normalize calls to ff_alloc_packet2
[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 FFmpeg.
8  *
9  * FFmpeg 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  * FFmpeg 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 FFmpeg; 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 void apply_window(void *dsp, SampleType *output,
38                          const SampleType *input, const SampleType *window,
39                          unsigned int len);
40
41 static int normalize_samples(AC3EncodeContext *s);
42
43 static void clip_coefficients(DSPContext *dsp, CoefType *coef, unsigned int len);
44
45 static CoefType calc_cpl_coord(CoefSumType energy_ch, CoefSumType energy_cpl);
46
47 static void sum_square_butterfly(AC3EncodeContext *s, CoefSumType sum[4],
48                                  const CoefType *coef0, const CoefType *coef1,
49                                  int len);
50
51 int AC3_NAME(allocate_sample_buffers)(AC3EncodeContext *s)
52 {
53     int ch;
54
55     FF_ALLOC_OR_GOTO(s->avctx, s->windowed_samples, AC3_WINDOW_SIZE *
56                      sizeof(*s->windowed_samples), alloc_fail);
57     FF_ALLOC_OR_GOTO(s->avctx, s->planar_samples, s->channels * sizeof(*s->planar_samples),
58                      alloc_fail);
59     for (ch = 0; ch < s->channels; ch++) {
60         FF_ALLOCZ_OR_GOTO(s->avctx, s->planar_samples[ch],
61                           (AC3_FRAME_SIZE+AC3_BLOCK_SIZE) * sizeof(**s->planar_samples),
62                           alloc_fail);
63     }
64
65     return 0;
66 alloc_fail:
67     return AVERROR(ENOMEM);
68 }
69
70
71 /*
72  * Copy input samples.
73  * Channels are reordered from FFmpeg's default order to AC-3 order.
74  */
75 static void copy_input_samples(AC3EncodeContext *s, SampleType **samples)
76 {
77     int ch;
78
79     /* copy and remap input samples */
80     for (ch = 0; ch < s->channels; ch++) {
81         /* copy last 256 samples of previous frame to the start of the current frame */
82         memcpy(&s->planar_samples[ch][0], &s->planar_samples[ch][AC3_BLOCK_SIZE * s->num_blocks],
83                AC3_BLOCK_SIZE * sizeof(s->planar_samples[0][0]));
84
85         /* copy new samples for current frame */
86         memcpy(&s->planar_samples[ch][AC3_BLOCK_SIZE],
87                samples[s->channel_map[ch]],
88                AC3_BLOCK_SIZE * s->num_blocks * sizeof(s->planar_samples[0][0]));
89     }
90 }
91
92
93 /*
94  * Apply the MDCT to input samples to generate frequency coefficients.
95  * This applies the KBD window and normalizes the input to reduce precision
96  * loss due to fixed-point calculations.
97  */
98 static void apply_mdct(AC3EncodeContext *s)
99 {
100     int blk, ch;
101
102     for (ch = 0; ch < s->channels; ch++) {
103         for (blk = 0; blk < s->num_blocks; blk++) {
104             AC3Block *block = &s->blocks[blk];
105             const SampleType *input_samples = &s->planar_samples[ch][blk * AC3_BLOCK_SIZE];
106
107 #if CONFIG_AC3ENC_FLOAT
108             apply_window(&s->fdsp, s->windowed_samples, input_samples,
109                          s->mdct_window, AC3_WINDOW_SIZE);
110 #else
111             apply_window(&s->dsp, s->windowed_samples, input_samples,
112                          s->mdct_window, AC3_WINDOW_SIZE);
113 #endif
114
115             if (s->fixed_point)
116                 block->coeff_shift[ch+1] = normalize_samples(s);
117
118             s->mdct.mdct_calcw(&s->mdct, block->mdct_coef[ch+1],
119                                s->windowed_samples);
120         }
121     }
122 }
123
124
125 /*
126  * Calculate coupling channel and coupling coordinates.
