Imported Upstream version 6.1
[platform/upstream/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 "config_components.h"
30
31 #include <stdint.h>
32
33 #include "libavutil/attributes.h"
34 #include "libavutil/internal.h"
35 #include "libavutil/mem_internal.h"
36
37 #include "audiodsp.h"
38 #include "ac3enc.h"
39 #include "eac3enc.h"
40
41
42 static int allocate_sample_buffers(AC3EncodeContext *s)
43 {
44     int ch;
45
46     if (!FF_ALLOC_TYPED_ARRAY(s->windowed_samples, AC3_WINDOW_SIZE) ||
47         !FF_ALLOCZ_TYPED_ARRAY(s->planar_samples,  s->channels))
48         return AVERROR(ENOMEM);
49
50     for (ch = 0; ch < s->channels; ch++) {
51         if (!(s->planar_samples[ch] = av_mallocz((AC3_FRAME_SIZE + AC3_BLOCK_SIZE) *
52                                                   sizeof(**s->planar_samples))))
53             return AVERROR(ENOMEM);
54     }
55     return 0;
56 }
57
58
59 /*
60  * Copy input samples.
61  * Channels are reordered from FFmpeg's default order to AC-3 order.
62  */
63 static void copy_input_samples(AC3EncodeContext *s, SampleType **samples)
64 {
65     int ch;
66
67     /* copy and remap input samples */
68     for (ch = 0; ch < s->channels; ch++) {
69         /* copy last 256 samples of previous frame to the start of the current frame */
70         memcpy(&s->planar_samples[ch][0], &s->planar_samples[ch][AC3_BLOCK_SIZE * s->num_blocks],
71                AC3_BLOCK_SIZE * sizeof(s->planar_samples[0][0]));
72
73         /* copy new samples for current frame */
74         memcpy(&s->planar_samples[ch][AC3_BLOCK_SIZE],
75                samples[s->channel_map[ch]],
76                AC3_BLOCK_SIZE * s->num_blocks * sizeof(s->planar_samples[0][0]));
77     }
78 }
79
80
81 /*
82  * Apply the MDCT to input samples to generate frequency coefficients.
83  * This applies the KBD window and normalizes the input to reduce precision
84  * loss due to fixed-point calculations.
85  */
86 static void apply_mdct(AC3EncodeContext *s)
87 {
88     int blk, ch;
89
90     for (ch = 0; ch < s->channels; ch++) {
91         for (blk = 0; blk < s->num_blocks; blk++) {
92             AC3Block *block = &s->blocks[blk];
93             const SampleType *input_samples = &s->planar_samples[ch][blk * AC3_BLOCK_SIZE];
94
95             s->fdsp->vector_fmul(s->windowed_samples, input_samples,
96                                  s->mdct_window, AC3_BLOCK_SIZE);
97             s->fdsp->vector_fmul_reverse(s->windowed_samples + AC3_BLOCK_SIZE,
98                                          &input_samples[AC3_BLOCK_SIZE],
99                                          s->mdct_window, AC3_BLOCK_SIZE);
100
101             s->tx_fn(s->tx, block->mdct_coef[ch+1],
102                      s->windowed_samples, sizeof(float));
103         }
104     }
105 }
106
107
108 /*
109  * Calculate coupling channel and coupling coordinates.
110  */
111 static void apply_channel_coupling(AC3EncodeContext *s)
112 {
113     LOCAL_ALIGNED_16(CoefType, cpl_coords,      [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
114 #if AC3ENC_FLOAT
115     LOCAL_ALIGNED_16(int32_t, fixed_cpl_coords, [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
116 #else
117     int32_t (*fixed_cpl_coords)[AC3_MAX_CHANNELS][16] = cpl_coords;
118 #endif
119     int av_uninit(blk), ch, bnd, i, j;
120     CoefSumType energy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][16] = {{{0}}};
121     int cpl_start, num_cpl_coefs;
122
123     memset(cpl_coords,       0, AC3_MAX_BLOCKS * sizeof(*cpl_coords));
124 #if AC3ENC_FLOAT
125     memset(fixed_cpl_coords, 0, AC3_MAX_BLOCKS * sizeof(*cpl_coords));
126 #endif
127
128     /* align start to 16-byte boundary. align length to multiple of 32.
