Actually use all the codebooks we are iterating over in the two-loop scalefactor...
[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]*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     int codebook; ///< codebook for coding band run
351     float cost;   ///< path cost
352     int run;
353 } BandCodingPath;
354
355 /**
356  * Encode band info for single window group bands.
357  */
358 static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
359                                      int win, int group_len, const float lambda)
360 {
361     BandCodingPath path[120][12];
362     int w, swb, cb, start, start2, size;
363     int i, j;
364     const int max_sfb  = sce->ics.max_sfb;
365     const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
366     const int run_esc  = (1 << run_bits) - 1;
367     int idx, ppos, count;
368     int stackrun[120], stackcb[120], stack_len;
369     float next_minrd = INFINITY;
370     int next_mincb = 0;
371
372     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
373     start = win*128;
374     for (cb = 0; cb < 12; cb++) {
375         path[0][cb].cost     = 0.0f;
376         path[0][cb].prev_idx = -1;
377         path[0][cb].run      = 0;
378     }
379     for (swb = 0; swb < max_sfb; swb++) {
380         start2 = start;
381         size = sce->ics.swb_sizes[swb];
382         if (sce->zeroes[win*16 + swb]) {
383             for (cb = 0; cb < 12; cb++) {
384                 path[swb+1][cb].prev_idx = cb;
385                 path[swb+1][cb].cost     = path[swb][cb].cost;
386                 path[swb+1][cb].run      = path[swb][cb].run + 1;
387             }
388         } else {
389             float minrd = next_minrd;
390             int mincb = next_mincb;
391             next_minrd = INFINITY;
392             next_mincb = 0;
393             for (cb = 0; cb < 12; cb++) {
394                 float cost_stay_here, cost_get_here;
395                 float rd = 0.0f;
396                 for (w = 0; w < group_len; w++) {
397                     FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(win+w)*16+swb];
398                     rd += quantize_band_cost(s, sce->coeffs + start + w*128,
399                                              s->scoefs + start + w*128, size,
400                                              sce->sf_idx[(win+w)*16+swb], cb,
401                                              lambda / band->threshold, INFINITY, NULL);
402                 }
403                 cost_stay_here = path[swb][cb].cost + rd;
404                 cost_get_here  = minrd              + rd + run_bits + 4;
405                 if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
406                     != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
407                     cost_stay_here += run_bits;
408                 if (cost_get_here < cost_stay_here) {
409                     path[swb+1][cb].prev_idx = mincb;
410                     path[swb+1][cb].cost     = cost_get_here;
411                     path[swb+1][cb].run      = 1;
412                 } else {
413                     path[swb+1][cb].prev_idx = cb;
414                     path[swb+1][cb].cost     = cost_stay_here;
415                     path[swb+1][cb].run      = path[swb][cb].run + 1;
416                 }
417                 if (path[swb+1][cb].cost < next_minrd) {
418                     next_minrd = path[swb+1][cb].cost;
419                     next_mincb = cb;
420                 }
421             }
422         }
423         start += sce->ics.swb_sizes[swb];
424     }
425
426     //convert resulting path from backward-linked list
427     stack_len = 0;
428     idx       = 0;
429     for (cb = 1; cb < 12; cb++)
430         if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
431             idx = cb;
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 typedef struct TrellisPath {
461     float cost;
462     int prev;
463     int min_val;
464     int max_val;
465 } TrellisPath;
466
467 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
468                                        SingleChannelElement *sce,
469                                        const float lambda)
470 {
471     int q, w, w2, g, start = 0;
472     int i;
473     int idx;
474     TrellisPath paths[256*121];
475     int bandaddr[121];
476     int minq;
477     float mincost;
478
479     for (i = 0; i < 256; i++) {
480         paths[i].cost    = 0.0f;
481         paths[i].prev    = -1;
482         paths[i].min_val = i;
483         paths[i].max_val = i;
484     }
485     for (i = 256; i < 256*121; i++) {
486         paths[i].cost    = INFINITY;
487         paths[i].prev    = -2;
488         paths[i].min_val = INT_MAX;
489         paths[i].max_val = 0;
490     }
491     idx = 256;
492     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
493     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
494         start = w*128;
495         for (g = 0; g < sce->ics.num_swb; g++) {
496             const float *coefs = sce->coeffs + start;
497             float qmin, qmax;
498             int nz = 0;
499
500             bandaddr[idx >> 8] = w * 16 + g;
501             qmin = INT_MAX;
502             qmax = 0.0f;
503             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
504                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
505                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
506                     sce->zeroes[(w+w2)*16+g] = 1;
507                     continue;
508                 }
509                 sce->zeroes[(w+w2)*16+g] = 0;
510                 nz = 1;
511                 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
512                     float t = fabsf(coefs[w2*128+i]);
513                     if (t > 0.0f)
514                         qmin = FFMIN(qmin, t);
515                     qmax = FFMAX(qmax, t);
516                 }
517             }
518             if (nz) {
519                 int minscale, maxscale;
