When calculating AAC quantized band cost, don't leave garbage in the bit count
[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,
68                            int size, float Q34, int is_signed, int maxval)
69 {
70     int i;
71     double qc;
72     for (i = 0; i < size; i++) {
73         qc = scaled[i] * Q34;
74         out[i][0] = (int)FFMIN((int)qc,            maxval);
75         out[i][1] = (int)FFMIN((int)(qc + 0.4054), maxval);
76         if (is_signed && in[i] < 0.0f) {
77             out[i][0] = -out[i][0];
78             out[i][1] = -out[i][1];
79         }
80     }
81 }
82
83 static void abs_pow34_v(float *out, const float *in, const int size)
84 {
85 #ifndef USE_REALLY_FULL_SEARCH
86     int i;
87     for (i = 0; i < size; i++)
88         out[i] = pow(fabsf(in[i]), 0.75);
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,
106                                 const float *scaled, int size, int scale_idx,
107                                 int cb, const float lambda, const float uplim,
108                                 int *bits)
109 {
110     const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
111     const float  Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
112     const float CLIPPED_ESCAPE = 165140.0f*IQ;
113     int i, j, k;
114     float cost = 0;
115     const int dim = cb < FIRST_PAIR_BT ? 4 : 2;
116     int resbits = 0;
117 #ifndef USE_REALLY_FULL_SEARCH
118     const float  Q34 = pow(Q, 0.75);
119     const int range  = aac_cb_range[cb];
120     const int maxval = aac_cb_maxval[cb];
121     int offs[4];
122 #endif /* USE_REALLY_FULL_SEARCH */
123
124     if (!cb) {
125         for (i = 0; i < size; i++)
126             cost += in[i]*in[i]*lambda;
127         if (bits)
128             *bits = 0;
129         return cost;
130     }
131 #ifndef USE_REALLY_FULL_SEARCH
132     offs[0] = 1;
133     for (i = 1; i < dim; i++)
134         offs[i] = offs[i-1]*range;
135     quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
136 #endif /* USE_REALLY_FULL_SEARCH */
137     for (i = 0; i < size; i += dim) {
138         float mincost;
139         int minidx  = 0;
140         int minbits = 0;
141         const float *vec;
142 #ifndef USE_REALLY_FULL_SEARCH
143         int (*quants)[2] = &s->qcoefs[i];
144         mincost = 0.0f;
145         for (j = 0; j < dim; j++)
146             mincost += in[i+j]*in[i+j]*lambda;
147         minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
148         minbits = ff_aac_spectral_bits[cb-1][minidx];
149         mincost += minbits;
150         for (j = 0; j < (1<<dim); j++) {
151             float rd = 0.0f;
152             int curbits;
153             int curidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
154             int same   = 0;
155             for (k = 0; k < dim; k++) {
156                 if ((j & (1 << k)) && quants[k][0] == quants[k][1]) {
157                     same = 1;
158                     break;
159                 }
160             }
161             if (same)
162                 continue;
163             for (k = 0; k < dim; k++)
164                 curidx += quants[k][!!(j & (1 << k))] * offs[dim - 1 - k];
165             curbits =  ff_aac_spectral_bits[cb-1][curidx];
166             vec     = &ff_aac_codebook_vectors[cb-1][curidx*dim];
167 #else
168         mincost = INFINITY;
169         vec = ff_aac_codebook_vectors[cb-1];
170         for (j = 0; j < ff_aac_spectral_sizes[cb-1]; j++, vec += dim) {
171             float rd = 0.0f;
172             int curbits = ff_aac_spectral_bits[cb-1][j];
173 #endif /* USE_REALLY_FULL_SEARCH */
174             if (IS_CODEBOOK_UNSIGNED(cb)) {
175                 for (k = 0; k < dim; k++) {
176                     float t = fabsf(in[i+k]);
177                     float di;
178                     //do not code with escape sequence small values
179                     if (vec[k] == 64.0f && t < 39.0f*IQ) {
180                         rd = INFINITY;
181                         break;
182                     }
183                     if (vec[k] == 64.0f) { //FIXME: slow
184                         if (t >= CLIPPED_ESCAPE) {
185                             di = t - CLIPPED_ESCAPE;
186                             curbits += 21;
187                         } else {
188                             int c = av_clip(quant(t, Q), 0, 8191);
189                             di = t - c*cbrt(c)*IQ;
190                             curbits += av_log2(c)*2 - 4 + 1;
191                         }
192                     } else {
193                         di = t - vec[k]*IQ;
194                     }
195                     if (vec[k] != 0.0f)
196                         curbits++;
197                     rd += di*di*lambda;
198                 }
199             } else {
200                 for (k = 0; k < dim; k++) {
201                     float di = in[i+k] - vec[k]*IQ;
202                     rd += di*di*lambda;
203                 }
204             }
205             rd += curbits;
206             if (rd < mincost) {
207                 mincost = rd;
208                 minidx  = j;
209                 minbits = curbits;
210             }
211         }
212         cost    += mincost;
213         resbits += minbits;
214         if (cost >= uplim)
215             return uplim;
216     }