129         note: coupling start bin % 4 will always be 1 */
130     cpl_start     = s->start_freq[CPL_CH] - 1;
131     num_cpl_coefs = FFALIGN(s->num_cpl_subbands * 12 + 1, 32);
132     cpl_start     = FFMIN(256, cpl_start + num_cpl_coefs) - num_cpl_coefs;
133
134     /* calculate coupling channel from fbw channels */
135     for (blk = 0; blk < s->num_blocks; blk++) {
136         AC3Block *block = &s->blocks[blk];
137         CoefType *cpl_coef = &block->mdct_coef[CPL_CH][cpl_start];
138         if (!block->cpl_in_use)
139             continue;
140         memset(cpl_coef, 0, num_cpl_coefs * sizeof(*cpl_coef));
141         for (ch = 1; ch <= s->fbw_channels; ch++) {
142             CoefType *ch_coef = &block->mdct_coef[ch][cpl_start];
143             if (!block->channel_in_cpl[ch])
144                 continue;
145             for (i = 0; i < num_cpl_coefs; i++)
146                 cpl_coef[i] += ch_coef[i];
147         }
148
149         /* coefficients must be clipped in order to be encoded */
150         clip_coefficients(&s->adsp, cpl_coef, num_cpl_coefs);
151     }
152
153     /* calculate energy in each band in coupling channel and each fbw channel */
154     /* TODO: possibly use SIMD to speed up energy calculation */
155     bnd = 0;
156     i = s->start_freq[CPL_CH];
157     while (i < s->cpl_end_freq) {
158         int band_size = s->cpl_band_sizes[bnd];
159         for (ch = CPL_CH; ch <= s->fbw_channels; ch++) {
160             for (blk = 0; blk < s->num_blocks; blk++) {
161                 AC3Block *block = &s->blocks[blk];
162                 if (!block->cpl_in_use || (ch > CPL_CH && !block->channel_in_cpl[ch]))
163                     continue;
164                 for (j = 0; j < band_size; j++) {
165                     CoefType v = block->mdct_coef[ch][i+j];
166                     MAC_COEF(energy[blk][ch][bnd], v, v);
167                 }
168             }
169         }
170         i += band_size;
171         bnd++;
172     }
173
174     /* calculate coupling coordinates for all blocks for all channels */
175     for (blk = 0; blk < s->num_blocks; blk++) {
176         AC3Block *block  = &s->blocks[blk];
177         if (!block->cpl_in_use)
178             continue;
179         for (ch = 1; ch <= s->fbw_channels; ch++) {
180             if (!block->channel_in_cpl[ch])
181                 continue;
182             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
183                 cpl_coords[blk][ch][bnd] = calc_cpl_coord(energy[blk][ch][bnd],
184                                                           energy[blk][CPL_CH][bnd]);
185             }
186         }
187     }
188
189     /* determine which blocks to send new coupling coordinates for */
190     for (blk = 0; blk < s->num_blocks; blk++) {
191         AC3Block *block  = &s->blocks[blk];
192         AC3Block *block0 = blk ? &s->blocks[blk-1] : NULL;
193
194         memset(block->new_cpl_coords, 0, sizeof(block->new_cpl_coords));
195
196         if (block->cpl_in_use) {
197             /* send new coordinates if this is the first block, if previous
198              * block did not use coupling but this block does, the channels
199              * using coupling has changed from the previous block, or the
200              * coordinate difference from the last block for any channel is
201              * greater than a threshold value. */
202             if (blk == 0 || !block0->cpl_in_use) {
203                 for (ch = 1; ch <= s->fbw_channels; ch++)
204                     block->new_cpl_coords[ch] = 1;
205             } else {
206                 for (ch = 1; ch <= s->fbw_channels; ch++) {
207                     if (!block->channel_in_cpl[ch])
208                         continue;
209                     if (!block0->channel_in_cpl[ch]) {
210                         block->new_cpl_coords[ch] = 1;
211                     } else {
212                         CoefSumType coord_diff = 0;
213                         for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
214                             coord_diff += FFABS(cpl_coords[blk-1][ch][bnd] -
