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(void *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 FFmpeg'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 #if CONFIG_AC3ENC_FLOAT
114             apply_window(&s->fdsp, s->windowed_samples, input_samples,
115                          s->mdct_window, AC3_WINDOW_SIZE);
116 #else
117             apply_window(&s->dsp, s->windowed_samples, input_samples,
118                          s->mdct_window, AC3_WINDOW_SIZE);
119 #endif
120
121             if (s->fixed_point)
122                 block->coeff_shift[ch+1] = normalize_samples(s);
123
124             s->mdct.mdct_calcw(&s->mdct, block->mdct_coef[ch+1],
125                                s->windowed_samples);
126         }
127     }
128 }
129
130
131 /*
132  * Calculate coupling channel and coupling coordinates.
133  */
134 static void apply_channel_coupling(AC3EncodeContext *s)
135 {
136     LOCAL_ALIGNED_16(CoefType, cpl_coords,      [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
137 #if CONFIG_AC3ENC_FLOAT
138     LOCAL_ALIGNED_16(int32_t, fixed_cpl_coords, [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
139 #else
140     int32_t (*fixed_cpl_coords)[AC3_MAX_CHANNELS][16] = cpl_coords;
141 #endif
142     int blk, ch, bnd, i, j;
143     CoefSumType energy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][16] = {{{0}}};
144     int cpl_start, num_cpl_coefs;
145
146     memset(cpl_coords,       0, AC3_MAX_BLOCKS * sizeof(*cpl_coords));
147 #if CONFIG_AC3ENC_FLOAT
148     memset(fixed_cpl_coords, 0, AC3_MAX_BLOCKS * sizeof(*cpl_coords));
149 #endif
150
151     /* align start to 16-byte boundary. align length to multiple of 32.
152         note: coupling start bin % 4 will always be 1 */
153     cpl_start     = s->start_freq[CPL_CH] - 1;
154     num_cpl_coefs = FFALIGN(s->num_cpl_subbands * 12 + 1, 32);
155     cpl_start     = FFMIN(256, cpl_start + num_cpl_coefs) - num_cpl_coefs;
156
157     /* calculate coupling channel from fbw channels */
158     for (blk = 0; blk < s->num_blocks; blk++) {
159         AC3Block *block = &s->blocks[blk];
160         CoefType *cpl_coef = &block->mdct_coef[CPL_CH][cpl_start];
161         if (!block->cpl_in_use)
162             continue;
163         memset(cpl_coef, 0, num_cpl_coefs * sizeof(*cpl_coef));
164         for (ch = 1; ch <= s->fbw_channels; ch++) {
165             CoefType *ch_coef = &block->mdct_coef[ch][cpl_start];
166             if (!block->channel_in_cpl[ch])
167                 continue;
168             for (i = 0; i < num_cpl_coefs; i++)
169                 cpl_coef[i] += ch_coef[i];
170         }
171
172         /* coefficients must be clipped in order to be encoded */
173         clip_coefficients(&s->dsp, cpl_coef, num_cpl_coefs);
174     }
175
176     /* calculate energy in each band in coupling channel and each fbw channel */
177     /* TODO: possibly use SIMD to speed up energy calculation */
178     bnd = 0;
179     i = s->start_freq[CPL_CH];
180     while (i < s->cpl_end_freq) {
181         int band_size = s->cpl_band_sizes[bnd];
182         for (ch = CPL_CH; ch <= s->fbw_channels; ch++) {
183             for (blk = 0; blk < s->num_blocks; blk++) {
184                 AC3Block *block = &s->blocks[blk];
185                 if (!block->cpl_in_use || (ch > CPL_CH && !block->channel_in_cpl[ch]))
186                     continue;
187                 for (j = 0; j < band_size; j++) {
188                     CoefType v = block->mdct_coef[ch][i+j];
189                     MAC_COEF(energy[blk][ch][bnd], v, v);
190                 }
191             }
192         }
193         i += band_size;
194         bnd++;
195     }
196
197     /* calculate coupling coordinates for all blocks for all channels */
198     for (blk = 0; blk < s->num_blocks; blk++) {
199         AC3Block *block  = &s->blocks[blk];
200         if (!block->cpl_in_use)
201             continue;
202         for (ch = 1; ch <= s->fbw_channels; ch++) {
203             if (!block->channel_in_cpl[ch])
204                 continue;
205             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
206                 cpl_coords[blk][ch][bnd] = calc_cpl_coord(energy[blk][ch][bnd],
207                                                           energy[blk][CPL_CH][bnd]);
208             }
209         }
210     }
211
212     /* determine which blocks to send new coupling coordinates for */
213     for (blk = 0; blk < s->num_blocks; blk++) {
214         AC3Block *block  = &s->blocks[blk];
215         AC3Block *block0 = blk ? &s->blocks[blk-1] : NULL;
216
217         memset(block->new_cpl_coords, 0, sizeof(block->new_cpl_coords));
218
219         if (block->cpl_in_use) {
220             /* send new coordinates if this is the first block, if previous
221              * block did not use coupling but this block does, the channels
222              * using coupling has changed from the previous block, or the
223              * coordinate difference from the last block for any channel is
