]> rtime.felk.cvut.cz Git - frescor/ffmpeg.git/blob - libavcodec/aaccoder.c
Add VDPAU hardware accelerated decoding for MPEG-4 ASP which can be used
[frescor/ffmpeg.git] / libavcodec / aaccoder.c
1 /*
2  * AAC coefficients encoder
3  * Copyright (C) 2008-2009 Konstantin Shishkov
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file libavcodec/aaccoder.c
24  * AAC coefficients encoder
25  */
26
27 /***********************************
28  *              TODOs:
29  * speedup quantizer selection
30  * add sane pulse detection
31  ***********************************/
32
33 #include "avcodec.h"
34 #include "put_bits.h"
35 #include "aac.h"
36 #include "aacenc.h"
37 #include "aactab.h"
38
39 /** bits needed to code codebook run value for long windows */
40 static const uint8_t run_value_bits_long[64] = {
41      5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,
42      5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5, 10,
43     10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
44     10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 15
45 };
46
47 /** bits needed to code codebook run value for short windows */
48 static const uint8_t run_value_bits_short[16] = {
49     3, 3, 3, 3, 3, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 9
50 };
51
52 static const uint8_t *run_value_bits[2] = {
53     run_value_bits_long, run_value_bits_short
54 };
55
56
57 /**
58  * Quantize one coefficient.
59  * @return absolute value of the quantized coefficient
60  * @see 3GPP TS26.403 5.6.2 "Scalefactor determination"
61  */
62 static av_always_inline int quant(float coef, const float Q)
63 {
64     float a = coef * Q;
65     return sqrtf(a * sqrtf(a)) + 0.4054;
66 }
67
68 static void quantize_bands(int (*out)[2], const float *in, const float *scaled,
69                            int size, float Q34, int is_signed, int maxval)
70 {
71     int i;
72     double qc;
73     for (i = 0; i < size; i++) {
74         qc = scaled[i] * Q34;
75         out[i][0] = (int)FFMIN(qc,          (double)maxval);
76         out[i][1] = (int)FFMIN(qc + 0.4054, (double)maxval);
77         if (is_signed && in[i] < 0.0f) {
78             out[i][0] = -out[i][0];
79             out[i][1] = -out[i][1];
80         }
81     }
82 }
83
84 static void abs_pow34_v(float *out, const float *in, const int size)
85 {
86 #ifndef USE_REALLY_FULL_SEARCH
87     int i;
88     for (i = 0; i < size; i++) {
89         float a = fabsf(in[i]);
90         out[i] = sqrtf(a * sqrtf(a));
91     }
92 #endif /* USE_REALLY_FULL_SEARCH */
93 }
94
95 static const uint8_t aac_cb_range [12] = {0, 3, 3, 3, 3, 9, 9, 8, 8, 13, 13, 17};
96 static const uint8_t aac_cb_maxval[12] = {0, 1, 1, 2, 2, 4, 4, 7, 7, 12, 12, 16};
97
98 /**
99  * Calculate rate distortion cost for quantizing with given codebook
100  *
101  * @return quantization distortion
102  */
103 static float quantize_band_cost(struct AACEncContext *s, const float *in,
104                                 const float *scaled, int size, int scale_idx,
105                                 int cb, const float lambda, const float uplim,
106                                 int *bits)
107 {
108     const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
109     const float  Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
110     const float CLIPPED_ESCAPE = 165140.0f*IQ;
111     int i, j, k;
112     float cost = 0;
113     const int dim = cb < FIRST_PAIR_BT ? 4 : 2;
114     int resbits = 0;
115 #ifndef USE_REALLY_FULL_SEARCH
116     const float  Q34 = sqrtf(Q * sqrtf(Q));
117     const int range  = aac_cb_range[cb];
118     const int maxval = aac_cb_maxval[cb];
119     int offs[4];
120 #endif /* USE_REALLY_FULL_SEARCH */
121
122     if (!cb) {
123         for (i = 0; i < size; i++)
124             cost += in[i]*in[i];
125         if (bits)
126             *bits = 0;
127         return cost * lambda;
128     }
129 #ifndef USE_REALLY_FULL_SEARCH
130     offs[0] = 1;
131     for (i = 1; i < dim; i++)
132         offs[i] = offs[i-1]*range;
133     quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
134 #endif /* USE_REALLY_FULL_SEARCH */
135     for (i = 0; i < size; i += dim) {
136         float mincost;
137         int minidx  = 0;
138         int minbits = 0;
139         const float *vec;
140 #ifndef USE_REALLY_FULL_SEARCH