215                                                 cpl_coords[blk  ][ch][bnd]);
216                         }
217                         coord_diff /= s->num_cpl_bands;
218                         if (coord_diff > NEW_CPL_COORD_THRESHOLD)
219                             block->new_cpl_coords[ch] = 1;
220                     }
221                 }
222             }
223         }
224     }
225
226     /* calculate final coupling coordinates, taking into account reusing of
227        coordinates in successive blocks */
228     for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
229         blk = 0;
230         while (blk < s->num_blocks) {
231             int av_uninit(blk1);
232             AC3Block *block  = &s->blocks[blk];
233
234             if (!block->cpl_in_use) {
235                 blk++;
236                 continue;
237             }
238
239             for (ch = 1; ch <= s->fbw_channels; ch++) {
240                 CoefSumType energy_ch, energy_cpl;
241                 if (!block->channel_in_cpl[ch])
242                     continue;
243                 energy_cpl = energy[blk][CPL_CH][bnd];
244                 energy_ch = energy[blk][ch][bnd];
245                 blk1 = blk+1;
246                 while (blk1 < s->num_blocks && !s->blocks[blk1].new_cpl_coords[ch]) {
247                     if (s->blocks[blk1].cpl_in_use) {
248                         energy_cpl += energy[blk1][CPL_CH][bnd];
249                         energy_ch += energy[blk1][ch][bnd];
250                     }
251                     blk1++;
252                 }
253                 cpl_coords[blk][ch][bnd] = calc_cpl_coord(energy_ch, energy_cpl);
254             }
255             blk = blk1;
256         }
257     }
258
259     /* calculate exponents/mantissas for coupling coordinates */
260     for (blk = 0; blk < s->num_blocks; blk++) {
261         AC3Block *block = &s->blocks[blk];
262         if (!block->cpl_in_use)
263             continue;
264
265 #if AC3ENC_FLOAT
266         s->ac3dsp.float_to_fixed24(fixed_cpl_coords[blk][1],
267                                    cpl_coords[blk][1],
268                                    s->fbw_channels * 16);
269 #endif
270         s->ac3dsp.extract_exponents(block->cpl_coord_exp[1],
271                                     fixed_cpl_coords[blk][1],
272                                     s->fbw_channels * 16);
273
274         for (ch = 1; ch <= s->fbw_channels; ch++) {
275             int bnd, min_exp, max_exp, master_exp;
276
277             if (!block->new_cpl_coords[ch])
278                 continue;
279
280             /* determine master exponent */
281             min_exp = max_exp = block->cpl_coord_exp[ch][0];
282             for (bnd = 1; bnd < s->num_cpl_bands; bnd++) {
283                 int exp = block->cpl_coord_exp[ch][bnd];
284                 min_exp = FFMIN(exp, min_exp);
285                 max_exp = FFMAX(exp, max_exp);
286             }
287             master_exp = ((max_exp - 15) + 2) / 3;
288             master_exp = FFMAX(master_exp, 0);
289             while (min_exp < master_exp * 3)
290                 master_exp--;
291             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
292                 block->cpl_coord_exp[ch][bnd] = av_clip(block->cpl_coord_exp[ch][bnd] -
293                                                         master_exp * 3, 0, 15);
294             }
295             block->cpl_master_exp[ch] = master_exp;
296
297             /* quantize mantissas */
298             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
299                 int cpl_exp  = block->cpl_coord_exp[ch][bnd];
300                 int cpl_mant = (fixed_cpl_coords[blk][ch][bnd] << (5 + cpl_exp + master_exp * 3)) >> 24;
301                 if (cpl_exp == 15)
302                     cpl_mant >>= 1;
303                 else
304                     cpl_mant -= 16;
305
306                 block->cpl_coord_mant[ch][bnd] = cpl_mant;
307             }
308         }
309     }
310
311     if (AC3ENC_FLOAT && CONFIG_EAC3_ENCODER && s->eac3)
312         ff_eac3_set_cpl_states(s);
313 }
314
315
316 /*
317  * Determine rematrixing flags for each block and band.
