     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
837     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
838         start = w*128;
839         for (g = 0;  g < sce->ics.num_swb; g++) {
840             const float *coefs  = sce->coeffs + start;
841             const float *scaled = s->scoefs   + start;
842             const int size      = sce->ics.swb_sizes[g];
843             int scf, prev_scf, step;
844             int min_scf = 0, max_scf = 255;
845             float curdiff;
846             if (maxq[w*16+g] < 21.544) {
847                 sce->zeroes[w*16+g] = 1;
848                 start += size;
849                 continue;
850             }
851             sce->zeroes[w*16+g] = 0;
852             scf  = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2(1/maxq[w*16+g])*16/3, 60, 218);
853             step = 16;
854             for (;;) {
855                 float dist = 0.0f;
856                 int quant_max;
857
858                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
859                     int b;
860                     dist += quantize_band_cost(s, coefs + w2*128,
861                                                scaled + w2*128,
862                                                sce->ics.swb_sizes[g],
863                                                scf,
864                                                ESC_BT,
865                                                lambda,
866                                                INFINITY,
867                                                &b);
868                     dist -= b;
869                 }
870                 dist *= 1.0f / 512.0f / lambda;
871                 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[200 - scf + SCALE_ONE_POS - SCALE_DIV_512]);
872                 if (quant_max >= 8191) { // too much, return to the previous quantizer
873                     sce->sf_idx[w*16+g] = prev_scf;
874                     break;
875                 }
876                 prev_scf = scf;
877                 curdiff = fabsf(dist - uplim[w*16+g]);
878                 if (curdiff == 0.0f)
879                     step = 0;
880                 else
881                     step = fabsf(log2(curdiff));
882                 if (dist > uplim[w*16+g])
883                     step = -step;
884                 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
885                     sce->sf_idx[w*16+g] = scf;
886                     break;
887                 }
888                 scf += step;
889                 if (step > 0)
890                     min_scf = scf;
891                 else
892                     max_scf = scf;
893             }
894             start += size;
895         }
896     }
897     minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
898     for (i = 1; i < 128; i++) {
899         if (!sce->sf_idx[i])
900             sce->sf_idx[i] = sce->sf_idx[i-1];
901         else
902             minq = FFMIN(minq, sce->sf_idx[i]);
903     }
904     if (minq == INT_MAX)
905         minq = 0;
906     minq = FFMIN(minq, SCALE_MAX_POS);
907     maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
908     for (i = 126; i >= 0; i--) {
909         if (!sce->sf_idx[i])
910             sce->sf_idx[i] = sce->sf_idx[i+1];
911         sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
912     }
913 }
914
915 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
916                                        SingleChannelElement *sce,
917                                        const float lambda)
918 {
919     int start = 0, i, w, w2, g;
920     int minq = 255;
921
922     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
923     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
924         start = w*128;
925         for (g = 0; g < sce->ics.num_swb; g++) {
926             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
927                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
928                 if (band->energy <= band->threshold) {
929                     sce->sf_idx[(w+w2)*16+g] = 218;
930                     sce->zeroes[(w+w2)*16+g] = 1;
931                 } else {
932                     sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2(band->threshold), 80, 218);
933                     sce->zeroes[(w+w2)*16+g] = 0;
934                 }
935                 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
936             }
937         }
938     }
939     for (i = 0; i < 128; i++) {
940         sce->sf_idx[i] = 140;
941         //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
942     }
943     //set the same quantizers inside window groups
944     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
945         for (g = 0;  g < sce->ics.num_swb; g++)
946             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
947                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
948 }
949
950 static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
951                           const float lambda)
952 {
953     int start = 0, i, w, w2, g;
954     float M[128], S[128];
955     float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
956     SingleChannelElement *sce0 = &cpe->ch[0];
957     SingleChannelElement *sce1 = &cpe->ch[1];
958     if (!cpe->common_window)
959         return;
960     for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
961         for (g = 0;  g < sce0->ics.num_swb; g++) {
962             if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
963                 float dist1 = 0.0f, dist2 = 0.0f;
964                 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
965                     FFPsyBand *band0 = &s->psy.psy_bands[(s->cur_channel+0)*PSY_MAX_BANDS+(w+w2)*16+g];
966                     FFPsyBand *band1 = &s->psy.psy_bands[(s->cur_channel+1)*PSY_MAX_BANDS+(w+w2)*16+g];
967                     float minthr = FFMIN(band0->threshold, band1->threshold);
968                     float maxthr = FFMAX(band0->threshold, band1->threshold);
969                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
970                         M[i] = (sce0->coeffs[start+w2*128+i]
971                               + sce1->coeffs[start+w2*128+i]) * 0.5;
972                         S[i] =  sce0->coeffs[start+w2*128+i]
973                               - sce1->coeffs[start+w2*128+i];
974                     }
975                     abs_pow34_v(L34, sce0->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
976                     abs_pow34_v(R34, sce1->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
977                     abs_pow34_v(M34, M,                         sce0->ics.swb_sizes[g]);
978                     abs_pow34_v(S34, S,                         sce0->ics.swb_sizes[g]);
979                     dist1 += quantize_band_cost(s, sce0->coeffs + start + w2*128,
980                                                 L34,
981                                                 sce0->ics.swb_sizes[g],
982                                                 sce0->sf_idx[(w+w2)*16+g],
983                                                 sce0->band_type[(w+w2)*16+g],
984                                                 lambda / band0->threshold, INFINITY, NULL);
985                     dist1 += quantize_band_cost(s, sce1->coeffs + start + w2*128,
986                                                 R34,
987                                                 sce1->ics.swb_sizes[g],
988                                                 sce1->sf_idx[(w+w2)*16+g],
989                                                 sce1->band_type[(w+w2)*16+g],
990                                                 lambda / band1->threshold, INFINITY, NULL);
991                     dist2 += quantize_band_cost(s, M,
992                                                 M34,
993                                                 sce0->ics.swb_sizes[g],
994                                                 sce0->sf_idx[(w+w2)*16+g],
995                                                 sce0->band_type[(w+w2)*16+g],
996                                                 lambda / maxthr, INFINITY, NULL);
997                     dist2 += quantize_band_cost(s, S,
998                                                 S34,
999                                                 sce1->ics.swb_sizes[g],
1000                                                 sce1->sf_idx[(w+w2)*16+g],
1001                                                 sce1->band_type[(w+w2)*16+g],
1002                                                 lambda / minthr, INFINITY, NULL);
1003                 }
1004                 cpe->ms_mask[w*16+g] = dist2 < dist1;
1005             }
1006             start += sce0->ics.swb_sizes[g];
1007         }
1008     }
1009 }
1010
1011 AACCoefficientsEncoder ff_aac_coders[] = {
1012     {
1013         search_for_quantizers_faac,
1014         encode_window_bands_info,
1015         quantize_and_encode_band,
1016 //        search_for_ms,
1017     },
1018     {
1019         search_for_quantizers_anmr,
1020         encode_window_bands_info,
1021         quantize_and_encode_band,
1022 //        search_for_ms,
1023     },
1024     {
1025         search_for_quantizers_twoloop,
1026         encode_window_bands_info,
1027         quantize_and_encode_band,
1028 //        search_for_ms,
1029     },
1030     {
1031         search_for_quantizers_fast,
1032         encode_window_bands_info,
1033         quantize_and_encode_band,
1034 //        search_for_ms,
1035     },
1036 };