141         int (*quants)[2] = &s->qcoefs[i];
142         mincost = 0.0f;
143         for (j = 0; j < dim; j++)
144             mincost += in[i+j]*in[i+j];
145         minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
146         minbits = ff_aac_spectral_bits[cb-1][minidx];
147         mincost = mincost * lambda + minbits;
148         for (j = 0; j < (1<<dim); j++) {
149             float rd = 0.0f;
150             int curbits;
151             int curidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
152             int same   = 0;
153             for (k = 0; k < dim; k++) {
154                 if ((j & (1 << k)) && quants[k][0] == quants[k][1]) {
155                     same = 1;
156                     break;
157                 }
158             }
159             if (same)
160                 continue;
161             for (k = 0; k < dim; k++)
162                 curidx += quants[k][!!(j & (1 << k))] * offs[dim - 1 - k];
163             curbits =  ff_aac_spectral_bits[cb-1][curidx];
164             vec     = &ff_aac_codebook_vectors[cb-1][curidx*dim];
165 #else
166         mincost = INFINITY;
167         vec = ff_aac_codebook_vectors[cb-1];
168         for (j = 0; j < ff_aac_spectral_sizes[cb-1]; j++, vec += dim) {
169             float rd = 0.0f;
170             int curbits = ff_aac_spectral_bits[cb-1][j];
171 #endif /* USE_REALLY_FULL_SEARCH */
172             if (IS_CODEBOOK_UNSIGNED(cb)) {
173                 for (k = 0; k < dim; k++) {
174                     float t = fabsf(in[i+k]);
175                     float di;
176                     if (vec[k] == 64.0f) { //FIXME: slow
177                         //do not code with escape sequence small values
178                         if (t < 39.0f*IQ) {
179                             rd = INFINITY;
180                             break;
181                         }
182                         if (t >= CLIPPED_ESCAPE) {
183                             di = t - CLIPPED_ESCAPE;
184                             curbits += 21;
185                         } else {
186                             int c = av_clip(quant(t, Q), 0, 8191);
187                             di = t - c*cbrtf(c)*IQ;
188                             curbits += av_log2(c)*2 - 4 + 1;
189                         }
190                     } else {
191                         di = t - vec[k]*IQ;
192                     }
193                     if (vec[k] != 0.0f)
194                         curbits++;
195                     rd += di*di;
196                 }
197             } else {
198                 for (k = 0; k < dim; k++) {
199                     float di = in[i+k] - vec[k]*IQ;
200                     rd += di*di;
201                 }
202             }
203             rd = rd * lambda + curbits;
204             if (rd < mincost) {
205                 mincost = rd;
206                 minidx  = j;
207                 minbits = curbits;
208             }
209         }
210         cost    += mincost;
211         resbits += minbits;
212         if (cost >= uplim)
213             return uplim;
214     }
215
216     if (bits)
217         *bits = resbits;
218     return cost;
219 }
220
221 static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
222                                      const float *in, int size, int scale_idx,
223                                      int cb, const float lambda)
224 {
225     const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
226     const float  Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512];
227     const float CLIPPED_ESCAPE = 165140.0f*IQ;
228     const int dim = (cb < FIRST_PAIR_BT) ? 4 : 2;
229     int i, j, k;
230 #ifndef USE_REALLY_FULL_SEARCH
231     const float  Q34 = sqrtf(Q * sqrtf(Q));
232     const int range  = aac_cb_range[cb];
233     const int maxval = aac_cb_maxval[cb];
234     int offs[4];
235     float *scaled = s->scoefs;
236 #endif /* USE_REALLY_FULL_SEARCH */
237
238 //START_TIMER
239     if (!cb)
240         return;
241
242 #ifndef USE_REALLY_FULL_SEARCH
243     offs[0] = 1;
244     for (i = 1; i < dim; i++)
245         offs[i] = offs[i-1]*range;
246     abs_pow34_v(scaled, in, size);
247     quantize_bands(s->qcoefs, in, scaled, size, Q34, !IS_CODEBOOK_UNSIGNED(cb), maxval);
248 #endif /* USE_REALLY_FULL_SEARCH */
249     for (i = 0; i < size; i += dim) {
