Merge remote-tracking branch 'qatar/master'
[ffmpeg.git] / 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
32 /* prototypes for static functions in ac3enc_fixed.c and ac3enc_float.c */
33
34 static void scale_coefficients(AC3EncodeContext *s);
35
36 static void apply_window(DSPContext *dsp, SampleType *output,
37                          const SampleType *input, const SampleType *window,
38                          unsigned int len);
39
40 static int normalize_samples(AC3EncodeContext *s);
41
42 static void clip_coefficients(DSPContext *dsp, CoefType *coef, unsigned int len);
43
44 static CoefType calc_cpl_coord(CoefSumType energy_ch, CoefSumType energy_cpl);
45
46 static void sum_square_butterfly(AC3EncodeContext *s, CoefSumType sum[4],
47                                  const CoefType *coef0, const CoefType *coef1,
48                                  int len);
49
50 int AC3_NAME(allocate_sample_buffers)(AC3EncodeContext *s)
51 {
52     int ch;
53
54     FF_ALLOC_OR_GOTO(s->avctx, s->windowed_samples, AC3_WINDOW_SIZE *
55                      sizeof(*s->windowed_samples), alloc_fail);
56     FF_ALLOC_OR_GOTO(s->avctx, s->planar_samples, s->channels * sizeof(*s->planar_samples),
57                      alloc_fail);
58     for (ch = 0; ch < s->channels; ch++) {
59         FF_ALLOCZ_OR_GOTO(s->avctx, s->planar_samples[ch],
60                           (AC3_FRAME_SIZE+AC3_BLOCK_SIZE) * sizeof(**s->planar_samples),
61                           alloc_fail);
62     }
63
64     return 0;
65 alloc_fail:
66     return AVERROR(ENOMEM);
67 }
68
69
70 /**
71  * Deinterleave input samples.
72  * Channels are reordered from Libav's default order to AC-3 order.
73  */
74 static void deinterleave_input_samples(AC3EncodeContext *s,
75                                        const SampleType *samples)
76 {
77     int ch, i;
78
79     /* deinterleave and remap input samples */
80     for (ch = 0; ch < s->channels; ch++) {
81         const SampleType *sptr;
82         int sinc;
83
84         /* copy last 256 samples of previous frame to the start of the current frame */
85         memcpy(&s->planar_samples[ch][0], &s->planar_samples[ch][AC3_BLOCK_SIZE * s->num_blocks],
86                AC3_BLOCK_SIZE * sizeof(s->planar_samples[0][0]));
87
88         /* deinterleave */
89         sinc = s->channels;
90         sptr = samples + s->channel_map[ch];
91         for (i = AC3_BLOCK_SIZE; i < AC3_BLOCK_SIZE * (s->num_blocks + 1); i++) {
92             s->planar_samples[ch][i] = *sptr;
93             sptr += sinc;
94         }
95     }
96 }
97
98
99 /**
100  * Apply the MDCT to input samples to generate frequency coefficients.
101  * This applies the KBD window and normalizes the input to reduce precision
102  * loss due to fixed-point calculations.
103  */
104 static void apply_mdct(AC3EncodeContext *s)
105 {
106     int blk, ch;
107
108     for (ch = 0; ch < s->channels; ch++) {
109         for (blk = 0; blk < s->num_blocks; blk++) {
110             AC3Block *block = &s->blocks[blk];
111             const SampleType *input_samples = &s->planar_samples[ch][blk * AC3_BLOCK_SIZE];
112
113             apply_window(&s->dsp, s->windowed_samples, input_samples,
114                          s->mdct_window, AC3_WINDOW_SIZE);
115
116             if (s->fixed_point)
117                 block->coeff_shift[ch+1] = normalize_samples(s);
118
119             s->mdct.mdct_calcw(&s->mdct, block->mdct_coef[ch+1],
120                                s->windowed_samples);
121         }
122     }
123 }
124
125
126 /**
127  * Calculate coupling channel and coupling coordinates.