217
218     if (bits)
219         *bits = resbits;
220     return cost;
221 }
222
223 static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
224                                      const float *in, int size, int scale_idx,
225                                      int cb, const float lambda)
226 {
227     const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
228     const float  Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
229     const float CLIPPED_ESCAPE = 165140.0f*IQ;
230     const int dim = (cb < FIRST_PAIR_BT) ? 4 : 2;
231     int i, j, k;
232 #ifndef USE_REALLY_FULL_SEARCH
233     const float  Q34 = pow(Q, 0.75);
234     const int range  = aac_cb_range[cb];
235     const int maxval = aac_cb_maxval[cb];
236     int offs[4];
237     float *scaled = s->scoefs;
238 #endif /* USE_REALLY_FULL_SEARCH */
239
240 //START_TIMER
241     if (!cb)
242         return;
243
244 #ifndef USE_REALLY_FULL_SEARCH
245     offs[0] = 1;
246     for (i = 1; i < dim; i++)
247         offs[i] = offs[i-1]*range;
248     abs_pow34_v(scaled, in, size);
249     quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
250 #endif /* USE_REALLY_FULL_SEARCH */
251     for (i = 0; i < size; i += dim) {
252         float mincost;
253         int minidx  = 0;
254         int minbits = 0;
255         const float *vec;
256 #ifndef USE_REALLY_FULL_SEARCH
257         int (*quants)[2] = &s->qcoefs[i];
258         mincost = 0.0f;
259         for (j = 0; j < dim; j++)
260             mincost += in[i+j]*in[i+j]*lambda;
261         minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
262         minbits = ff_aac_spectral_bits[cb-1][minidx];
263         mincost += minbits;
264         for (j = 0; j < (1<<dim); j++) {
265             float rd = 0.0f;
266             int curbits;
267             int curidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
268             int same   = 0;
269             for (k = 0; k < dim; k++) {
270                 if ((j & (1 << k)) && quants[k][0] == quants[k][1]) {
271                     same = 1;
272                     break;
273                 }
274             }
275             if (same)
276                 continue;
277             for (k = 0; k < dim; k++)
278                 curidx += quants[k][!!(j & (1 << k))] * offs[dim - 1 - k];
279             curbits =  ff_aac_spectral_bits[cb-1][curidx];
280             vec     = &ff_aac_codebook_vectors[cb-1][curidx*dim];
281 #else
282         vec = ff_aac_codebook_vectors[cb-1];
283         mincost = INFINITY;
284         for (j = 0; j < ff_aac_spectral_sizes[cb-1]; j++, vec += dim) {
285             float rd = 0.0f;
286             int curbits = ff_aac_spectral_bits[cb-1][j];
287             int curidx  = j;
288 #endif /* USE_REALLY_FULL_SEARCH */
289             if (IS_CODEBOOK_UNSIGNED(cb)) {
290                 for (k = 0; k < dim; k++) {
291                     float t = fabsf(in[i+k]);
292                     float di;
293                     //do not code with escape sequence small values
294                     if (vec[k] == 64.0f && t < 39.0f*IQ) {
295                         rd = INFINITY;
296                         break;
297                     }
298                     if (vec[k] == 64.0f) { //FIXME: slow
299                         if (t >= CLIPPED_ESCAPE) {
300                             di = t - CLIPPED_ESCAPE;
301                             curbits += 21;
302                         } else {
303                             int c = av_clip(quant(t, Q), 0, 8191);
304                             di = t - c*cbrt(c)*IQ;
305                             curbits += av_log2(c)*2 - 4 + 1;
306                         }
307                     } else {
308                         di = t - vec[k]*IQ;
309                     }
310                     if (vec[k] != 0.0f)
311                         curbits++;
312                     rd += di*di*lambda;
313                 }
314             } else {
315                 for (k = 0; k < dim; k++) {
316                     float di = in[i+k] - vec[k]*IQ;
317                     rd += di*di*lambda;
318                 }
319             }
320             rd += curbits;
321             if (rd < mincost) {
322                 mincost = rd;
323                 minidx  = curidx;
324                 minbits = curbits;
325             }
326         }
327         put_bits(pb, ff_aac_spectral_bits[cb-1][minidx], ff_aac_spectral_codes[cb-1][minidx]);
328         if (IS_CODEBOOK_UNSIGNED(cb))
329             for (j = 0; j < dim; j++)
330                 if (ff_aac_codebook_vectors[cb-1][minidx*dim+j] != 0.0f)
331                     put_bits(pb, 1, in[i+j] < 0.0f);
332         if (cb == ESC_BT) {
333             for (j = 0; j < 2; j++) {
334                 if (ff_aac_codebook_vectors[cb-1][minidx*2+j] == 64.0f) {
335                     int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
336                     int len = av_log2(coef);
337
338                     put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
339                     put_bits(pb, len, coef & ((1 << len) - 1));
340                 }
341             }
342         }
343     }
344 //STOP_TIMER("quantize_and_encode")
345 }
346
347 /**
348  * structure used in optimal codebook search
349  */
350 typedef struct BandCodingPath {
