]> git.sesse.net Git - ffmpeg/blob - libavcodec/ac3enc_template.c
latm: Always reconfigure if no extradata was set previously
[ffmpeg] / libavcodec / ac3enc_template.c
1 /*
2  * AC-3 encoder float/fixed template
3  * Copyright (c) 2000 Fabrice Bellard
4  * Copyright (c) 2006-2011 Justin Ruggles <justin.ruggles@gmail.com>
5  * Copyright (c) 2006-2010 Prakash Punnoor <prakash@punnoor.de>
6  *
7  * This file is part of Libav.
8  *
9  * Libav is free software; you can redistribute it and/or
10  * modify it under the terms of the GNU Lesser General Public
11  * License as published by the Free Software Foundation; either
12  * version 2.1 of the License, or (at your option) any later version.
13  *
14  * Libav is distributed in the hope that it will be useful,
15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
17  * Lesser General Public License for more details.
18  *
19  * You should have received a copy of the GNU Lesser General Public
20  * License along with Libav; if not, write to the Free Software
21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22  */
23
24 /**
25  * @file
26  * AC-3 encoder float/fixed template
27  */
28
29 #include <stdint.h>
30
31 #include "libavutil/internal.h"
32
33 /* prototypes for static functions in ac3enc_fixed.c and ac3enc_float.c */
34
35 static void scale_coefficients(AC3EncodeContext *s);
36
37 static int normalize_samples(AC3EncodeContext *s);
38
39 static void clip_coefficients(DSPContext *dsp, CoefType *coef, unsigned int len);
40
41 static CoefType calc_cpl_coord(CoefSumType energy_ch, CoefSumType energy_cpl);
42
43
44 int AC3_NAME(allocate_sample_buffers)(AC3EncodeContext *s)
45 {
46     int ch;
47
48     FF_ALLOC_OR_GOTO(s->avctx, s->windowed_samples, AC3_WINDOW_SIZE *
49                      sizeof(*s->windowed_samples), alloc_fail);
50     FF_ALLOC_OR_GOTO(s->avctx, s->planar_samples, s->channels * sizeof(*s->planar_samples),
51                      alloc_fail);
52     for (ch = 0; ch < s->channels; ch++) {
53         FF_ALLOCZ_OR_GOTO(s->avctx, s->planar_samples[ch],
54                           (AC3_FRAME_SIZE+AC3_BLOCK_SIZE) * sizeof(**s->planar_samples),
55                           alloc_fail);
56     }
57
58     return 0;
59 alloc_fail:
60     return AVERROR(ENOMEM);
61 }
62
63
64 /*
65  * Copy input samples.
66  * Channels are reordered from Libav's default order to AC-3 order.
67  */
68 static void copy_input_samples(AC3EncodeContext *s, SampleType **samples)
69 {
70     int ch;
71
72     /* copy and remap input samples */
73     for (ch = 0; ch < s->channels; ch++) {
74         /* copy last 256 samples of previous frame to the start of the current frame */
75         memcpy(&s->planar_samples[ch][0], &s->planar_samples[ch][AC3_BLOCK_SIZE * s->num_blocks],
76                AC3_BLOCK_SIZE * sizeof(s->planar_samples[0][0]));
77
78         /* copy new samples for current frame */
79         memcpy(&s->planar_samples[ch][AC3_BLOCK_SIZE],
80                samples[s->channel_map[ch]],
81                AC3_BLOCK_SIZE * s->num_blocks * sizeof(s->planar_samples[0][0]));
82     }
83 }
84
85
86 /*
87  * Apply the MDCT to input samples to generate frequency coefficients.
88  * This applies the KBD window and normalizes the input to reduce precision
89  * loss due to fixed-point calculations.
90  */
91 static void apply_mdct(AC3EncodeContext *s)
92 {
93     int blk, ch;
94
95     for (ch = 0; ch < s->channels; ch++) {
96         for (blk = 0; blk < s->num_blocks; blk++) {
97             AC3Block *block = &s->blocks[blk];
98             const SampleType *input_samples = &s->planar_samples[ch][blk * AC3_BLOCK_SIZE];
99
100 #if CONFIG_AC3ENC_FLOAT
101             s->fdsp.vector_fmul(s->windowed_samples, input_samples,
102                                 s->mdct_window, AC3_WINDOW_SIZE);
103 #else
104             s->ac3dsp.apply_window_int16(s->windowed_samples, input_samples,
105                                          s->mdct_window, AC3_WINDOW_SIZE);
106 #endif
107
108             if (s->fixed_point)
109                 block->coeff_shift[ch+1] = normalize_samples(s);
110
111             s->mdct.mdct_calcw(&s->mdct, block->mdct_coef[ch+1],
112                                s->windowed_samples);
113         }
114     }
115 }
116
117
118 /*
119  * Calculate coupling channel and coupling coordinates.