128  */
129 static void apply_channel_coupling(AC3EncodeContext *s)
130 {
131     LOCAL_ALIGNED_16(CoefType, cpl_coords,      [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
132 #if CONFIG_AC3ENC_FLOAT
133     LOCAL_ALIGNED_16(int32_t, fixed_cpl_coords, [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
134 #else
135     int32_t (*fixed_cpl_coords)[AC3_MAX_CHANNELS][16] = cpl_coords;
136 #endif
137     int blk, ch, bnd, i, j;
138     CoefSumType energy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][16] = {{{0}}};
139     int cpl_start, num_cpl_coefs;
140
141     memset(cpl_coords,       0, AC3_MAX_BLOCKS * sizeof(*cpl_coords));
142 #if CONFIG_AC3ENC_FLOAT
143     memset(fixed_cpl_coords, 0, AC3_MAX_BLOCKS * sizeof(*cpl_coords));
144 #endif
145
146     /* align start to 16-byte boundary. align length to multiple of 32.
147         note: coupling start bin % 4 will always be 1 */
148     cpl_start     = s->start_freq[CPL_CH] - 1;
149     num_cpl_coefs = FFALIGN(s->num_cpl_subbands * 12 + 1, 32);
150     cpl_start     = FFMIN(256, cpl_start + num_cpl_coefs) - num_cpl_coefs;
151
152     /* calculate coupling channel from fbw channels */
153     for (blk = 0; blk < s->num_blocks; blk++) {
154         AC3Block *block = &s->blocks[blk];
155         CoefType *cpl_coef = &block->mdct_coef[CPL_CH][cpl_start];
156         if (!block->cpl_in_use)
157             continue;
158         memset(cpl_coef, 0, num_cpl_coefs * sizeof(*cpl_coef));
159         for (ch = 1; ch <= s->fbw_channels; ch++) {
160             CoefType *ch_coef = &block->mdct_coef[ch][cpl_start];
161             if (!block->channel_in_cpl[ch])
162                 continue;
163             for (i = 0; i < num_cpl_coefs; i++)
164                 cpl_coef[i] += ch_coef[i];
165         }
166
167         /* coefficients must be clipped in order to be encoded */
168         clip_coefficients(&s->dsp, cpl_coef, num_cpl_coefs);
169     }
170
171     /* calculate energy in each band in coupling channel and each fbw channel */
172     /* TODO: possibly use SIMD to speed up energy calculation */
173     bnd = 0;
174     i = s->start_freq[CPL_CH];
175     while (i < s->cpl_end_freq) {
176         int band_size = s->cpl_band_sizes[bnd];
177         for (ch = CPL_CH; ch <= s->fbw_channels; ch++) {
178             for (blk = 0; blk < s->num_blocks; blk++) {
179                 AC3Block *block = &s->blocks[blk];
180                 if (!block->cpl_in_use || (ch > CPL_CH && !block->channel_in_cpl[ch]))
181                     continue;
182                 for (j = 0; j < band_size; j++) {
183                     CoefType v = block->mdct_coef[ch][i+j];
184                     MAC_COEF(energy[blk][ch][bnd], v, v);
185                 }
186             }
187         }
188         i += band_size;
189         bnd++;
190     }
191
192     /* calculate coupling coordinates for all blocks for all channels */
193     for (blk = 0; blk < s->num_blocks; blk++) {
194         AC3Block *block  = &s->blocks[blk];
195         if (!block->cpl_in_use)
196             continue;
197         for (ch = 1; ch <= s->fbw_channels; ch++) {
198             if (!block->channel_in_cpl[ch])
199                 continue;
200             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
201                 cpl_coords[blk][ch][bnd] = calc_cpl_coord(energy[blk][ch][bnd],
202                                                           energy[blk][CPL_CH][bnd]);
203             }
204         }
205     }
206
207     /* determine which blocks to send new coupling coordinates for */
208     for (blk = 0; blk < s->num_blocks; blk++) {
209         AC3Block *block  = &s->blocks[blk];
210         AC3Block *block0 = blk ? &s->blocks[blk-1] : NULL;
211
212         memset(block->new_cpl_coords, 0, sizeof(block->new_cpl_coords));
213
214         if (block->cpl_in_use) {
215             /* send new coordinates if this is the first block, if previous
216              * block did not use coupling but this block does, the channels
217              * using coupling has changed from the previous block, or the
218              * coordinate difference from the last block for any channel is
219              * greater than a threshold value. */
220             if (blk == 0 || !block0->cpl_in_use) {
221                 for (ch = 1; ch <= s->fbw_channels; ch++)
222                     block->new_cpl_coords[ch] = 1;
223             } else {
224                 for (ch = 1; ch <= s->fbw_channels; ch++) {
225                     if (!block->channel_in_cpl[ch])