520                 float minrd = INFINITY;
521                 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
522                 minscale = av_clip_uint8(log2(qmin)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
523                 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
524                 maxscale = av_clip_uint8(log2(qmax)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
525                 for (q = minscale; q < maxscale; q++) {
526                     float dists[12], dist;
527                     memset(dists, 0, sizeof(dists));
528                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
529                         FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
530                         int cb;
531                         for (cb = 0; cb <= ESC_BT; cb++)
532                             dists[cb] += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
533                                                             q, cb, lambda / band->threshold, INFINITY, NULL);
534                     }
535                     dist = dists[0];
536                     for (i = 1; i <= ESC_BT; i++)
537                         dist = FFMIN(dist, dists[i]);
538                     minrd = FFMIN(minrd, dist);
539
540                     for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, 256); i++) {
541                         float cost;
542                         int minv, maxv;
543                         if (isinf(paths[idx - 256 + i].cost))
544                             continue;
545                         cost = paths[idx - 256 + i].cost + dist
546                                + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
547                         minv = FFMIN(paths[idx - 256 + i].min_val, q);
548                         maxv = FFMAX(paths[idx - 256 + i].max_val, q);
549                         if (cost < paths[idx + q].cost && maxv-minv < SCALE_MAX_DIFF) {
550                             paths[idx + q].cost    = cost;
551                             paths[idx + q].prev    = idx - 256 + i;
552                             paths[idx + q].min_val = minv;
553                             paths[idx + q].max_val = maxv;
554                         }
555                     }
556                 }
557             } else {
558                 for (q = 0; q < 256; q++) {
559                     if (!isinf(paths[idx - 256 + q].cost)) {
560                         paths[idx + q].cost = paths[idx - 256 + q].cost + 1;
561                         paths[idx + q].prev = idx - 256 + q;
562                         paths[idx + q].min_val = FFMIN(paths[idx - 256 + q].min_val, q);
563                         paths[idx + q].max_val = FFMAX(paths[idx - 256 + q].max_val, q);
564                         continue;
565                     }
566                     for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, 256); i++) {
567                         float cost;
568                         int minv, maxv;
569                         if (isinf(paths[idx - 256 + i].cost))
570                             continue;
571                         cost = paths[idx - 256 + i].cost + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
572                         minv = FFMIN(paths[idx - 256 + i].min_val, q);
573                         maxv = FFMAX(paths[idx - 256 + i].max_val, q);
574                         if (cost < paths[idx + q].cost && maxv-minv < SCALE_MAX_DIFF) {
575                             paths[idx + q].cost    = cost;
576                             paths[idx + q].prev    = idx - 256 + i;
577                             paths[idx + q].min_val = minv;
578                             paths[idx + q].max_val = maxv;
579                         }
580                     }
581                 }
582             }
583             sce->zeroes[w*16+g] = !nz;
584             start += sce->ics.swb_sizes[g];
585             idx   += 256;
586         }
587     }
588     idx -= 256;
589     mincost = paths[idx].cost;
590     minq    = idx;
591     for (i = 1; i < 256; i++) {
592         if (paths[idx + i].cost < mincost) {
593             mincost = paths[idx + i].cost;
594             minq = idx + i;
595         }
596     }
597     while (minq >= 256) {
598         sce->sf_idx[bandaddr[minq>>8]] = minq & 0xFF;
599         minq = paths[minq].prev;
600     }
601     //set the same quantizers inside window groups
602     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
603         for (g = 0;  g < sce->ics.num_swb; g++)
604             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
605                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
606 }
607
608 /**
609  * two-loop quantizers search taken from ISO 13818-7 Appendix C
610  */
611 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
612                                           AACEncContext *s,
613                                           SingleChannelElement *sce,
614                                           const float lambda)
615 {
616     int start = 0, i, w, w2, g;
617     int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels;
618     float dists[128], uplims[128];
619     int fflag, minscaler;
620     int its  = 0;
621     int allz = 0;
622     float minthr = INFINITY;
623
624     //XXX: some heuristic to determine initial quantizers will reduce search time
625     memset(dists, 0, sizeof(dists));
626     //determine zero bands and upper limits
627     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
628         for (g = 0;  g < sce->ics.num_swb; g++) {
629             int nz = 0;
630             float uplim = 0.0f;
631             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
632                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
633                 uplim += band->threshold;
634                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
635                     sce->zeroes[(w+w2)*16+g] = 1;
636                     continue;