224              * greater than a threshold value. */
225             if (blk == 0 || !block0->cpl_in_use) {
226                 for (ch = 1; ch <= s->fbw_channels; ch++)
227                     block->new_cpl_coords[ch] = 1;
228             } else {
229                 for (ch = 1; ch <= s->fbw_channels; ch++) {
230                     if (!block->channel_in_cpl[ch])
231                         continue;
232                     if (!block0->channel_in_cpl[ch]) {
233                         block->new_cpl_coords[ch] = 1;
234                     } else {
235                         CoefSumType coord_diff = 0;
236                         for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
237                             coord_diff += FFABS(cpl_coords[blk-1][ch][bnd] -
238                                                 cpl_coords[blk  ][ch][bnd]);
239                         }
240                         coord_diff /= s->num_cpl_bands;
241                         if (coord_diff > NEW_CPL_COORD_THRESHOLD)
242                             block->new_cpl_coords[ch] = 1;
243                     }
244                 }
245             }
246         }
247     }
248
249     /* calculate final coupling coordinates, taking into account reusing of
250        coordinates in successive blocks */
251     for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
252         blk = 0;
253         while (blk < s->num_blocks) {
254             int av_uninit(blk1);
255             AC3Block *block  = &s->blocks[blk];
256
257             if (!block->cpl_in_use) {
258                 blk++;
259                 continue;
260             }
261
262             for (ch = 1; ch <= s->fbw_channels; ch++) {
263                 CoefSumType energy_ch, energy_cpl;
264                 if (!block->channel_in_cpl[ch])
265                     continue;
266                 energy_cpl = energy[blk][CPL_CH][bnd];
267                 energy_ch = energy[blk][ch][bnd];
268                 blk1 = blk+1;
269                 while (!s->blocks[blk1].new_cpl_coords[ch] && blk1 < s->num_blocks) {
270                     if (s->blocks[blk1].cpl_in_use) {
271                         energy_cpl += energy[blk1][CPL_CH][bnd];
272                         energy_ch += energy[blk1][ch][bnd];
273                     }
274                     blk1++;
275                 }
276                 cpl_coords[blk][ch][bnd] = calc_cpl_coord(energy_ch, energy_cpl);
277             }
278             blk = blk1;
279         }
280     }
281
282     /* calculate exponents/mantissas for coupling coordinates */
283     for (blk = 0; blk < s->num_blocks; blk++) {
284         AC3Block *block = &s->blocks[blk];
285         if (!block->cpl_in_use)
286             continue;
287
288 #if CONFIG_AC3ENC_FLOAT
289         s->ac3dsp.float_to_fixed24(fixed_cpl_coords[blk][1],
290                                    cpl_coords[blk][1],
291                                    s->fbw_channels * 16);
292 #endif
293         s->ac3dsp.extract_exponents(block->cpl_coord_exp[1],
294                                     fixed_cpl_coords[blk][1],
295                                     s->fbw_channels * 16);
296
297         for (ch = 1; ch <= s->fbw_channels; ch++) {
298             int bnd, min_exp, max_exp, master_exp;
299
300             if (!block->new_cpl_coords[ch])
301                 continue;
302
303             /* determine master exponent */
304             min_exp = max_exp = block->cpl_coord_exp[ch][0];
305             for (bnd = 1; bnd < s->num_cpl_bands; bnd++) {
306                 int exp = block->cpl_coord_exp[ch][bnd];
307                 min_exp = FFMIN(exp, min_exp);
308                 max_exp = FFMAX(exp, max_exp);
309             }
310             master_exp = ((max_exp - 15) + 2) / 3;
311             master_exp = FFMAX(master_exp, 0);
312             while (min_exp < master_exp * 3)
313                 master_exp--;
314             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
315                 block->cpl_coord_exp[ch][bnd] = av_clip(block->cpl_coord_exp[ch][bnd] -
316                                                         master_exp * 3, 0, 15);
317             }
318             block->cpl_master_exp[ch] = master_exp;
319
320             /* quantize mantissas */
321             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
322                 int cpl_exp  = block->cpl_coord_exp[ch][bnd];
323                 int cpl_mant = (fixed_cpl_coords[blk][ch][bnd] << (5 + cpl_exp + master_exp * 3)) >> 24;
324                 if (cpl_exp == 15)
325                     cpl_mant >>= 1;
326                 else
327                     cpl_mant -= 16;
328
329                 block->cpl_coord_mant[ch][bnd] = cpl_mant;
330             }
331         }
332     }
333
334     if (CONFIG_EAC3_ENCODER && s->eac3)
335         ff_eac3_set_cpl_states(s);
336 }
337
338
339 /*
340  * Determine rematrixing flags for each block and band.