226                         continue;
227                     if (!block0->channel_in_cpl[ch]) {
228                         block->new_cpl_coords[ch] = 1;
229                     } else {
230                         CoefSumType coord_diff = 0;
231                         for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
232                             coord_diff += FFABS(cpl_coords[blk-1][ch][bnd] -
233                                                 cpl_coords[blk  ][ch][bnd]);
234                         }
235                         coord_diff /= s->num_cpl_bands;
236                         if (coord_diff > NEW_CPL_COORD_THRESHOLD)
237                             block->new_cpl_coords[ch] = 1;
238                     }
239                 }
240             }
241         }
242     }
243
244     /* calculate final coupling coordinates, taking into account reusing of
245        coordinates in successive blocks */
246     for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
247         blk = 0;
248         while (blk < s->num_blocks) {
249             int av_uninit(blk1);
250             AC3Block *block  = &s->blocks[blk];
251
252             if (!block->cpl_in_use) {
253                 blk++;
254                 continue;
255             }
256
257             for (ch = 1; ch <= s->fbw_channels; ch++) {
258                 CoefSumType energy_ch, energy_cpl;
259                 if (!block->channel_in_cpl[ch])
260                     continue;
261                 energy_cpl = energy[blk][CPL_CH][bnd];
262                 energy_ch = energy[blk][ch][bnd];
263                 blk1 = blk+1;
264                 while (!s->blocks[blk1].new_cpl_coords[ch] && blk1 < s->num_blocks) {
265                     if (s->blocks[blk1].cpl_in_use) {
266                         energy_cpl += energy[blk1][CPL_CH][bnd];
267                         energy_ch += energy[blk1][ch][bnd];
268                     }
269                     blk1++;
270                 }
271                 cpl_coords[blk][ch][bnd] = calc_cpl_coord(energy_ch, energy_cpl);
272             }
273             blk = blk1;
274         }
275     }
276
277     /* calculate exponents/mantissas for coupling coordinates */
278     for (blk = 0; blk < s->num_blocks; blk++) {
279         AC3Block *block = &s->blocks[blk];
280         if (!block->cpl_in_use)
281             continue;
282
283 #if CONFIG_AC3ENC_FLOAT
284         s->ac3dsp.float_to_fixed24(fixed_cpl_coords[blk][1],
285                                    cpl_coords[blk][1],
286                                    s->fbw_channels * 16);
287 #endif
288         s->ac3dsp.extract_exponents(block->cpl_coord_exp[1],
289                                     fixed_cpl_coords[blk][1],
290                                     s->fbw_channels * 16);
291
292         for (ch = 1; ch <= s->fbw_channels; ch++) {
293             int bnd, min_exp, max_exp, master_exp;
294
295             if (!block->new_cpl_coords[ch])
296                 continue;
297
298             /* determine master exponent */
299             min_exp = max_exp = block->cpl_coord_exp[ch][0];
300             for (bnd = 1; bnd < s->num_cpl_bands; bnd++) {
301                 int exp = block->cpl_coord_exp[ch][bnd];
302                 min_exp = FFMIN(exp, min_exp);
303                 max_exp = FFMAX(exp, max_exp);
304             }
305             master_exp = ((max_exp - 15) + 2) / 3;
306             master_exp = FFMAX(master_exp, 0);
307             while (min_exp < master_exp * 3)
308                 master_exp--;
309             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
310                 block->cpl_coord_exp[ch][bnd] = av_clip(block->cpl_coord_exp[ch][bnd] -
311                                                         master_exp * 3, 0, 15);
312             }
313             block->cpl_master_exp[ch] = master_exp;
314
315             /* quantize mantissas */
316             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
317                 int cpl_exp  = block->cpl_coord_exp[ch][bnd];
318                 int cpl_mant = (fixed_cpl_coords[blk][ch][bnd] << (5 + cpl_exp + master_exp * 3)) >> 24;
319                 if (cpl_exp == 15)
320                     cpl_mant >>= 1;
321                 else
322                     cpl_mant -= 16;
323
324                 block->cpl_coord_mant[ch][bnd] = cpl_mant;
325             }
326         }
327     }
328
329     if (CONFIG_EAC3_ENCODER && s->eac3)
330         ff_eac3_set_cpl_states(s);
331 }
332
333
334 /**
335  * Determine rematrixing flags for each block and band.