351     int prev_idx; ///< pointer to the previous path point
352     int codebook; ///< codebook for coding band run
353     float cost;   ///< path cost
354     int run;
355 } BandCodingPath;
356
357 /**
358  * Encode band info for single window group bands.
359  */
360 static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
361                                      int win, int group_len, const float lambda)
362 {
363     BandCodingPath path[120][12];
364     int w, swb, cb, start, start2, size;
365     int i, j;
366     const int max_sfb  = sce->ics.max_sfb;
367     const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
368     const int run_esc  = (1 << run_bits) - 1;
369     int idx, ppos, count;
370     int stackrun[120], stackcb[120], stack_len;
371     float next_minrd = INFINITY;
372     int next_mincb = 0;
373
374     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
375     start = win*128;
376     for (cb = 0; cb < 12; cb++) {
377         path[0][cb].cost     = 0.0f;
378         path[0][cb].prev_idx = -1;
379         path[0][cb].run      = 0;
380     }
381     for (swb = 0; swb < max_sfb; swb++) {
382         start2 = start;
383         size = sce->ics.swb_sizes[swb];
384         if (sce->zeroes[win*16 + swb]) {
385             for (cb = 0; cb < 12; cb++) {
386                 path[swb+1][cb].prev_idx = cb;
387                 path[swb+1][cb].cost     = path[swb][cb].cost;
388                 path[swb+1][cb].run      = path[swb][cb].run + 1;
389             }
390         } else {
391             float minrd = next_minrd;
392             int mincb = next_mincb;
393             next_minrd = INFINITY;
394             next_mincb = 0;
395             for (cb = 0; cb < 12; cb++) {
396                 float cost_stay_here, cost_get_here;
397                 float rd = 0.0f;
398                 for (w = 0; w < group_len; w++) {
399                     FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(win+w)*16+swb];
400                     rd += quantize_band_cost(s, sce->coeffs + start + w*128,
401                                              s->scoefs + start + w*128, size,
402                                              sce->sf_idx[(win+w)*16+swb], cb,
403                                              lambda / band->threshold, INFINITY, NULL);
404                 }
405                 cost_stay_here = path[swb][cb].cost + rd;
406                 cost_get_here  = minrd              + rd + run_bits + 4;
407                 if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
408                     != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
409                     cost_stay_here += run_bits;
410                 if (cost_get_here < cost_stay_here) {
411                     path[swb+1][cb].prev_idx = mincb;
412                     path[swb+1][cb].cost     = cost_get_here;
413                     path[swb+1][cb].run      = 1;
414                 } else {
415                     path[swb+1][cb].prev_idx = cb;
416                     path[swb+1][cb].cost     = cost_stay_here;
417                     path[swb+1][cb].run      = path[swb][cb].run + 1;
418                 }
419                 if (path[swb+1][cb].cost < next_minrd) {
420                     next_minrd = path[swb+1][cb].cost;
421                     next_mincb = cb;
422                 }
423             }
424         }
425         start += sce->ics.swb_sizes[swb];
426     }
427
428     //convert resulting path from backward-linked list
429     stack_len = 0;
430     idx       = 0;
431     for (cb = 1; cb < 12; cb++)
432         if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
433             idx = cb;
434     ppos = max_sfb;
435     while (ppos > 0) {
436         cb = idx;
437         stackrun[stack_len] = path[ppos][cb].run;
438         stackcb [stack_len] = cb;
439         idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
440         ppos -= path[ppos][cb].run;
441         stack_len++;
442     }
443     //perform actual band info encoding
444     start = 0;
445     for (i = stack_len - 1; i >= 0; i--) {
446         put_bits(&s->pb, 4, stackcb[i]);
447         count = stackrun[i];
448         memset(sce->zeroes + win*16 + start, !stackcb[i], count);
449         //XXX: memset when band_type is also uint8_t
450         for (j = 0; j < count; j++) {
451             sce->band_type[win*16 + start] =  stackcb[i];
452             start++;
453         }
454         while (count >= run_esc) {
455             put_bits(&s->pb, run_bits, run_esc);
456             count -= run_esc;
457         }
458         put_bits(&s->pb, run_bits, count);
459     }
460 }
461
462 typedef struct TrellisPath {
463     float cost;
464     int prev;
465     int min_val;
466     int max_val;
467 } TrellisPath;
468
469 static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
470                                        SingleChannelElement *sce,
471                                        const float lambda)
472 {
473     int q, w, w2, g, start = 0;
474     int i;
475     int idx;
476     TrellisPath paths[256*121];
477     int bandaddr[121];
478     int minq;
479     float mincost;
480
481     for (i = 0; i < 256; i++) {
482         paths[i].cost    = 0.0f;
483         paths[i].prev    = -1;
484         paths[i].min_val = i;