120  */
121 static void apply_channel_coupling(AC3EncodeContext *s)
122 {
123     LOCAL_ALIGNED_16(CoefType, cpl_coords,      [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
124 #if CONFIG_AC3ENC_FLOAT
125     LOCAL_ALIGNED_16(int32_t, fixed_cpl_coords, [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
126 #else
127     int32_t (*fixed_cpl_coords)[AC3_MAX_CHANNELS][16] = cpl_coords;
128 #endif
129     int blk, ch, bnd, i, j;
130     CoefSumType energy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][16] = {{{0}}};
131     int cpl_start, num_cpl_coefs;
132
133     memset(cpl_coords,       0, AC3_MAX_BLOCKS * sizeof(*cpl_coords));
134 #if CONFIG_AC3ENC_FLOAT
135     memset(fixed_cpl_coords, 0, AC3_MAX_BLOCKS * sizeof(*cpl_coords));
136 #endif
137
138     /* align start to 16-byte boundary. align length to multiple of 32.
139         note: coupling start bin % 4 will always be 1 */
140     cpl_start     = s->start_freq[CPL_CH] - 1;
141     num_cpl_coefs = FFALIGN(s->num_cpl_subbands * 12 + 1, 32);
142     cpl_start     = FFMIN(256, cpl_start + num_cpl_coefs) - num_cpl_coefs;
143
144     /* calculate coupling channel from fbw channels */
145     for (blk = 0; blk < s->num_blocks; blk++) {
146         AC3Block *block = &s->blocks[blk];
147         CoefType *cpl_coef = &block->mdct_coef[CPL_CH][cpl_start];
148         if (!block->cpl_in_use)
149             continue;
150         memset(cpl_coef, 0, num_cpl_coefs * sizeof(*cpl_coef));
151         for (ch = 1; ch <= s->fbw_channels; ch++) {
152             CoefType *ch_coef = &block->mdct_coef[ch][cpl_start];
153             if (!block->channel_in_cpl[ch])
154                 continue;
155             for (i = 0; i < num_cpl_coefs; i++)
156                 cpl_coef[i] += ch_coef[i];
157         }
158
159         /* coefficients must be clipped in order to be encoded */
160         clip_coefficients(&s->dsp, cpl_coef, num_cpl_coefs);
161     }
162
163     /* calculate energy in each band in coupling channel and each fbw channel */
164     /* TODO: possibly use SIMD to speed up energy calculation */
165     bnd = 0;
166     i = s->start_freq[CPL_CH];
167     while (i < s->cpl_end_freq) {
168         int band_size = s->cpl_band_sizes[bnd];
169         for (ch = CPL_CH; ch <= s->fbw_channels; ch++) {
170             for (blk = 0; blk < s->num_blocks; blk++) {
171                 AC3Block *block = &s->blocks[blk];
172                 if (!block->cpl_in_use || (ch > CPL_CH && !block->channel_in_cpl[ch]))
173                     continue;
174                 for (j = 0; j < band_size; j++) {
175                     CoefType v = block->mdct_coef[ch][i+j];
176                     MAC_COEF(energy[blk][ch][bnd], v, v);
177                 }
178             }
179         }
180         i += band_size;
181         bnd++;
182     }
183
184     /* calculate coupling coordinates for all blocks for all channels */
185     for (blk = 0; blk < s->num_blocks; blk++) {
186         AC3Block *block  = &s->blocks[blk];
187         if (!block->cpl_in_use)
188             continue;
189         for (ch = 1; ch <= s->fbw_channels; ch++) {
190             if (!block->channel_in_cpl[ch])
191                 continue;
192             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
193                 cpl_coords[blk][ch][bnd] = calc_cpl_coord(energy[blk][ch][bnd],
194                                                           energy[blk][CPL_CH][bnd]);
195             }
196         }
197     }
198
199     /* determine which blocks to send new coupling coordinates for */
200     for (blk = 0; blk < s->num_blocks; blk++) {
201         AC3Block *block  = &s->blocks[blk];
202         AC3Block *block0 = blk ? &s->blocks[blk-1] : NULL;
203
204         memset(block->new_cpl_coords, 0, sizeof(block->new_cpl_coords));
205
206         if (block->cpl_in_use) {
207             /* send new coordinates if this is the first block, if previous
208              * block did not use coupling but this block does, the channels
209              * using coupling has changed from the previous block, or the
210              * coordinate difference from the last block for any channel is
211              * greater than a threshold value. */
212             if (blk == 0 || !block0->cpl_in_use) {
213                 for (ch = 1; ch <= s->fbw_channels; ch++)
214                     block->new_cpl_coords[ch] = 1;
215             } else {
216                 for (ch = 1; ch <= s->fbw_channels; ch++) {
217                     if (!block->channel_in_cpl[ch])