127  */
128 static void apply_channel_coupling(AC3EncodeContext *s)
129 {
130     LOCAL_ALIGNED_16(CoefType, cpl_coords,      [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
131 #if CONFIG_AC3ENC_FLOAT
132     LOCAL_ALIGNED_16(int32_t, fixed_cpl_coords, [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
133 #else
134     int32_t (*fixed_cpl_coords)[AC3_MAX_CHANNELS][16] = cpl_coords;
135 #endif
136     int av_uninit(blk), ch, bnd, i, j;
137     CoefSumType energy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][16] = {{{0}}};
138     int cpl_start, num_cpl_coefs;
139
140     memset(cpl_coords,       0, AC3_MAX_BLOCKS * sizeof(*cpl_coords));
141 #if CONFIG_AC3ENC_FLOAT
142     memset(fixed_cpl_coords, 0, AC3_MAX_BLOCKS * sizeof(*cpl_coords));
143 #endif
144
145     /* align start to 16-byte boundary. align length to multiple of 32.
146         note: coupling start bin % 4 will always be 1 */
147     cpl_start     = s->start_freq[CPL_CH] - 1;
148     num_cpl_coefs = FFALIGN(s->num_cpl_subbands * 12 + 1, 32);
149     cpl_start     = FFMIN(256, cpl_start + num_cpl_coefs) - num_cpl_coefs;
150
151     /* calculate coupling channel from fbw channels */
152     for (blk = 0; blk < s->num_blocks; blk++) {
153         AC3Block *block = &s->blocks[blk];
154         CoefType *cpl_coef = &block->mdct_coef[CPL_CH][cpl_start];
155         if (!block->cpl_in_use)
156             continue;
157         memset(cpl_coef, 0, num_cpl_coefs * sizeof(*cpl_coef));
158         for (ch = 1; ch <= s->fbw_channels; ch++) {
159             CoefType *ch_coef = &block->mdct_coef[ch][cpl_start];
160             if (!block->channel_in_cpl[ch])
161                 continue;
162             for (i = 0; i < num_cpl_coefs; i++)
163                 cpl_coef[i] += ch_coef[i];
164         }
165
166         /* coefficients must be clipped in order to be encoded */
167         clip_coefficients(&s->dsp, cpl_coef, num_cpl_coefs);
168     }
169
170     /* calculate energy in each band in coupling channel and each fbw channel */
171     /* TODO: possibly use SIMD to speed up energy calculation */
172     bnd = 0;
173     i = s->start_freq[CPL_CH];
174     while (i < s->cpl_end_freq) {
175         int band_size = s->cpl_band_sizes[bnd];
176         for (ch = CPL_CH; ch <= s->fbw_channels; ch++) {
177             for (blk = 0; blk < s->num_blocks; blk++) {
178                 AC3Block *block = &s->blocks[blk];
179                 if (!block->cpl_in_use || (ch > CPL_CH && !block->channel_in_cpl[ch]))
180                     continue;
181                 for (j = 0; j < band_size; j++) {
182                     CoefType v = block->mdct_coef[ch][i+j];
183                     MAC_COEF(energy[blk][ch][bnd], v, v);
184                 }
185             }
186         }
187         i += band_size;
188         bnd++;
189     }
190
191     /* calculate coupling coordinates for all blocks for all channels */
192     for (blk = 0; blk < s->num_blocks; blk++) {
193         AC3Block *block  = &s->blocks[blk];
194         if (!block->cpl_in_use)
195             continue;
196         for (ch = 1; ch <= s->fbw_channels; ch++) {
197             if (!block->channel_in_cpl[ch])
198                 continue;
199             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
200                 cpl_coords[blk][ch][bnd] = calc_cpl_coord(energy[blk][ch][bnd],
201                                                           energy[blk][CPL_CH][bnd]);
202             }
203         }
204     }
205
206     /* determine which blocks to send new coupling coordinates for */
207     for (blk = 0; blk < s->num_blocks; blk++) {
208         AC3Block *block  = &s->blocks[blk];
209         AC3Block *block0 = blk ? &s->blocks[blk-1] : NULL;
210
211         memset(block->new_cpl_coords, 0, sizeof(block->new_cpl_coords));
212
213         if (block->cpl_in_use) {
214             /* send new coordinates if this is the first block, if previous
215              * block did not use coupling but this block does, the channels
216              * using coupling has changed from the previous block, or the
217              * coordinate difference from the last block for any channel is
218              * greater than a threshold value. */
219             if (blk == 0 || !block0->cpl_in_use) {
220                 for (ch = 1; ch <= s->fbw_channels; ch++)
221                     block->new_cpl_coords[ch] = 1;
222             } else {
223                 for (ch = 1; ch <= s->fbw_channels; ch++) {
224                     if (!block->channel_in_cpl[ch])