485         paths[i].max_val = i;
486     }
487     for (i = 256; i < 256*121; i++) {
488         paths[i].cost    = INFINITY;
489         paths[i].prev    = -2;
490         paths[i].min_val = INT_MAX;
491         paths[i].max_val = 0;
492     }
493     idx = 256;
494     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
495     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
496         start = w*128;
497         for (g = 0; g < sce->ics.num_swb; g++) {
498             const float *coefs = sce->coeffs + start;
499             float qmin, qmax;
500             int nz = 0;
501
502             bandaddr[idx >> 8] = w * 16 + g;
503             qmin = INT_MAX;
504             qmax = 0.0f;
505             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
506                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
507                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
508                     sce->zeroes[(w+w2)*16+g] = 1;
509                     continue;
510                 }
511                 sce->zeroes[(w+w2)*16+g] = 0;
512                 nz = 1;
513                 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
514                     float t = fabsf(coefs[w2*128+i]);
515                     if (t > 0.0f)
516                         qmin = FFMIN(qmin, t);
517                     qmax = FFMAX(qmax, t);
518                 }
519             }
520             if (nz) {
521                 int minscale, maxscale;
522                 float minrd = INFINITY;
523                 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
524                 minscale = av_clip_uint8(log2(qmin)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
525                 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
526                 maxscale = av_clip_uint8(log2(qmax)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
527                 for (q = minscale; q < maxscale; q++) {
528                     float dists[12], dist;
529                     memset(dists, 0, sizeof(dists));
530                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
531                         FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
532                         int cb;
533                         for (cb = 0; cb <= ESC_BT; cb++)
534                             dists[cb] += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
535                                                             q, cb, lambda / band->threshold, INFINITY, NULL);
536                     }
537                     dist = dists[0];
538                     for (i = 1; i <= ESC_BT; i++)
539                         dist = FFMIN(dist, dists[i]);
540                     minrd = FFMIN(minrd, dist);
541
542                     for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, 256); i++) {
543                         float cost;
544                         int minv, maxv;
545                         if (isinf(paths[idx - 256 + i].cost))
546                             continue;
547                         cost = paths[idx - 256 + i].cost + dist
548                                + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
549                         minv = FFMIN(paths[idx - 256 + i].min_val, q);
550                         maxv = FFMAX(paths[idx - 256 + i].max_val, q);
551                         if (cost < paths[idx + q].cost && maxv-minv < SCALE_MAX_DIFF) {
552                             paths[idx + q].cost    = cost;
553                             paths[idx + q].prev    = idx - 256 + i;
554                             paths[idx + q].min_val = minv;
555                             paths[idx + q].max_val = maxv;
556                         }
557                     }
558                 }
559             } else {
560                 for (q = 0; q < 256; q++) {
561                     if (!isinf(paths[idx - 256 + q].cost)) {
562                         paths[idx + q].cost = paths[idx - 256 + q].cost + 1;
563                         paths[idx + q].prev = idx - 256 + q;
564                         paths[idx + q].min_val = FFMIN(paths[idx - 256 + q].min_val, q);
565                         paths[idx + q].max_val = FFMAX(paths[idx - 256 + q].max_val, q);
566                         continue;
567                     }
568                     for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, 256); i++) {
569                         float cost;
570                         int minv, maxv;
571                         if (isinf(paths[idx - 256 + i].cost))
572                             continue;
573                         cost = paths[idx - 256 + i].cost + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
574                         minv = FFMIN(paths[idx - 256 + i].min_val, q);
575                         maxv = FFMAX(paths[idx - 256 + i].max_val, q);
576                         if (cost < paths[idx + q].cost && maxv-minv < SCALE_MAX_DIFF) {
577                             paths[idx + q].cost    = cost;
578                             paths[idx + q].prev    = idx - 256 + i;
579                             paths[idx + q].min_val = minv;
580                             paths[idx + q].max_val = maxv;
581                         }
582                     }
583                 }