637                 }
638                 nz = 1;
639             }
640             uplims[w*16+g] = uplim *512;
641             sce->zeroes[w*16+g] = !nz;
642             if (nz)
643                 minthr = FFMIN(minthr, uplim);
644             allz = FFMAX(allz, nz);
645         }
646     }
647     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
648         for (g = 0;  g < sce->ics.num_swb; g++) {
649             if (sce->zeroes[w*16+g]) {
650                 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
651                 continue;
652             }
653             sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2(uplims[w*16+g]/minthr)*4,59);
654         }
655     }
656
657     if (!allz)
658         return;
659     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
660     //perform two-loop search
661     //outer loop - improve quality
662     do {
663         int tbits, qstep;
664         minscaler = sce->sf_idx[0];
665         //inner loop - quantize spectrum to fit into given number of bits
666         qstep = its ? 1 : 32;
667         do {
668             int prev = -1;
669             tbits = 0;
670             fflag = 0;
671             for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
672                 start = w*128;
673                 for (g = 0;  g < sce->ics.num_swb; g++) {
674                     const float *coefs = sce->coeffs + start;
675                     const float *scaled = s->scoefs + start;
676                     int bits = 0;
677                     int cb;
678                     float mindist = INFINITY;
679                     int minbits = 0;
680
681                     if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218)
682                         continue;
683                     minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
684                     for (cb = 0; cb <= ESC_BT; cb++) {
685                         float dist = 0.0f;
686                         int bb = 0;
687                         for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
688                             int b;
689                             dist += quantize_band_cost(s, coefs + w2*128,
690                                                        scaled + w2*128,
691                                                        sce->ics.swb_sizes[g],
692                                                        sce->sf_idx[w*16+g],
693                                                        cb,
694                                                        lambda,
695                                                        INFINITY,
696                                                        &b);
697                             bb += b;
698                         }
699                         if (dist < mindist) {
700                             mindist = dist;
701                             minbits = bb;
702                         }
703                     }
704                     dists[w*16+g] = (mindist - minbits) / lambda;
705                     bits = minbits;
706                     if (prev != -1) {
707                         bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
708                     }
709                     tbits += bits;
710                     start += sce->ics.swb_sizes[g];
711                     prev = sce->sf_idx[w*16+g];
712                 }
713             }
714             if (tbits > destbits) {
715                 for (i = 0; i < 128; i++)
716                     if (sce->sf_idx[i] < 218 - qstep)
717                         sce->sf_idx[i] += qstep;
718             } else {
719                 for (i = 0; i < 128; i++)
720                     if (sce->sf_idx[i] > 60 - qstep)
721                         sce->sf_idx[i] -= qstep;
722             }
723             qstep >>= 1;
724             if (!qstep && tbits > destbits*1.02)
725                 qstep = 1;
726             if (sce->sf_idx[0] >= 217)
727                 break;
728         } while (qstep);
729
730         fflag = 0;
731         minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
732         for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
733             start = w*128;
734             for (g = 0; g < sce->ics.num_swb; g++) {
735                 int prevsc = sce->sf_idx[w*16+g];
736                 if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60)
737                     sce->sf_idx[w*16+g]--;
738                 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
739                 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
740                 if (sce->sf_idx[w*16+g] != prevsc)
741                     fflag = 1;
742             }
743         }
744         its++;
745     } while (fflag && its < 10);
746 }
747
748 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
749                                        SingleChannelElement *sce,
750                                        const float lambda)
751 {
752     int start = 0, i, w, w2, g;
753     float uplim[128], maxq[128];
754     int minq, maxsf;
755     float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
756     int last = 0, lastband = 0, curband = 0;
757     float avg_energy = 0.0;
758     if (sce->ics.num_windows == 1) {
759         start = 0;
760         for (i = 0; i < 1024; i++) {
761             if (i - start >= sce->ics.swb_sizes[curband]) {
762                 start += sce->ics.swb_sizes[curband];
763                 curband++;
764             }
765             if (sce->coeffs[i]) {
766                 avg_energy += sce->coeffs[i] * sce->coeffs[i];
767                 last = i;
768                 lastband = curband;
769             }
770         }
771     } else {
772         for (w = 0; w < 8; w++) {
773             const float *coeffs = sce->coeffs + w*128;
774             start = 0;
775             for (i = 0; i < 128; i++) {
776                 if (i - start >= sce->ics.swb_sizes[curband]) {
777                     start += sce->ics.swb_sizes[curband];