250         float mincost;
251         int minidx  = 0;
252         int minbits = 0;
253         const float *vec;
254 #ifndef USE_REALLY_FULL_SEARCH
255         int (*quants)[2] = &s->qcoefs[i];
256         mincost = 0.0f;
257         for (j = 0; j < dim; j++)
258             mincost += in[i+j]*in[i+j];
259         minidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
260         minbits = ff_aac_spectral_bits[cb-1][minidx];
261         mincost = mincost * lambda + minbits;
262         for (j = 0; j < (1<<dim); j++) {
263             float rd = 0.0f;
264             int curbits;
265             int curidx = IS_CODEBOOK_UNSIGNED(cb) ? 0 : 40;
266             int same   = 0;
267             for (k = 0; k < dim; k++) {
268                 if ((j & (1 << k)) && quants[k][0] == quants[k][1]) {
269                     same = 1;
270                     break;
271                 }
272             }
273             if (same)
274                 continue;
275             for (k = 0; k < dim; k++)
276                 curidx += quants[k][!!(j & (1 << k))] * offs[dim - 1 - k];
277             curbits =  ff_aac_spectral_bits[cb-1][curidx];
278             vec     = &ff_aac_codebook_vectors[cb-1][curidx*dim];
279 #else
280         vec = ff_aac_codebook_vectors[cb-1];
281         mincost = INFINITY;
282         for (j = 0; j < ff_aac_spectral_sizes[cb-1]; j++, vec += dim) {
283             float rd = 0.0f;
284             int curbits = ff_aac_spectral_bits[cb-1][j];
285             int curidx  = j;
286 #endif /* USE_REALLY_FULL_SEARCH */
287             if (IS_CODEBOOK_UNSIGNED(cb)) {
288                 for (k = 0; k < dim; k++) {
289                     float t = fabsf(in[i+k]);
290                     float di;
291                     if (vec[k] == 64.0f) { //FIXME: slow
292                         //do not code with escape sequence small values
293                         if (t < 39.0f*IQ) {
294                             rd = INFINITY;
295                             break;
296                         }
297                         if (t >= CLIPPED_ESCAPE) {
298                             di = t - CLIPPED_ESCAPE;
299                             curbits += 21;
300                         } else {
301                             int c = av_clip(quant(t, Q), 0, 8191);
302                             di = t - c*cbrtf(c)*IQ;
303                             curbits += av_log2(c)*2 - 4 + 1;
304                         }
305                     } else {
306                         di = t - vec[k]*IQ;
307                     }
308                     if (vec[k] != 0.0f)
309                         curbits++;
310                     rd += di*di;
311                 }
312             } else {
313                 for (k = 0; k < dim; k++) {
314                     float di = in[i+k] - vec[k]*IQ;
315                     rd += di*di;
316                 }
317             }
318             rd = rd * lambda + curbits;
319             if (rd < mincost) {
320                 mincost = rd;
321                 minidx  = curidx;
322                 minbits = curbits;
323             }
324         }
325         put_bits(pb, ff_aac_spectral_bits[cb-1][minidx], ff_aac_spectral_codes[cb-1][minidx]);
326         if (IS_CODEBOOK_UNSIGNED(cb))
327             for (j = 0; j < dim; j++)
328                 if (ff_aac_codebook_vectors[cb-1][minidx*dim+j] != 0.0f)
329                     put_bits(pb, 1, in[i+j] < 0.0f);
330         if (cb == ESC_BT) {
331             for (j = 0; j < 2; j++) {
332                 if (ff_aac_codebook_vectors[cb-1][minidx*2+j] == 64.0f) {
333                     int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
334                     int len = av_log2(coef);
335
336                     put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
337                     put_bits(pb, len, coef & ((1 << len) - 1));
338                 }
339             }
340         }
341     }
342 //STOP_TIMER("quantize_and_encode")
343 }
344
345 /**
346  * structure used in optimal codebook search
347  */
348 typedef struct BandCodingPath {
349     int prev_idx; ///< pointer to the previous path point
350     float cost;   ///< path cost
351     int run;
352 } BandCodingPath;
353
354 /**
355  * Encode band info for single window group bands.