318  */
319 static void compute_rematrixing_strategy(AC3EncodeContext *s)
320 {
321     int nb_coefs;
322     int blk, bnd;
323     AC3Block *block, *block0 = NULL;
324
325     if (s->channel_mode != AC3_CHMODE_STEREO)
326         return;
327
328     for (blk = 0; blk < s->num_blocks; blk++) {
329         block = &s->blocks[blk];
330         block->new_rematrixing_strategy = !blk;
331
332         block->num_rematrixing_bands = 4;
333         if (block->cpl_in_use) {
334             block->num_rematrixing_bands -= (s->start_freq[CPL_CH] <= 61);
335             block->num_rematrixing_bands -= (s->start_freq[CPL_CH] == 37);
336             if (blk && block->num_rematrixing_bands != block0->num_rematrixing_bands)
337                 block->new_rematrixing_strategy = 1;
338         }
339         nb_coefs = FFMIN(block->end_freq[1], block->end_freq[2]);
340
341         if (!s->rematrixing_enabled) {
342             block0 = block;
343             continue;
344         }
345
346         for (bnd = 0; bnd < block->num_rematrixing_bands; bnd++) {
347             /* calculate sum of squared coeffs for one band in one block */
348             int start = ff_ac3_rematrix_band_tab[bnd];
349             int end   = FFMIN(nb_coefs, ff_ac3_rematrix_band_tab[bnd+1]);
350             CoefSumType sum[4];
351             sum_square_butterfly(s, sum, block->mdct_coef[1] + start,
352                                  block->mdct_coef[2] + start, end - start);
353
354             /* compare sums to determine if rematrixing will be used for this band */
355             if (FFMIN(sum[2], sum[3]) < FFMIN(sum[0], sum[1]))
356                 block->rematrixing_flags[bnd] = 1;
357             else
358                 block->rematrixing_flags[bnd] = 0;
359
360             /* determine if new rematrixing flags will be sent */
361             if (blk &&
362                 block->rematrixing_flags[bnd] != block0->rematrixing_flags[bnd]) {
363                 block->new_rematrixing_strategy = 1;
364             }
365         }
366         block0 = block;
367     }
368 }
369
370
371 int AC3_NAME(encode_frame)(AVCodecContext *avctx, AVPacket *avpkt,
372                            const AVFrame *frame, int *got_packet_ptr)
373 {
374     AC3EncodeContext *s = avctx->priv_data;
375     int ret;
376
377     if (s->options.allow_per_frame_metadata) {
378         ret = ff_ac3_validate_metadata(s);
379         if (ret)
380             return ret;
381     }
382
383     if (s->bit_alloc.sr_code == 1 || (AC3ENC_FLOAT && s->eac3))
384         ff_ac3_adjust_frame_size(s);
385
386     copy_input_samples(s, (SampleType **)frame->extended_data);
387
388     apply_mdct(s);
389
390     s->cpl_on = s->cpl_enabled;
391     ff_ac3_compute_coupling_strategy(s);
392
393     if (s->cpl_on)
394         apply_channel_coupling(s);
395
396     compute_rematrixing_strategy(s);
397
398 #if AC3ENC_FLOAT
399     scale_coefficients(s);
400 #endif
401
402     return ff_ac3_encode_frame_common_end(avctx, avpkt, frame, got_packet_ptr);
403 }