218                         continue;
219                     if (!block0->channel_in_cpl[ch]) {
220                         block->new_cpl_coords[ch] = 1;
221                     } else {
222                         CoefSumType coord_diff = 0;
223                         for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
224                             coord_diff += FFABS(cpl_coords[blk-1][ch][bnd] -
225                                                 cpl_coords[blk  ][ch][bnd]);
226                         }
227                         coord_diff /= s->num_cpl_bands;
228                         if (coord_diff > NEW_CPL_COORD_THRESHOLD)
229                             block->new_cpl_coords[ch] = 1;
230                     }
231                 }
232             }
233         }
234     }
235
236     /* calculate final coupling coordinates, taking into account reusing of
237        coordinates in successive blocks */
238     for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
239         blk = 0;
240         while (blk < s->num_blocks) {
241             int av_uninit(blk1);
242             AC3Block *block  = &s->blocks[blk];
243
244             if (!block->cpl_in_use) {
245                 blk++;
246                 continue;
247             }
248
249             for (ch = 1; ch <= s->fbw_channels; ch++) {
250                 CoefSumType energy_ch, energy_cpl;
251                 if (!block->channel_in_cpl[ch])
252                     continue;
253                 energy_cpl = energy[blk][CPL_CH][bnd];
254                 energy_ch = energy[blk][ch][bnd];
255                 blk1 = blk+1;
256                 while (!s->blocks[blk1].new_cpl_coords[ch] && blk1 < s->num_blocks) {
257                     if (s->blocks[blk1].cpl_in_use) {
258                         energy_cpl += energy[blk1][CPL_CH][bnd];
259                         energy_ch += energy[blk1][ch][bnd];
260                     }
261                     blk1++;
262                 }
263                 cpl_coords[blk][ch][bnd] = calc_cpl_coord(energy_ch, energy_cpl);
264             }
265             blk = blk1;
266         }
267     }
268
269     /* calculate exponents/mantissas for coupling coordinates */
270     for (blk = 0; blk < s->num_blocks; blk++) {
271         AC3Block *block = &s->blocks[blk];
272         if (!block->cpl_in_use)
273             continue;
274
275 #if CONFIG_AC3ENC_FLOAT
276         s->ac3dsp.float_to_fixed24(fixed_cpl_coords[blk][1],
277                                    cpl_coords[blk][1],
278                                    s->fbw_channels * 16);
279 #endif
280         s->ac3dsp.extract_exponents(block->cpl_coord_exp[1],
281                                     fixed_cpl_coords[blk][1],
282                                     s->fbw_channels * 16);
283
284         for (ch = 1; ch <= s->fbw_channels; ch++) {
285             int bnd, min_exp, max_exp, master_exp;
286
287             if (!block->new_cpl_coords[ch])
288                 continue;
289
290             /* determine master exponent */
291             min_exp = max_exp = block->cpl_coord_exp[ch][0];
292             for (bnd = 1; bnd < s->num_cpl_bands; bnd++) {
293                 int exp = block->cpl_coord_exp[ch][bnd];
294                 min_exp = FFMIN(exp, min_exp);
295                 max_exp = FFMAX(exp, max_exp);
296             }
297             master_exp = ((max_exp - 15) + 2) / 3;
298             master_exp = FFMAX(master_exp, 0);
299             while (min_exp < master_exp * 3)
300                 master_exp--;
301             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
302                 block->cpl_coord_exp[ch][bnd] = av_clip(block->cpl_coord_exp[ch][bnd] -
303                                                         master_exp * 3, 0, 15);
304             }
305             block->cpl_master_exp[ch] = master_exp;
306
307             /* quantize mantissas */
308             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
309                 int cpl_exp  = block->cpl_coord_exp[ch][bnd];
310                 int cpl_mant = (fixed_cpl_coords[blk][ch][bnd] << (5 + cpl_exp + master_exp * 3)) >> 24;
311                 if (cpl_exp == 15)
312                     cpl_mant >>= 1;
313                 else
314                     cpl_mant -= 16;
315
316                 block->cpl_coord_mant[ch][bnd] = cpl_mant;
317             }
318         }
319     }
320
321     if (CONFIG_EAC3_ENCODER && s->eac3)
322         ff_eac3_set_cpl_states(s);
323 }
324
325
326 /*
327  * Determine rematrixing flags for each block and band.