356  */
357 static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
358                                      int win, int group_len, const float lambda)
359 {
360     BandCodingPath path[120][12];
361     int w, swb, cb, start, start2, size;
362     int i, j;
363     const int max_sfb  = sce->ics.max_sfb;
364     const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
365     const int run_esc  = (1 << run_bits) - 1;
366     int idx, ppos, count;
367     int stackrun[120], stackcb[120], stack_len;
368     float next_minrd = INFINITY;
369     int next_mincb = 0;
370
371     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
372     start = win*128;
373     for (cb = 0; cb < 12; cb++) {
374         path[0][cb].cost     = 0.0f;
375         path[0][cb].prev_idx = -1;
376         path[0][cb].run      = 0;
377     }
378     for (swb = 0; swb < max_sfb; swb++) {
379         start2 = start;
380         size = sce->ics.swb_sizes[swb];
381         if (sce->zeroes[win*16 + swb]) {
382             for (cb = 0; cb < 12; cb++) {
383                 path[swb+1][cb].prev_idx = cb;
384                 path[swb+1][cb].cost     = path[swb][cb].cost;
385                 path[swb+1][cb].run      = path[swb][cb].run + 1;
386             }
387         } else {
388             float minrd = next_minrd;
389             int mincb = next_mincb;
390             next_minrd = INFINITY;
391             next_mincb = 0;
392             for (cb = 0; cb < 12; cb++) {
393                 float cost_stay_here, cost_get_here;
394                 float rd = 0.0f;
395                 for (w = 0; w < group_len; w++) {
396                     FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(win+w)*16+swb];
397                     rd += quantize_band_cost(s, sce->coeffs + start + w*128,
398                                              s->scoefs + start + w*128, size,
399                                              sce->sf_idx[(win+w)*16+swb], cb,
400                                              lambda / band->threshold, INFINITY, NULL);
401                 }
402                 cost_stay_here = path[swb][cb].cost + rd;
403                 cost_get_here  = minrd              + rd + run_bits + 4;
404                 if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
405                     != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
406                     cost_stay_here += run_bits;
407                 if (cost_get_here < cost_stay_here) {
408                     path[swb+1][cb].prev_idx = mincb;
409                     path[swb+1][cb].cost     = cost_get_here;
410                     path[swb+1][cb].run      = 1;
411                 } else {
412                     path[swb+1][cb].prev_idx = cb;
413                     path[swb+1][cb].cost     = cost_stay_here;
414                     path[swb+1][cb].run      = path[swb][cb].run + 1;
415                 }
416                 if (path[swb+1][cb].cost < next_minrd) {
417                     next_minrd = path[swb+1][cb].cost;
418                     next_mincb = cb;
419                 }
420             }
421         }
422         start += sce->ics.swb_sizes[swb];
423     }
424
425     //convert resulting path from backward-linked list
426     stack_len = 0;
427     idx       = 0;
428     for (cb = 1; cb < 12; cb++)
429         if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
430             idx = cb;
431     ppos = max_sfb;
432     while (ppos > 0) {
433         cb = idx;
434         stackrun[stack_len] = path[ppos][cb].run;
435         stackcb [stack_len] = cb;
436         idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
437         ppos -= path[ppos][cb].run;
438         stack_len++;
439     }
440     //perform actual band info encoding
441     start = 0;
442     for (i = stack_len - 1; i >= 0; i--) {
443         put_bits(&s->pb, 4, stackcb[i]);
444         count = stackrun[i];
445         memset(sce->zeroes + win*16 + start, !stackcb[i], count);
446         //XXX: memset when band_type is also uint8_t
447         for (j = 0; j < count; j++) {
448             sce->band_type[win*16 + start] =  stackcb[i];
449             start++;
450         }
451         while (count >= run_esc) {
452             put_bits(&s->pb, run_bits, run_esc);
453             count -= run_esc;
454         }
455         put_bits(&s->pb, run_bits, count);
456     }
457 }
458
459 typedef struct TrellisPath {
460     float cost;
461     int prev;
462     int min_val;
463     int max_val;
464 } TrellisPath;
465
466 #define TRELLIS_STAGES 121
467 #define TRELLIS_STATES 256
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, j;
475     int idx;
476     TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
477     int bandaddr[TRELLIS_STAGES];
478     int minq;
479     float mincost;
480
481     for (i = 0; i < TRELLIS_STATES; i++) {
482         paths[0][i].cost    = 0.0f;
483         paths[0][i].prev    = -1;
484         paths[0][i].min_val = i;
485         paths[0][i].max_val = i;
486     }
487     for (j = 1; j < TRELLIS_STAGES; j++) {
488         for (i = 0; i < TRELLIS_STATES; i++) {
489             paths[j][i].cost    = INFINITY;
490             paths[j][i].prev    = -2;
491             paths[j][i].min_val = INT_MAX;
492             paths[j][i].max_val = 0;
493         }
494     }
495     idx = 1;
496     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
497     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
498         start = w*128;
499         for (g = 0; g < sce->ics.num_swb; g++) {
500             const float *coefs = sce->coeffs + start;
501             float qmin, qmax;
502             int nz = 0;
503
504             bandaddr[idx] = w * 16 + g;
505             qmin = INT_MAX;
506             qmax = 0.0f;
507             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
508                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
509                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
510                     sce->zeroes[(w+w2)*16+g] = 1;
511                     continue;
512                 }
513                 sce->zeroes[(w+w2)*16+g] = 0;