225                         continue;
226                     if (!block0->channel_in_cpl[ch]) {
227                         block->new_cpl_coords[ch] = 1;
228                     } else {
229                         CoefSumType coord_diff = 0;
230                         for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
231                             coord_diff += FFABS(cpl_coords[blk-1][ch][bnd] -
232                                                 cpl_coords[blk  ][ch][bnd]);
233                         }
234                         coord_diff /= s->num_cpl_bands;
235                         if (coord_diff > NEW_CPL_COORD_THRESHOLD)
236                             block->new_cpl_coords[ch] = 1;
237                     }
238                 }
239             }
240         }
241     }
242
243     /* calculate final coupling coordinates, taking into account reusing of
244        coordinates in successive blocks */
245     for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
246         blk = 0;
247         while (blk < s->num_blocks) {
248             int av_uninit(blk1);
249             AC3Block *block  = &s->blocks[blk];
250
251             if (!block->cpl_in_use) {
252                 blk++;
253                 continue;
254             }
255
256             for (ch = 1; ch <= s->fbw_channels; ch++) {
257                 CoefSumType energy_ch, energy_cpl;
258                 if (!block->channel_in_cpl[ch])
259                     continue;
260                 energy_cpl = energy[blk][CPL_CH][bnd];
261                 energy_ch = energy[blk][ch][bnd];
262                 blk1 = blk+1;
263                 while (!s->blocks[blk1].new_cpl_coords[ch] && blk1 < s->num_blocks) {
264                     if (s->blocks[blk1].cpl_in_use) {
265                         energy_cpl += energy[blk1][CPL_CH][bnd];
266                         energy_ch += energy[blk1][ch][bnd];
267                     }
268                     blk1++;
269                 }
270                 cpl_coords[blk][ch][bnd] = calc_cpl_coord(energy_ch, energy_cpl);
271             }
272             blk = blk1;
273         }
274     }
275
276     /* calculate exponents/mantissas for coupling coordinates */
277     for (blk = 0; blk < s->num_blocks; blk++) {
278         AC3Block *block = &s->blocks[blk];
279         if (!block->cpl_in_use)
280             continue;
281
282 #if CONFIG_AC3ENC_FLOAT
283         s->ac3dsp.float_to_fixed24(fixed_cpl_coords[blk][1],
284                                    cpl_coords[blk][1],
285                                    s->fbw_channels * 16);
286 #endif
287         s->ac3dsp.extract_exponents(block->cpl_coord_exp[1],
288                                     fixed_cpl_coords[blk][1],
289                                     s->fbw_channels * 16);
290
291         for (ch = 1; ch <= s->fbw_channels; ch++) {
292             int bnd, min_exp, max_exp, master_exp;
293
294             if (!block->new_cpl_coords[ch])
295                 continue;
296
297             /* determine master exponent */
298             min_exp = max_exp = block->cpl_coord_exp[ch][0];
299             for (bnd = 1; bnd < s->num_cpl_bands; bnd++) {
300                 int exp = block->cpl_coord_exp[ch][bnd];
301                 min_exp = FFMIN(exp, min_exp);
302                 max_exp = FFMAX(exp, max_exp);
303             }
304             master_exp = ((max_exp - 15) + 2) / 3;
305             master_exp = FFMAX(master_exp, 0);
306             while (min_exp < master_exp * 3)
307                 master_exp--;
308             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
309                 block->cpl_coord_exp[ch][bnd] = av_clip(block->cpl_coord_exp[ch][bnd] -
310                                                         master_exp * 3, 0, 15);
311             }
312             block->cpl_master_exp[ch] = master_exp;
313
314             /* quantize mantissas */
315             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
316                 int cpl_exp  = block->cpl_coord_exp[ch][bnd];
317                 int cpl_mant = (fixed_cpl_coords[blk][ch][bnd] << (5 + cpl_exp + master_exp * 3)) >> 24;
318                 if (cpl_exp == 15)
319                     cpl_mant >>= 1;
320                 else
321                     cpl_mant -= 16;
322
323                 block->cpl_coord_mant[ch][bnd] = cpl_mant;
324             }
325         }
326     }
327
328     if (CONFIG_EAC3_ENCODER && s->eac3)
329         ff_eac3_set_cpl_states(s);
330 }
331
332
333 /*
334  * Determine rematrixing flags for each block and band.