328  */
329 static void compute_rematrixing_strategy(AC3EncodeContext *s)
330 {
331     int nb_coefs;
332     int blk, bnd, i;
333     AC3Block *block, *block0;
334
335     if (s->channel_mode != AC3_CHMODE_STEREO)
336         return;
337
338     for (blk = 0; blk < s->num_blocks; blk++) {
339         block = &s->blocks[blk];
340         block->new_rematrixing_strategy = !blk;
341
342         block->num_rematrixing_bands = 4;
343         if (block->cpl_in_use) {
344             block->num_rematrixing_bands -= (s->start_freq[CPL_CH] <= 61);
345             block->num_rematrixing_bands -= (s->start_freq[CPL_CH] == 37);
346             if (blk && block->num_rematrixing_bands != block0->num_rematrixing_bands)
347                 block->new_rematrixing_strategy = 1;
348         }
349         nb_coefs = FFMIN(block->end_freq[1], block->end_freq[2]);
350
351         if (!s->rematrixing_enabled) {
352             block0 = block;
353             continue;
354         }
355
356         for (bnd = 0; bnd < block->num_rematrixing_bands; bnd++) {
357             /* calculate calculate sum of squared coeffs for one band in one block */
358             int start = ff_ac3_rematrix_band_tab[bnd];
359             int end   = FFMIN(nb_coefs, ff_ac3_rematrix_band_tab[bnd+1]);
360             CoefSumType sum[4] = {0,};
361             for (i = start; i < end; i++) {
362                 CoefType lt = block->mdct_coef[1][i];
363                 CoefType rt = block->mdct_coef[2][i];
364                 CoefType md = lt + rt;
365                 CoefType sd = lt - rt;
366                 MAC_COEF(sum[0], lt, lt);
367                 MAC_COEF(sum[1], rt, rt);
368                 MAC_COEF(sum[2], md, md);
369                 MAC_COEF(sum[3], sd, sd);
370             }
371
372             /* compare sums to determine if rematrixing will be used for this band */
373             if (FFMIN(sum[2], sum[3]) < FFMIN(sum[0], sum[1]))
374                 block->rematrixing_flags[bnd] = 1;
375             else
376                 block->rematrixing_flags[bnd] = 0;
377
378             /* determine if new rematrixing flags will be sent */
379             if (blk &&
380                 block->rematrixing_flags[bnd] != block0->rematrixing_flags[bnd]) {
381                 block->new_rematrixing_strategy = 1;
382             }
383         }
384         block0 = block;
385     }
386 }
387
388
389 int AC3_NAME(encode_frame)(AVCodecContext *avctx, AVPacket *avpkt,
390                            const AVFrame *frame, int *got_packet_ptr)
391 {
392     AC3EncodeContext *s = avctx->priv_data;
393     int ret;
394
395     if (s->options.allow_per_frame_metadata) {
396         ret = ff_ac3_validate_metadata(s);
397         if (ret)
398             return ret;
399     }
400
401     if (s->bit_alloc.sr_code == 1 || s->eac3)
402         ff_ac3_adjust_frame_size(s);
403
404     copy_input_samples(s, (SampleType **)frame->extended_data);
405
406     apply_mdct(s);
407
408     if (s->fixed_point)
409         scale_coefficients(s);
410
411     clip_coefficients(&s->dsp, s->blocks[0].mdct_coef[1],
412                       AC3_MAX_COEFS * s->num_blocks * s->channels);
413
414     s->cpl_on = s->cpl_enabled;
415     ff_ac3_compute_coupling_strategy(s);
416
417     if (s->cpl_on)
418         apply_channel_coupling(s);
419
420     compute_rematrixing_strategy(s);
421
422     if (!s->fixed_point)
423         scale_coefficients(s);
424
425     ff_ac3_apply_rematrixing(s);
426
427     ff_ac3_process_exponents(s);
428
429     ret = ff_ac3_compute_bit_allocation(s);
430     if (ret) {
431         av_log(avctx, AV_LOG_ERROR, "Bit allocation failed. Try increasing the bitrate.\n");
432         return ret;
433     }
434
435     ff_ac3_group_exponents(s);
436
437     ff_ac3_quantize_mantissas(s);
438
439     if ((ret = ff_alloc_packet(avpkt, s->frame_size))) {
440         av_log(avctx, AV_LOG_ERROR, "Error getting output packet\n");
441         return ret;
442     }
443     ff_ac3_output_frame(s, avpkt->data);
444
445     if (frame->pts != AV_NOPTS_VALUE)
446         avpkt->pts = frame->pts - ff_samples_to_time_base(avctx, avctx->delay);
447
448     *got_packet_ptr = 1;
449     return 0;
450 }