584             }
585             sce->zeroes[w*16+g] = !nz;
586             start += sce->ics.swb_sizes[g];
587             idx   += 256;
588         }
589     }
590     idx -= 256;
591     mincost = paths[idx].cost;
592     minq    = idx;
593     for (i = 1; i < 256; i++) {
594         if (paths[idx + i].cost < mincost) {
595             mincost = paths[idx + i].cost;
596             minq = idx + i;
597         }
598     }
599     while (minq >= 256) {
600         sce->sf_idx[bandaddr[minq>>8]] = minq & 0xFF;
601         minq = paths[minq].prev;
602     }
603     //set the same quantizers inside window groups
604     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
605         for (g = 0;  g < sce->ics.num_swb; g++)
606             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
607                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
608 }
609
610 /**
611  * two-loop quantizers search taken from ISO 13818-7 Appendix C
612  */
613 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
614                                           AACEncContext *s,
615                                           SingleChannelElement *sce,
616                                           const float lambda)
617 {
618     int start = 0, i, w, w2, g;
619     int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels;
620     float dists[128], uplims[128];
621     int fflag, minscaler;
622     int its  = 0;
623     int allz = 0;
624     float minthr = INFINITY;
625
626     //XXX: some heuristic to determine initial quantizers will reduce search time
627     memset(dists, 0, sizeof(dists));
628     //determine zero bands and upper limits
629     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
630         for (g = 0;  g < sce->ics.num_swb; g++) {
631             int nz = 0;
632             float uplim = 0.0f;
633             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
634                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
635                 uplim += band->threshold;
636                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
637                     sce->zeroes[(w+w2)*16+g] = 1;
638                     continue;
639                 }
640                 nz = 1;
641             }
642             uplims[w*16+g] = uplim *512;
643             sce->zeroes[w*16+g] = !nz;
644             if (nz)
645                 minthr = FFMIN(minthr, uplim);
646             allz = FFMAX(allz, nz);
647         }
648     }
649     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
650         for (g = 0;  g < sce->ics.num_swb; g++) {
651             if (sce->zeroes[w*16+g]) {
652                 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
653                 continue;
654             }
655             sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2(uplims[w*16+g]/minthr)*4,59);
656         }
657     }
658
659     if (!allz)
660         return;
661     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
662     //perform two-loop search
663     //outer loop - improve quality
664     do {
665         int tbits, qstep;
666         minscaler = sce->sf_idx[0];
667         //inner loop - quantize spectrum to fit into given number of bits
668         qstep = its ? 1 : 32;
669         do {
670             int prev = -1;
671             tbits = 0;
672             fflag = 0;
673             for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
674                 start = w*128;
675                 for (g = 0;  g < sce->ics.num_swb; g++) {
676                     const float *coefs = sce->coeffs + start;
677                     const float *scaled = s->scoefs + start;
678                     int bits = 0;
679                     int cb;
680                     float mindist = INFINITY;
681                     int minbits = 0;
682
683                     if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218)
684                         continue;
685                     minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
686                     for (cb = 0; cb <= ESC_BT; cb++) {
687                         float dist = 0.0f;
688                         int bb = 0;
689                         for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
690                             int b;
691                             dist += quantize_band_cost(s, coefs + w2*128,
692                                                        scaled + w2*128,
693                                                        sce->ics.swb_sizes[g],
694                                                        sce->sf_idx[w*16+g],
695                                                        ESC_BT,
696                                                        lambda,
697                                                        INFINITY,
698                                                        &b);
699                             bb += b;
700                         }
701                         if (dist < mindist) {
702                             mindist = dist;
703                             minbits = bb;
704                         }
705                     }
706                     dists[w*16+g] = (mindist - minbits) / lambda;
707                     bits = minbits;
708                     if (prev != -1) {