514                 nz = 1;
515                 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
516                     float t = fabsf(coefs[w2*128+i]);
517                     if (t > 0.0f)
518                         qmin = FFMIN(qmin, t);
519                     qmax = FFMAX(qmax, t);
520                 }
521             }
522             if (nz) {
523                 int minscale, maxscale;
524                 float minrd = INFINITY;
525                 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
526                 minscale = av_clip_uint8(log2(qmin)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
527                 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
528                 maxscale = av_clip_uint8(log2(qmax)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
529                 for (q = minscale; q < maxscale; q++) {
530                     float dists[12], dist;
531                     memset(dists, 0, sizeof(dists));
532                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
533                         FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
534                         int cb;
535                         for (cb = 0; cb <= ESC_BT; cb++)
536                             dists[cb] += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
537                                                             q, cb, lambda / band->threshold, INFINITY, NULL);
538                     }
539                     dist = dists[0];
540                     for (i = 1; i <= ESC_BT; i++)
541                         dist = FFMIN(dist, dists[i]);
542                     minrd = FFMIN(minrd, dist);
543
544                     for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, TRELLIS_STATES); i++) {
545                         float cost;
546                         int minv, maxv;
547                         if (isinf(paths[idx - 1][i].cost))
548                             continue;
549                         cost = paths[idx - 1][i].cost + dist
550                                + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
551                         minv = FFMIN(paths[idx - 1][i].min_val, q);
552                         maxv = FFMAX(paths[idx - 1][i].max_val, q);
553                         if (cost < paths[idx][q].cost && maxv-minv < SCALE_MAX_DIFF) {
554                             paths[idx][q].cost    = cost;
555                             paths[idx][q].prev    = i;
556                             paths[idx][q].min_val = minv;
557                             paths[idx][q].max_val = maxv;
558                         }
559                     }
560                 }
561             } else {
562                 for (q = 0; q < TRELLIS_STATES; q++) {
563                     if (!isinf(paths[idx - 1][q].cost)) {
564                         paths[idx][q].cost = paths[idx - 1][q].cost + 1;
565                         paths[idx][q].prev = q;
566                         paths[idx][q].min_val = FFMIN(paths[idx - 1][q].min_val, q);
567                         paths[idx][q].max_val = FFMAX(paths[idx - 1][q].max_val, q);
568                         continue;
569                     }
570                     for (i = FFMAX(q - SCALE_MAX_DIFF, 0); i < FFMIN(q + SCALE_MAX_DIFF, TRELLIS_STATES); i++) {
571                         float cost;
572                         int minv, maxv;
573                         if (isinf(paths[idx - 1][i].cost))
574                             continue;
575                         cost = paths[idx - 1][i].cost + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
576                         minv = FFMIN(paths[idx - 1][i].min_val, q);
577                         maxv = FFMAX(paths[idx - 1][i].max_val, q);
578                         if (cost < paths[idx][q].cost && maxv-minv < SCALE_MAX_DIFF) {
579                             paths[idx][q].cost    = cost;
580                             paths[idx][q].prev    = i;
581                             paths[idx][q].min_val = minv;
582                             paths[idx][q].max_val = maxv;
583                         }
584                     }
585                 }
586             }
587             sce->zeroes[w*16+g] = !nz;
588             start += sce->ics.swb_sizes[g];
589             idx++;
590         }
591     }
592     idx--;
593     mincost = paths[idx][0].cost;
594     minq    = 0;
595     for (i = 1; i < TRELLIS_STATES; i++) {
596         if (paths[idx][i].cost < mincost) {
597             mincost = paths[idx][i].cost;
598             minq = i;
599         }
600     }
601     while (idx) {
602         sce->sf_idx[bandaddr[idx]] = minq;
603         minq = paths[idx][minq].prev;
604         idx--;
605     }
606     //set the same quantizers inside window groups
607     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
608         for (g = 0;  g < sce->ics.num_swb; g++)
609             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
610                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
611 }
612
613 /**
614  * two-loop quantizers search taken from ISO 13818-7 Appendix C
615  */
616 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
617                                           AACEncContext *s,
618                                           SingleChannelElement *sce,
619                                           const float lambda)
620 {
621     int start = 0, i, w, w2, g;
622     int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels;
623     float dists[128], uplims[128];
624     int fflag, minscaler;
625     int its  = 0;
626     int allz = 0;
627     float minthr = INFINITY;
628
629     //XXX: some heuristic to determine initial quantizers will reduce search time
630     memset(dists, 0, sizeof(dists));
631     //determine zero bands and upper limits
632     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
633         for (g = 0;  g < sce->ics.num_swb; g++) {
634             int nz = 0;
635             float uplim = 0.0f;
636             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
