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>
7 * This file is part of Libav.
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.
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.
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
26 * AC-3 encoder float/fixed template
32 /* prototypes for static functions in ac3enc_fixed.c and ac3enc_float.c */
34 static void scale_coefficients(AC3EncodeContext *s);
36 static void apply_window(DSPContext *dsp, SampleType *output,
37 const SampleType *input, const SampleType *window,
40 static int normalize_samples(AC3EncodeContext *s);
42 static void clip_coefficients(DSPContext *dsp, CoefType *coef, unsigned int len);
45 int AC3_NAME(allocate_sample_buffers)(AC3EncodeContext *s)
49 FF_ALLOC_OR_GOTO(s->avctx, s->windowed_samples, AC3_WINDOW_SIZE *
50 sizeof(*s->windowed_samples), alloc_fail);
51 FF_ALLOC_OR_GOTO(s->avctx, s->planar_samples, s->channels * sizeof(*s->planar_samples),
53 for (ch = 0; ch < s->channels; ch++) {
54 FF_ALLOCZ_OR_GOTO(s->avctx, s->planar_samples[ch],
55 (AC3_FRAME_SIZE+AC3_BLOCK_SIZE) * sizeof(**s->planar_samples),
61 return AVERROR(ENOMEM);
66 * Deinterleave input samples.
67 * Channels are reordered from Libav's default order to AC-3 order.
69 static void deinterleave_input_samples(AC3EncodeContext *s,
70 const SampleType *samples)
74 /* deinterleave and remap input samples */
75 for (ch = 0; ch < s->channels; ch++) {
76 const SampleType *sptr;
79 /* copy last 256 samples of previous frame to the start of the current frame */
80 memcpy(&s->planar_samples[ch][0], &s->planar_samples[ch][AC3_BLOCK_SIZE * s->num_blocks],
81 AC3_BLOCK_SIZE * sizeof(s->planar_samples[0][0]));
85 sptr = samples + s->channel_map[ch];
86 for (i = AC3_BLOCK_SIZE; i < AC3_BLOCK_SIZE * (s->num_blocks + 1); i++) {
87 s->planar_samples[ch][i] = *sptr;
95 * Apply the MDCT to input samples to generate frequency coefficients.
96 * This applies the KBD window and normalizes the input to reduce precision
97 * loss due to fixed-point calculations.
99 static void apply_mdct(AC3EncodeContext *s)
103 for (ch = 0; ch < s->channels; ch++) {
104 for (blk = 0; blk < s->num_blocks; blk++) {
105 AC3Block *block = &s->blocks[blk];
106 const SampleType *input_samples = &s->planar_samples[ch][blk * AC3_BLOCK_SIZE];
108 apply_window(&s->dsp, s->windowed_samples, input_samples,
109 s->mdct_window, AC3_WINDOW_SIZE);
112 block->coeff_shift[ch+1] = normalize_samples(s);
114 s->mdct.mdct_calcw(&s->mdct, block->mdct_coef[ch+1],
115 s->windowed_samples);
122 * Calculate a single coupling coordinate.
124 static inline float calc_cpl_coord(float energy_ch, float energy_cpl)
128 coord *= sqrtf(energy_ch / energy_cpl);
129 return FFMIN(coord, COEF_MAX);
134 * Calculate coupling channel and coupling coordinates.
136 static void apply_channel_coupling(AC3EncodeContext *s)
138 #if CONFIG_AC3ENC_FLOAT
139 LOCAL_ALIGNED_16(float, cpl_coords, [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
140 LOCAL_ALIGNED_16(int32_t, fixed_cpl_coords, [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
141 int blk, ch, bnd, i, j;
142 CoefSumType energy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][16] = {{{0}}};
143 int cpl_start, num_cpl_coefs;
145 memset(cpl_coords, 0, AC3_MAX_BLOCKS * sizeof(*cpl_coords));
146 memset(fixed_cpl_coords, 0, AC3_MAX_BLOCKS * sizeof(*fixed_cpl_coords));
148 /* align start to 16-byte boundary. align length to multiple of 32.