335  */
336 static void compute_rematrixing_strategy(AC3EncodeContext *s)
337 {
338     int nb_coefs;
339     int blk, bnd;
340     AC3Block *block, *block0 = NULL;
341
342     if (s->channel_mode != AC3_CHMODE_STEREO)
343         return;
344
345     for (blk = 0; blk < s->num_blocks; blk++) {
346         block = &s->blocks[blk];
347         block->new_rematrixing_strategy = !blk;
348
349         block->num_rematrixing_bands = 4;
350         if (block->cpl_in_use) {
351             block->num_rematrixing_bands -= (s->start_freq[CPL_CH] <= 61);
352             block->num_rematrixing_bands -= (s->start_freq[CPL_CH] == 37);
353             if (blk && block->num_rematrixing_bands != block0->num_rematrixing_bands)
354                 block->new_rematrixing_strategy = 1;
355         }
356         nb_coefs = FFMIN(block->end_freq[1], block->end_freq[2]);
357
358         if (!s->rematrixing_enabled) {
359             block0 = block;
360             continue;
361         }
362
363         for (bnd = 0; bnd < block->num_rematrixing_bands; bnd++) {
364             /* calculate calculate sum of squared coeffs for one band in one block */
365             int start = ff_ac3_rematrix_band_tab[bnd];
366             int end   = FFMIN(nb_coefs, ff_ac3_rematrix_band_tab[bnd+1]);
367             CoefSumType sum[4];
368             sum_square_butterfly(s, sum, block->mdct_coef[1] + start,
369                                  block->mdct_coef[2] + start, end - start);
370
371             /* compare sums to determine if rematrixing will be used for this band */
372             if (FFMIN(sum[2], sum[3]) < FFMIN(sum[0], sum[1]))
373                 block->rematrixing_flags[bnd] = 1;
374             else
375                 block->rematrixing_flags[bnd] = 0;
376
377             /* determine if new rematrixing flags will be sent */
378             if (blk &&
379                 block->rematrixing_flags[bnd] != block0->rematrixing_flags[bnd]) {
380                 block->new_rematrixing_strategy = 1;
381             }
382         }
383         block0 = block;
384     }
385 }
386
387
388 int AC3_NAME(encode_frame)(AVCodecContext *avctx, AVPacket *avpkt,
389                            const AVFrame *frame, int *got_packet_ptr)
390 {
391     AC3EncodeContext *s = avctx->priv_data;
392     int ret;
393
394     if (s->options.allow_per_frame_metadata) {
395         ret = ff_ac3_validate_metadata(s);
396         if (ret)
397             return ret;
398     }
399
400     if (s->bit_alloc.sr_code == 1 || s->eac3)
401         ff_ac3_adjust_frame_size(s);
402
403     copy_input_samples(s, (SampleType **)frame->extended_data);
404
405     apply_mdct(s);
406
407     if (s->fixed_point)
408         scale_coefficients(s);
409
410     clip_coefficients(&s->dsp, s->blocks[0].mdct_coef[1],
411                       AC3_MAX_COEFS * s->num_blocks * s->channels);
412
413     s->cpl_on = s->cpl_enabled;
414     ff_ac3_compute_coupling_strategy(s);
415
416     if (s->cpl_on)
417         apply_channel_coupling(s);
418
419     compute_rematrixing_strategy(s);
420
421     if (!s->fixed_point)
422         scale_coefficients(s);
423
424     ff_ac3_apply_rematrixing(s);
425
426     ff_ac3_process_exponents(s);
427
428     ret = ff_ac3_compute_bit_allocation(s);
429     if (ret) {
430         av_log(avctx, AV_LOG_ERROR, "Bit allocation failed. Try increasing the bitrate.\n");
431         return ret;
432     }
433
434     ff_ac3_group_exponents(s);
435
436     ff_ac3_quantize_mantissas(s);
437
438     if ((ret = ff_alloc_packet2(avctx, avpkt, s->frame_size)) < 0)
439         return ret;
440     ff_ac3_output_frame(s, avpkt->data);
441
442     if (frame->pts != AV_NOPTS_VALUE)
443         avpkt->pts = frame->pts - ff_samples_to_time_base(avctx, avctx->delay);
444
445     *got_packet_ptr = 1;
446     return 0;
447 }