709                         bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
710                     }
711                     tbits += bits;
712                     start += sce->ics.swb_sizes[g];
713                     prev = sce->sf_idx[w*16+g];
714                 }
715             }
716             if (tbits > destbits) {
717                 for (i = 0; i < 128; i++)
718                     if (sce->sf_idx[i] < 218 - qstep)
719                         sce->sf_idx[i] += qstep;
720             } else {
721                 for (i = 0; i < 128; i++)
722                     if (sce->sf_idx[i] > 60 - qstep)
723                         sce->sf_idx[i] -= qstep;
724             }
725             qstep >>= 1;
726             if (!qstep && tbits > destbits*1.02)
727                 qstep = 1;
728             if (sce->sf_idx[0] >= 217)
729                 break;
730         } while (qstep);
731
732         fflag = 0;
733         minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
734         for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
735             start = w*128;
736             for (g = 0; g < sce->ics.num_swb; g++) {
737                 int prevsc = sce->sf_idx[w*16+g];
738                 if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60)
739                     sce->sf_idx[w*16+g]--;
740                 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
741                 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
742                 if (sce->sf_idx[w*16+g] != prevsc)
743                     fflag = 1;
744             }
745         }
746         its++;
747     } while (fflag && its < 10);
748 }
749
750 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
751                                        SingleChannelElement *sce,
752                                        const float lambda)
753 {
754     int start = 0, i, w, w2, g;
755     float uplim[128], maxq[128];
756     int minq, maxsf;
757     float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
758     int last = 0, lastband = 0, curband = 0;
759     float avg_energy = 0.0;
760     if (sce->ics.num_windows == 1) {
761         start = 0;
762         for (i = 0; i < 1024; i++) {
763             if (i - start >= sce->ics.swb_sizes[curband]) {
764                 start += sce->ics.swb_sizes[curband];
765                 curband++;
766             }
767             if (sce->coeffs[i]) {
768                 avg_energy += sce->coeffs[i] * sce->coeffs[i];
769                 last = i;
770                 lastband = curband;
771             }
772         }
773     } else {
774         for (w = 0; w < 8; w++) {
775             const float *coeffs = sce->coeffs + w*128;
776             start = 0;
777             for (i = 0; i < 128; i++) {
778                 if (i - start >= sce->ics.swb_sizes[curband]) {
779                     start += sce->ics.swb_sizes[curband];
780                     curband++;
781                 }
782                 if (coeffs[i]) {
783                     avg_energy += coeffs[i] * coeffs[i];
784                     last = FFMAX(last, i);
785                     lastband = FFMAX(lastband, curband);
786                 }
787             }
788         }
789     }
790     last++;
791     avg_energy /= last;
792     if (avg_energy == 0.0f) {
793         for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
794             sce->sf_idx[i] = SCALE_ONE_POS;
795         return;
796     }
797     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
798         start = w*128;
799         for (g = 0; g < sce->ics.num_swb; g++) {
800             float *coefs   = sce->coeffs + start;
801             const int size = sce->ics.swb_sizes[g];
802             int start2 = start, end2 = start + size, peakpos = start;
803             float maxval = -1, thr = 0.0f, t;
804             maxq[w*16+g] = 0.0f;
805             if (g > lastband) {
806                 maxq[w*16+g] = 0.0f;
807                 start += size;
808                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
809                     memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
810                 continue;
811             }
812             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
813                 for (i = 0; i < size; i++) {
814                     float t = coefs[w2*128+i]*coefs[w2*128+i];
815                     maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
816                     thr += t;
817                     if (sce->ics.num_windows == 1 && maxval < t) {
818                         maxval  = t;
819                         peakpos = start+i;
820                     }
821                 }
822             }
823             if (sce->ics.num_windows == 1) {
824                 start2 = FFMAX(peakpos - 2, start2);
825                 end2   = FFMIN(peakpos + 3, end2);
826             } else {
827                 start2 -= start;
828                 end2   -= start;
829             }
830             start += size;
831             thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
832             t   = 1.0 - (1.0 * start2 / last);
833             uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
834         }
835     }
836     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