778                     curband++;
779                 }
780                 if (coeffs[i]) {
781                     avg_energy += coeffs[i] * coeffs[i];
782                     last = FFMAX(last, i);
783                     lastband = FFMAX(lastband, curband);
784                 }
785             }
786         }
787     }
788     last++;
789     avg_energy /= last;
790     if (avg_energy == 0.0f) {
791         for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
792             sce->sf_idx[i] = SCALE_ONE_POS;
793         return;
794     }
795     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
796         start = w*128;
797         for (g = 0; g < sce->ics.num_swb; g++) {
798             float *coefs   = sce->coeffs + start;
799             const int size = sce->ics.swb_sizes[g];
800             int start2 = start, end2 = start + size, peakpos = start;
801             float maxval = -1, thr = 0.0f, t;
802             maxq[w*16+g] = 0.0f;
803             if (g > lastband) {
804                 maxq[w*16+g] = 0.0f;
805                 start += size;
806                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
807                     memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
808                 continue;
809             }
810             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
811                 for (i = 0; i < size; i++) {
812                     float t = coefs[w2*128+i]*coefs[w2*128+i];
813                     maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
814                     thr += t;
815                     if (sce->ics.num_windows == 1 && maxval < t) {
816                         maxval  = t;
817                         peakpos = start+i;
818                     }
819                 }
820             }
821             if (sce->ics.num_windows == 1) {
822                 start2 = FFMAX(peakpos - 2, start2);
823                 end2   = FFMIN(peakpos + 3, end2);
824             } else {
825                 start2 -= start;
826                 end2   -= start;
827             }
828             start += size;
829             thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
830             t   = 1.0 - (1.0 * start2 / last);
831             uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
832         }
833     }
834     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
835     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
836     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
837         start = w*128;
838         for (g = 0;  g < sce->ics.num_swb; g++) {
839             const float *coefs  = sce->coeffs + start;
840             const float *scaled = s->scoefs   + start;
841             const int size      = sce->ics.swb_sizes[g];
842             int scf, prev_scf, step;
843             int min_scf = 0, max_scf = 255;
844             float curdiff;
845             if (maxq[w*16+g] < 21.544) {
846                 sce->zeroes[w*16+g] = 1;
847                 start += size;
848                 continue;
849             }
850             sce->zeroes[w*16+g] = 0;
851             scf  = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2(1/maxq[w*16+g])*16/3, 60, 218);
852             step = 16;
853             for (;;) {
854                 float dist = 0.0f;
855                 int quant_max;
856
857                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
858                     int b;
859                     dist += quantize_band_cost(s, coefs + w2*128,
860                                                scaled + w2*128,
861                                                sce->ics.swb_sizes[g],
862                                                scf,
863                                                ESC_BT,
864                                                lambda,
865                                                INFINITY,
866                                                &b);
867                     dist -= b;
868                 }
869                 dist *= 1.0f / 512.0f / lambda;
870                 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[200 - scf + SCALE_ONE_POS - SCALE_DIV_512]);
871                 if (quant_max >= 8191) { // too much, return to the previous quantizer
872                     sce->sf_idx[w*16+g] = prev_scf;
873                     break;
874                 }
875                 prev_scf = scf;
876                 curdiff = fabsf(dist - uplim[w*16+g]);
877                 if (curdiff == 0.0f)
878                     step = 0;
879                 else
880                     step = fabsf(log2(curdiff));
881                 if (dist > uplim[w*16+g])
882                     step = -step;
883                 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
884                     sce->sf_idx[w*16+g] = scf;
885                     break;
886                 }
887                 scf += step;
888                 if (step > 0)
889                     min_scf = scf;
890                 else
891                     max_scf = scf;
892             }
893             start += size;
894         }
895     }
896     minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
897     for (i = 1; i < 128; i++) {
898         if (!sce->sf_idx[i])
899             sce->sf_idx[i] = sce->sf_idx[i-1];
900         else
901             minq = FFMIN(minq, sce->sf_idx[i]);
902     }
903     if (minq == INT_MAX)
904         minq = 0;
905     minq = FFMIN(minq, SCALE_MAX_POS);
906     maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
907     for (i = 126; i >= 0; i--) {
908         if (!sce->sf_idx[i])
909             sce->sf_idx[i] = sce->sf_idx[i+1];
910         sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
911     }
912 }
913
914 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
915                                        SingleChannelElement *sce,
916                                        const float lambda)
917 {