637                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
638                 uplim += band->threshold;
639                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
640                     sce->zeroes[(w+w2)*16+g] = 1;
641                     continue;
642                 }
643                 nz = 1;
644             }
645             uplims[w*16+g] = uplim *512;
646             sce->zeroes[w*16+g] = !nz;
647             if (nz)
648                 minthr = FFMIN(minthr, uplim);
649             allz = FFMAX(allz, nz);
650         }
651     }
652     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
653         for (g = 0;  g < sce->ics.num_swb; g++) {
654             if (sce->zeroes[w*16+g]) {
655                 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
656                 continue;
657             }
658             sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2(uplims[w*16+g]/minthr)*4,59);
659         }
660     }
661
662     if (!allz)
663         return;
664     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
665     //perform two-loop search
666     //outer loop - improve quality
667     do {
668         int tbits, qstep;
669         minscaler = sce->sf_idx[0];
670         //inner loop - quantize spectrum to fit into given number of bits
671         qstep = its ? 1 : 32;
672         do {
673             int prev = -1;
674             tbits = 0;
675             fflag = 0;
676             for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
677                 start = w*128;
678                 for (g = 0;  g < sce->ics.num_swb; g++) {
679                     const float *coefs = sce->coeffs + start;
680                     const float *scaled = s->scoefs + start;
681                     int bits = 0;
682                     int cb;
683                     float mindist = INFINITY;
684                     int minbits = 0;
685
686                     if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
687                         start += sce->ics.swb_sizes[g];
688                         continue;
689                     }
690                     minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
691                     for (cb = 0; cb <= ESC_BT; cb++) {
692                         float dist = 0.0f;
693                         int bb = 0;
694                         for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
695                             int b;
696                             dist += quantize_band_cost(s, coefs + w2*128,
697                                                        scaled + w2*128,
698                                                        sce->ics.swb_sizes[g],
699                                                        sce->sf_idx[w*16+g],
700                                                        cb,
701                                                        lambda,
702                                                        INFINITY,
703                                                        &b);
704                             bb += b;
705                         }
706                         if (dist < mindist) {
707                             mindist = dist;
708                             minbits = bb;
709                         }
710                     }
711                     dists[w*16+g] = (mindist - minbits) / lambda;
712                     bits = minbits;
713                     if (prev != -1) {
714                         bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
715                     }
716                     tbits += bits;
717                     start += sce->ics.swb_sizes[g];
718                     prev = sce->sf_idx[w*16+g];
719                 }
720             }
721             if (tbits > destbits) {
722                 for (i = 0; i < 128; i++)
723                     if (sce->sf_idx[i] < 218 - qstep)
724                         sce->sf_idx[i] += qstep;
725             } else {
726                 for (i = 0; i < 128; i++)
727                     if (sce->sf_idx[i] > 60 - qstep)
728                         sce->sf_idx[i] -= qstep;
729             }
730             qstep >>= 1;
731             if (!qstep && tbits > destbits*1.02)
732                 qstep = 1;
733             if (sce->sf_idx[0] >= 217)
734                 break;
735         } while (qstep);
736
737         fflag = 0;
738         minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
739         for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
740             start = w*128;
741             for (g = 0; g < sce->ics.num_swb; g++) {
742                 int prevsc = sce->sf_idx[w*16+g];
743                 if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60)
744                     sce->sf_idx[w*16+g]--;
745                 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
746                 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
747                 if (sce->sf_idx[w*16+g] != prevsc)
748                     fflag = 1;
749             }
750         }
751         its++;
752     } while (fflag && its < 10);
753 }
754
755 static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
756                                        SingleChannelElement *sce,
757                                        const float lambda)
758 {
759     int start = 0, i, w, w2, g;
760     float uplim[128], maxq[128];
761     int minq, maxsf;
762     float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
763     int last = 0, lastband = 0, curband = 0;
764     float avg_energy = 0.0;
765     if (sce->ics.num_windows == 1) {
766         start = 0;
767         for (i = 0; i < 1024; i++) {
768             if (i - start >= sce->ics.swb_sizes[curband]) {
769                 start += sce->ics.swb_sizes[curband];
770                 curband++;
771             }
772             if (sce->coeffs[i]) {
773                 avg_energy += sce->coeffs[i] * sce->coeffs[i];
774                 last = i;
775                 lastband = curband;
776             }
777         }
778     } else {
779         for (w = 0; w < 8; w++) {