336  */
337 static void compute_rematrixing_strategy(AC3EncodeContext *s)
338 {
339     int nb_coefs;
340     int blk, bnd;
341     AC3Block *block, *av_uninit(block0);
342
343     if (s->channel_mode != AC3_CHMODE_STEREO)
344         return;
345
346     for (blk = 0; blk < s->num_blocks; blk++) {
347         block = &s->blocks[blk];
348         block->new_rematrixing_strategy = !blk;
349
350         block->num_rematrixing_bands = 4;
351         if (block->cpl_in_use) {
352             block->num_rematrixing_bands -= (s->start_freq[CPL_CH] <= 61);
353             block->num_rematrixing_bands -= (s->start_freq[CPL_CH] == 37);
354             if (blk && block->num_rematrixing_bands != block0->num_rematrixing_bands)
355                 block->new_rematrixing_strategy = 1;
356         }
357         nb_coefs = FFMIN(block->end_freq[1], block->end_freq[2]);
358
359         if (!s->rematrixing_enabled) {
360             block0 = block;
361             continue;
362         }
363
364         for (bnd = 0; bnd < block->num_rematrixing_bands; bnd++) {
365             /* calculate calculate sum of squared coeffs for one band in one block */
366             int start = ff_ac3_rematrix_band_tab[bnd];
367             int end   = FFMIN(nb_coefs, ff_ac3_rematrix_band_tab[bnd+1]);
368             CoefSumType sum[4];
369             sum_square_butterfly(s, sum, block->mdct_coef[1] + start,
370                                  block->mdct_coef[2] + start, end - start);
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 /**
390  * Encode a single AC-3 frame.
391  */
392 int AC3_NAME(encode_frame)(AVCodecContext *avctx, unsigned char *frame,
393                            int buf_size, void *data)
394 {
395     AC3EncodeContext *s = avctx->priv_data;
396     const SampleType *samples = data;
397     int ret;
398
399     if (s->options.allow_per_frame_metadata) {
400         ret = ff_ac3_validate_metadata(s);
401         if (ret)
402             return ret;
403     }
404
405     if (s->bit_alloc.sr_code == 1 || s->eac3)
406         ff_ac3_adjust_frame_size(s);
407
408     deinterleave_input_samples(s, samples);
409
410     apply_mdct(s);
411
412     if (s->fixed_point)
413         scale_coefficients(s);
414
415     clip_coefficients(&s->dsp, s->blocks[0].mdct_coef[1],
416                       AC3_MAX_COEFS * s->num_blocks * s->channels);
417
418     s->cpl_on = s->cpl_enabled;
419     ff_ac3_compute_coupling_strategy(s);
420
421     if (s->cpl_on)
422         apply_channel_coupling(s);
423
424     compute_rematrixing_strategy(s);
425
426     if (!s->fixed_point)
427         scale_coefficients(s);
428
429     ff_ac3_apply_rematrixing(s);
430
431     ff_ac3_process_exponents(s);
432
433     ret = ff_ac3_compute_bit_allocation(s);
434     if (ret) {
435         av_log(avctx, AV_LOG_ERROR, "Bit allocation failed. Try increasing the bitrate.\n");
436         return ret;
437     }
438
439     ff_ac3_group_exponents(s);
440
441     ff_ac3_quantize_mantissas(s);
442
443     ff_ac3_output_frame(s, frame);
444
445     return s->frame_size;
446 }