837     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
838     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
839         start = w*128;
840         for (g = 0;  g < sce->ics.num_swb; g++) {
841             const float *coefs  = sce->coeffs + start;
842             const float *scaled = s->scoefs   + start;
843             const int size      = sce->ics.swb_sizes[g];
844             int scf, prev_scf, step;
845             int min_scf = 0, max_scf = 255;
846             float curdiff;
847             if (maxq[w*16+g] < 21.544) {
848                 sce->zeroes[w*16+g] = 1;
849                 start += size;
850                 continue;
851             }
852             sce->zeroes[w*16+g] = 0;
853             scf  = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2(1/maxq[w*16+g])*16/3, 60, 218);
854             step = 16;
855             for (;;) {
856                 float dist = 0.0f;
857                 int quant_max;
858
859                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
860                     int b;
861                     dist += quantize_band_cost(s, coefs + w2*128,
862                                                scaled + w2*128,
863                                                sce->ics.swb_sizes[g],
864                                                scf,
865                                                ESC_BT,
866                                                lambda,
867                                                INFINITY,
868                                                &b);
869                     dist -= b;
870                 }
871                 dist *= 1.0f / 512.0f / lambda;
872                 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[200 - scf + SCALE_ONE_POS - SCALE_DIV_512]);
873                 if (quant_max >= 8191) { // too much, return to the previous quantizer
874                     sce->sf_idx[w*16+g] = prev_scf;
875                     break;
876                 }
877                 prev_scf = scf;
878                 curdiff = fabsf(dist - uplim[w*16+g]);
879                 if (curdiff == 0.0f)
880                     step = 0;
881                 else
882                     step = fabsf(log2(curdiff));
883                 if (dist > uplim[w*16+g])
884                     step = -step;
885                 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
886                     sce->sf_idx[w*16+g] = scf;
887                     break;
888                 }
889                 scf += step;
890                 if (step > 0)
891                     min_scf = scf;
892                 else
893                     max_scf = scf;
894             }
895             start += size;
896         }
897     }
898     minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
899     for (i = 1; i < 128; i++) {
900         if (!sce->sf_idx[i])
901             sce->sf_idx[i] = sce->sf_idx[i-1];
902         else
903             minq = FFMIN(minq, sce->sf_idx[i]);
904     }
905     if (minq == INT_MAX)
906         minq = 0;
907     minq = FFMIN(minq, SCALE_MAX_POS);
908     maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
909     for (i = 126; i >= 0; i--) {
910         if (!sce->sf_idx[i])
911             sce->sf_idx[i] = sce->sf_idx[i+1];
912         sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
913     }
914 }
915
916 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
917                                        SingleChannelElement *sce,
918                                        const float lambda)
919 {
920     int start = 0, i, w, w2, g;
921     int minq = 255;
922
923     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
924     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
925         start = w*128;
926         for (g = 0; g < sce->ics.num_swb; g++) {
927             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
928                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
929                 if (band->energy <= band->threshold) {
930                     sce->sf_idx[(w+w2)*16+g] = 218;
931                     sce->zeroes[(w+w2)*16+g] = 1;
932                 } else {
933                     sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2(band->threshold), 80, 218);
934                     sce->zeroes[(w+w2)*16+g] = 0;
935                 }
936                 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
937             }
938         }
939     }
940     for (i = 0; i < 128; i++) {
941         sce->sf_idx[i] = 140;
942         //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
943     }
944     //set the same quantizers inside window groups
945     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
946         for (g = 0;  g < sce->ics.num_swb; g++)
947             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
948                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
949 }
950
951 static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
952                           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 = FFMIN(band0->threshold, band1->threshold);
969                     float maxthr = FFMAX(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,
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 };