780             const float *coeffs = sce->coeffs + w*128;
781             start = 0;
782             for (i = 0; i < 128; i++) {
783                 if (i - start >= sce->ics.swb_sizes[curband]) {
784                     start += sce->ics.swb_sizes[curband];
785                     curband++;
786                 }
787                 if (coeffs[i]) {
788                     avg_energy += coeffs[i] * coeffs[i];
789                     last = FFMAX(last, i);
790                     lastband = FFMAX(lastband, curband);
791                 }
792             }
793         }
794     }
795     last++;
796     avg_energy /= last;
797     if (avg_energy == 0.0f) {
798         for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
799             sce->sf_idx[i] = SCALE_ONE_POS;
800         return;
801     }
802     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
803         start = w*128;
804         for (g = 0; g < sce->ics.num_swb; g++) {
805             float *coefs   = sce->coeffs + start;
806             const int size = sce->ics.swb_sizes[g];
807             int start2 = start, end2 = start + size, peakpos = start;
808             float maxval = -1, thr = 0.0f, t;
809             maxq[w*16+g] = 0.0f;
810             if (g > lastband) {
811                 maxq[w*16+g] = 0.0f;
812                 start += size;
813                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
814                     memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
815                 continue;
816             }
817             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
818                 for (i = 0; i < size; i++) {
819                     float t = coefs[w2*128+i]*coefs[w2*128+i];
820                     maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
821                     thr += t;
822                     if (sce->ics.num_windows == 1 && maxval < t) {
823                         maxval  = t;
824                         peakpos = start+i;
825                     }
826                 }
827             }
828             if (sce->ics.num_windows == 1) {
829                 start2 = FFMAX(peakpos - 2, start2);
830                 end2   = FFMIN(peakpos + 3, end2);
831             } else {
832                 start2 -= start;
833                 end2   -= start;
834             }
835             start += size;
836             thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
837             t   = 1.0 - (1.0 * start2 / last);
838             uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
839         }
840     }
841     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
842     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
843     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
844         start = w*128;
845         for (g = 0;  g < sce->ics.num_swb; g++) {
846             const float *coefs  = sce->coeffs + start;
847             const float *scaled = s->scoefs   + start;
848             const int size      = sce->ics.swb_sizes[g];
849             int scf, prev_scf, step;
850             int min_scf = 0, max_scf = 255;
851             float curdiff;
852             if (maxq[w*16+g] < 21.544) {
853                 sce->zeroes[w*16+g] = 1;
854                 start += size;
855                 continue;
856             }
857             sce->zeroes[w*16+g] = 0;
858             scf  = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2(1/maxq[w*16+g])*16/3, 60, 218);
859             step = 16;
860             for (;;) {
861                 float dist = 0.0f;
862                 int quant_max;
863
864                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
865                     int b;
866                     dist += quantize_band_cost(s, coefs + w2*128,
867                                                scaled + w2*128,
868                                                sce->ics.swb_sizes[g],
869                                                scf,
870                                                ESC_BT,
871                                                lambda,
872                                                INFINITY,
873                                                &b);
874                     dist -= b;
875                 }
876                 dist *= 1.0f / 512.0f / lambda;
877                 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[200 - scf + SCALE_ONE_POS - SCALE_DIV_512]);
878                 if (quant_max >= 8191) { // too much, return to the previous quantizer
879                     sce->sf_idx[w*16+g] = prev_scf;
880                     break;
881                 }
882                 prev_scf = scf;
883                 curdiff = fabsf(dist - uplim[w*16+g]);
884                 if (curdiff == 0.0f)
885                     step = 0;
886                 else
887                     step = fabsf(log2(curdiff));
888                 if (dist > uplim[w*16+g])
889                     step = -step;
890                 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
891                     sce->sf_idx[w*16+g] = scf;
892                     break;
893                 }
894                 scf += step;
895                 if (step > 0)
896                     min_scf = scf;
897                 else
898                     max_scf = scf;
899             }
900             start += size;
901         }
902     }
903     minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
904     for (i = 1; i < 128; i++) {
905         if (!sce->sf_idx[i])
906             sce->sf_idx[i] = sce->sf_idx[i-1];
907         else
908             minq = FFMIN(minq, sce->sf_idx[i]);
909     }
910     if (minq == INT_MAX)
911         minq = 0;
912     minq = FFMIN(minq, SCALE_MAX_POS);
913     maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
914     for (i = 126; i >= 0; i--) {
915         if (!sce->sf_idx[i])
916             sce->sf_idx[i] = sce->sf_idx[i+1];
917         sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
918     }
919 }
920
921 static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
922                                        SingleChannelElement *sce,