918     int start = 0, i, w, w2, g;
919     int minq = 255;
920
921     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
922     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
923         start = w*128;
924         for (g = 0; g < sce->ics.num_swb; g++) {
925             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
926                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
927                 if (band->energy <= band->threshold) {
928                     sce->sf_idx[(w+w2)*16+g] = 218;
929                     sce->zeroes[(w+w2)*16+g] = 1;
930                 } else {
931                     sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2(band->threshold), 80, 218);
932                     sce->zeroes[(w+w2)*16+g] = 0;
933                 }
934                 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
935             }
936         }
937     }
938     for (i = 0; i < 128; i++) {
939         sce->sf_idx[i] = 140;
940         //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
941     }
942     //set the same quantizers inside window groups
943     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
944         for (g = 0;  g < sce->ics.num_swb; g++)
945             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
946                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
947 }
948
949 static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
950                           const float lambda)
951 {
952     int start = 0, i, w, w2, g;
953     float M[128], S[128];
954     float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
955     SingleChannelElement *sce0 = &cpe->ch[0];
956     SingleChannelElement *sce1 = &cpe->ch[1];
957     if (!cpe->common_window)
958         return;
959     for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
960         for (g = 0;  g < sce0->ics.num_swb; g++) {
961             if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
962                 float dist1 = 0.0f, dist2 = 0.0f;
963                 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
964                     FFPsyBand *band0 = &s->psy.psy_bands[(s->cur_channel+0)*PSY_MAX_BANDS+(w+w2)*16+g];
965                     FFPsyBand *band1 = &s->psy.psy_bands[(s->cur_channel+1)*PSY_MAX_BANDS+(w+w2)*16+g];
966                     float minthr = FFMIN(band0->threshold, band1->threshold);
967                     float maxthr = FFMAX(band0->threshold, band1->threshold);
968                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
969                         M[i] = (sce0->coeffs[start+w2*128+i]
970                               + sce1->coeffs[start+w2*128+i]) * 0.5;
971                         S[i] =  sce0->coeffs[start+w2*128+i]
972                               - sce1->coeffs[start+w2*128+i];
973                     }
974                     abs_pow34_v(L34, sce0->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
975                     abs_pow34_v(R34, sce1->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
976                     abs_pow34_v(M34, M,                         sce0->ics.swb_sizes[g]);
977                     abs_pow34_v(S34, S,                         sce0->ics.swb_sizes[g]);
978                     dist1 += quantize_band_cost(s, sce0->coeffs + start + w2*128,
979                                                 L34,
980                                                 sce0->ics.swb_sizes[g],
981                                                 sce0->sf_idx[(w+w2)*16+g],
982                                                 sce0->band_type[(w+w2)*16+g],
983                                                 lambda / band0->threshold, INFINITY, NULL);
984                     dist1 += quantize_band_cost(s, sce1->coeffs + start + w2*128,
985                                                 R34,
986                                                 sce1->ics.swb_sizes[g],
987                                                 sce1->sf_idx[(w+w2)*16+g],
988                                                 sce1->band_type[(w+w2)*16+g],
989                                                 lambda / band1->threshold, INFINITY, NULL);
990                     dist2 += quantize_band_cost(s, M,
991                                                 M34,
992                                                 sce0->ics.swb_sizes[g],
993                                                 sce0->sf_idx[(w+w2)*16+g],
994                                                 sce0->band_type[(w+w2)*16+g],
995                                                 lambda / maxthr, INFINITY, NULL);
996                     dist2 += quantize_band_cost(s, S,
997                                                 S34,
998                                                 sce1->ics.swb_sizes[g],
999                                                 sce1->sf_idx[(w+w2)*16+g],
1000                                                 sce1->band_type[(w+w2)*16+g],
1001                                                 lambda / minthr, INFINITY, NULL);
1002                 }
1003                 cpe->ms_mask[w*16+g] = dist2 < dist1;
1004             }
1005             start += sce0->ics.swb_sizes[g];
1006         }
1007     }
1008 }
1009
1010 AACCoefficientsEncoder ff_aac_coders[] = {
1011     {
1012         search_for_quantizers_faac,
1013         encode_window_bands_info,
1014         quantize_and_encode_band,
1015 //        search_for_ms,
1016     },
1017     {
1018         search_for_quantizers_anmr,
1019         encode_window_bands_info,
1020         quantize_and_encode_band,
1021 //        search_for_ms,
1022     },
1023     {
1024         search_for_quantizers_twoloop,
1025         encode_window_bands_info,
1026         quantize_and_encode_band,
1027 //        search_for_ms,
1028     },
1029     {
1030         search_for_quantizers_fast,
1031         encode_window_bands_info,
1032         quantize_and_encode_band,
1033 //        search_for_ms,
1034     },
1035 };