341  */
342 static void compute_rematrixing_strategy(AC3EncodeContext *s)
343 {
344     int nb_coefs;
345     int blk, bnd;
346     AC3Block *block, *block0;
347
348     if (s->channel_mode != AC3_CHMODE_STEREO)
349         return;
350
351     for (blk = 0; blk < s->num_blocks; blk++) {
352         block = &s->blocks[blk];
353         block->new_rematrixing_strategy = !blk;
354
355         block->num_rematrixing_bands = 4;
356         if (block->cpl_in_use) {
357             block->num_rematrixing_bands -= (s->start_freq[CPL_CH] <= 61);
358             block->num_rematrixing_bands -= (s->start_freq[CPL_CH] == 37);
359             if (blk && block->num_rematrixing_bands != block0->num_rematrixing_bands)
360                 block->new_rematrixing_strategy = 1;
361         }
362         nb_coefs = FFMIN(block->end_freq[1], block->end_freq[2]);
363
364         if (!s->rematrixing_enabled) {
365             block0 = block;
366             continue;
367         }
368
369         for (bnd = 0; bnd < block->num_rematrixing_bands; bnd++) {
370             /* calculate calculate sum of squared coeffs for one band in one block */
371             int start = ff_ac3_rematrix_band_tab[bnd];
372             int end   = FFMIN(nb_coefs, ff_ac3_rematrix_band_tab[bnd+1]);
373             CoefSumType sum[4];
374             sum_square_butterfly(s, sum, block->mdct_coef[1] + start,
375                                  block->mdct_coef[2] + start, end - start);
376
377             /* compare sums to determine if rematrixing will be used for this band */
378             if (FFMIN(sum[2], sum[3]) < FFMIN(sum[0], sum[1]))
379                 block->rematrixing_flags[bnd] = 1;
380             else
381                 block->rematrixing_flags[bnd] = 0;
382
383             /* determine if new rematrixing flags will be sent */
384             if (blk &&
385                 block->rematrixing_flags[bnd] != block0->rematrixing_flags[bnd]) {
386                 block->new_rematrixing_strategy = 1;
387             }
388         }
389         block0 = block;
390     }
391 }
392
393
394 int AC3_NAME(encode_frame)(AVCodecContext *avctx, AVPacket *avpkt,
395                            const AVFrame *frame, int *got_packet_ptr)
396 {
397     AC3EncodeContext *s = avctx->priv_data;
398     const SampleType *samples = (const SampleType *)frame->data[0];
399     int ret;
400
401     if (s->options.allow_per_frame_metadata) {
402         ret = ff_ac3_validate_metadata(s);
403         if (ret)
404             return ret;
405     }
406
407     if (s->bit_alloc.sr_code == 1 || s->eac3)
408         ff_ac3_adjust_frame_size(s);
409
410     deinterleave_input_samples(s, samples);
411
412     apply_mdct(s);
413
414     if (s->fixed_point)
415         scale_coefficients(s);
416
417     clip_coefficients(&s->dsp, s->blocks[0].mdct_coef[1],
418                       AC3_MAX_COEFS * s->num_blocks * s->channels);
419
420     s->cpl_on = s->cpl_enabled;
421     ff_ac3_compute_coupling_strategy(s);
422
423     if (s->cpl_on)
424         apply_channel_coupling(s);
425
426     compute_rematrixing_strategy(s);
427
428     if (!s->fixed_point)
429         scale_coefficients(s);
430
431     ff_ac3_apply_rematrixing(s);
432
433     ff_ac3_process_exponents(s);
434
435     ret = ff_ac3_compute_bit_allocation(s);
436     if (ret) {
437         av_log(avctx, AV_LOG_ERROR, "Bit allocation failed. Try increasing the bitrate.\n");
438         return ret;
439     }
440
441     ff_ac3_group_exponents(s);
442
443     ff_ac3_quantize_mantissas(s);
444
445     if ((ret = ff_alloc_packet2(avctx, avpkt, s->frame_size)))
446         return ret;
447     ff_ac3_output_frame(s, avpkt->data);
448
449     if (frame->pts != AV_NOPTS_VALUE)
450         avpkt->pts = frame->pts - ff_samples_to_time_base(avctx, avctx->delay);
451
452     *got_packet_ptr = 1;
453     return 0;
454 }