923                                        const float lambda)
924 {
925     int start = 0, i, w, w2, g;
926     int minq = 255;
927
928     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
929     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
930         start = w*128;
931         for (g = 0; g < sce->ics.num_swb; g++) {
932             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
933                 FFPsyBand *band = &s->psy.psy_bands[s->cur_channel*PSY_MAX_BANDS+(w+w2)*16+g];
934                 if (band->energy <= band->threshold) {
935                     sce->sf_idx[(w+w2)*16+g] = 218;
936                     sce->zeroes[(w+w2)*16+g] = 1;
937                 } else {
938                     sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2(band->threshold), 80, 218);
939                     sce->zeroes[(w+w2)*16+g] = 0;
940                 }
941                 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
942             }
943         }
944     }
945     for (i = 0; i < 128; i++) {
946         sce->sf_idx[i] = 140;
947         //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
948     }
949     //set the same quantizers inside window groups
950     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
951         for (g = 0;  g < sce->ics.num_swb; g++)
952             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
953                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
954 }
955
956 static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
957                           const float lambda)
958 {
959     int start = 0, i, w, w2, g;
960     float M[128], S[128];
961     float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
962     SingleChannelElement *sce0 = &cpe->ch[0];
963     SingleChannelElement *sce1 = &cpe->ch[1];
964     if (!cpe->common_window)
965         return;
966     for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
967         for (g = 0;  g < sce0->ics.num_swb; g++) {
968             if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
969                 float dist1 = 0.0f, dist2 = 0.0f;
970                 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
971                     FFPsyBand *band0 = &s->psy.psy_bands[(s->cur_channel+0)*PSY_MAX_BANDS+(w+w2)*16+g];
972                     FFPsyBand *band1 = &s->psy.psy_bands[(s->cur_channel+1)*PSY_MAX_BANDS+(w+w2)*16+g];
973                     float minthr = FFMIN(band0->threshold, band1->threshold);
974                     float maxthr = FFMAX(band0->threshold, band1->threshold);
975                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
976                         M[i] = (sce0->coeffs[start+w2*128+i]
977                               + sce1->coeffs[start+w2*128+i]) * 0.5;
978                         S[i] =  sce0->coeffs[start+w2*128+i]
979                               - sce1->coeffs[start+w2*128+i];
980                     }
981                     abs_pow34_v(L34, sce0->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
982                     abs_pow34_v(R34, sce1->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
983                     abs_pow34_v(M34, M,                         sce0->ics.swb_sizes[g]);
984                     abs_pow34_v(S34, S,                         sce0->ics.swb_sizes[g]);
985                     dist1 += quantize_band_cost(s, sce0->coeffs + start + w2*128,
986                                                 L34,
987                                                 sce0->ics.swb_sizes[g],
988                                                 sce0->sf_idx[(w+w2)*16+g],
989                                                 sce0->band_type[(w+w2)*16+g],
990                                                 lambda / band0->threshold, INFINITY, NULL);
991                     dist1 += quantize_band_cost(s, sce1->coeffs + start + w2*128,
992                                                 R34,
993                                                 sce1->ics.swb_sizes[g],
994                                                 sce1->sf_idx[(w+w2)*16+g],
995                                                 sce1->band_type[(w+w2)*16+g],
996                                                 lambda / band1->threshold, INFINITY, NULL);
997                     dist2 += quantize_band_cost(s, M,
998                                                 M34,
999                                                 sce0->ics.swb_sizes[g],
1000                                                 sce0->sf_idx[(w+w2)*16+g],
1001                                                 sce0->band_type[(w+w2)*16+g],
1002                                                 lambda / maxthr, INFINITY, NULL);
1003                     dist2 += quantize_band_cost(s, S,
1004                                                 S34,
1005                                                 sce1->ics.swb_sizes[g],
1006                                                 sce1->sf_idx[(w+w2)*16+g],
1007                                                 sce1->band_type[(w+w2)*16+g],
1008                                                 lambda / minthr, INFINITY, NULL);
1009                 }
1010                 cpe->ms_mask[w*16+g] = dist2 < dist1;
1011             }
1012             start += sce0->ics.swb_sizes[g];
1013         }
1014     }
1015 }
1016
1017 AACCoefficientsEncoder ff_aac_coders[] = {
1018     {
1019         search_for_quantizers_faac,
1020         encode_window_bands_info,
1021         quantize_and_encode_band,
1022         search_for_ms,
1023     },
1024     {
1025         search_for_quantizers_anmr,
1026         encode_window_bands_info,
1027         quantize_and_encode_band,
1028         search_for_ms,
1029     },
1030     {
1031         search_for_quantizers_twoloop,
1032         encode_window_bands_info,
1033         quantize_and_encode_band,
1034         search_for_ms,
1035     },
1036     {
1037         search_for_quantizers_fast,
1038         encode_window_bands_info,
1039         quantize_and_encode_band,
1040         search_for_ms,
1041     },
1042 };