149 note: coupling start bin % 4 will always be 1 */
150 cpl_start = s->start_freq[CPL_CH] - 1;
151 num_cpl_coefs = FFALIGN(s->num_cpl_subbands * 12 + 1, 32);
152 cpl_start = FFMIN(256, cpl_start + num_cpl_coefs) - num_cpl_coefs;
154 /* calculate coupling channel from fbw channels */
155 for (blk = 0; blk < s->num_blocks; blk++) {
156 AC3Block *block = &s->blocks[blk];
157 CoefType *cpl_coef = &block->mdct_coef[CPL_CH][cpl_start];
158 if (!block->cpl_in_use)
160 memset(cpl_coef, 0, num_cpl_coefs * sizeof(*cpl_coef));
161 for (ch = 1; ch <= s->fbw_channels; ch++) {
162 CoefType *ch_coef = &block->mdct_coef[ch][cpl_start];
163 if (!block->channel_in_cpl[ch])
165 for (i = 0; i < num_cpl_coefs; i++)
166 cpl_coef[i] += ch_coef[i];
169 /* coefficients must be clipped in order to be encoded */
170 clip_coefficients(&s->dsp, cpl_coef, num_cpl_coefs);
172 /* scale coupling coefficients from float to 24-bit fixed-point */
173 s->ac3dsp.float_to_fixed24(&block->fixed_coef[CPL_CH][cpl_start],
174 cpl_coef, num_cpl_coefs);
177 /* calculate energy in each band in coupling channel and each fbw channel */
178 /* TODO: possibly use SIMD to speed up energy calculation */
180 i = s->start_freq[CPL_CH];
181 while (i < s->cpl_end_freq) {
182 int band_size = s->cpl_band_sizes[bnd];
183 for (ch = CPL_CH; ch <= s->fbw_channels; ch++) {
184 for (blk = 0; blk < s->num_blocks; blk++) {
185 AC3Block *block = &s->blocks[blk];
186 if (!block->cpl_in_use || (ch > CPL_CH && !block->channel_in_cpl[ch]))
188 for (j = 0; j < band_size; j++) {
189 CoefType v = block->mdct_coef[ch][i+j];
190 MAC_COEF(energy[blk][ch][bnd], v, v);
198 /* calculate coupling coordinates for all blocks for all channels */
199 for (blk = 0; blk < s->num_blocks; blk++) {
200 AC3Block *block = &s->blocks[blk];
201 if (!block->cpl_in_use)
203 for (ch = 1; ch <= s->fbw_channels; ch++) {
204 if (!block->channel_in_cpl[ch])
206 for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
207 cpl_coords[blk][ch][bnd] = calc_cpl_coord(energy[blk][ch][bnd],
208 energy[blk][CPL_CH][bnd]);
213 /* determine which blocks to send new coupling coordinates for */
214 for (blk = 0; blk < s->num_blocks; blk++) {
215 AC3Block *block = &s->blocks[blk];
216 AC3Block *block0 = blk ? &s->blocks[blk-1] : NULL;
218 memset(block->new_cpl_coords, 0, sizeof(block->new_cpl_coords));
220 if (block->cpl_in_use) {
221 /* send new coordinates if this is the first block, if previous
222 * block did not use coupling but this block does, the channels
223 * using coupling has changed from the previous block, or the
224 * coordinate difference from the last block for any channel is
225 * greater than a threshold value. */
226 if (blk == 0 || !block0->cpl_in_use) {
227 for (ch = 1; ch <= s->fbw_channels; ch++)
228 block->new_cpl_coords[ch] = 1;
230 for (ch = 1; ch <= s->fbw_channels; ch++) {
231 if (!block->channel_in_cpl[ch])
233 if (!block0->channel_in_cpl[ch]) {
234 block->new_cpl_coords[ch] = 1;
236 CoefSumType coord_diff = 0;
237 for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
238 coord_diff += fabs(cpl_coords[blk-1][ch][bnd] -
239 cpl_coords[blk ][ch][bnd]);
241 coord_diff /= s->num_cpl_bands;
242 if (coord_diff > 0.03)
243 block->new_cpl_coords[ch] = 1;
250 /* calculate final coupling coordinates, taking into account reusing of
251 coordinates in successive blocks */
252 for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
254 while (blk < s->num_blocks) {
256 AC3Block *block = &s->blocks[blk];
258 if (!block->cpl_in_use) {
263 for (ch = 1; ch <= s->fbw_channels; ch++) {
264 CoefSumType energy_ch, energy_cpl;
265 if (!block->channel_in_cpl[ch])
267 energy_cpl = energy[blk][CPL_CH][bnd];
268 energy_ch = energy[blk][ch][bnd];
270 while (!s->blocks[blk1].new_cpl_coords[ch] && blk1 < s->num_blocks) {
271 if (s->blocks[blk1].cpl_in_use) {
272 energy_cpl += energy[blk1][CPL_CH][bnd];
273 energy_ch += energy[blk1][ch][bnd];
277 cpl_coords[blk][ch][bnd] = calc_cpl_coord(energy_ch, energy_cpl);
283 /* calculate exponents/mantissas for coupling coordinates */
284 for (blk = 0; blk < s->num_blocks; blk++) {
285 AC3Block *block = &s->blocks[blk];
286 if (!block->cpl_in_use)
289 s->ac3dsp.float_to_fixed24(fixed_cpl_coords[blk][1],
291 s->fbw_channels * 16);
292 s->ac3dsp.extract_exponents(block->cpl_coord_exp[1],
293 fixed_cpl_coords[blk][1],
294 s->fbw_channels * 16);
296 for (ch = 1; ch <= s->fbw_channels; ch++) {
297 int bnd, min_exp, max_exp, master_exp;
299 if (!block->new_cpl_coords[ch])
302 /* determine master exponent */
303 min_exp = max_exp = block->cpl_coord_exp[ch][0];
304 for (bnd = 1; bnd < s->num_cpl_bands; bnd++) {
305 int exp = block->cpl_coord_exp[ch][bnd];
306 min_exp = FFMIN(exp, min_exp);
307 max_exp = FFMAX(exp, max_exp);
309 master_exp = ((max_exp - 15) + 2) / 3;
310 master_exp = FFMAX(master_exp, 0);
311 while (min_exp < master_exp * 3)
313 for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
314 block->cpl_coord_exp[ch][bnd] = av_clip(block->cpl_coord_exp[ch][bnd] -
315 master_exp * 3, 0, 15);
317 block->cpl_master_exp[ch] = master_exp;
319 /* quantize mantissas */
320 for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
321 int cpl_exp = block->cpl_coord_exp[ch][bnd];
322 int cpl_mant = (fixed_cpl_coords[blk][ch][bnd] << (5 + cpl_exp + master_exp * 3)) >> 24;
328 block->cpl_coord_mant[ch][bnd] = cpl_mant;
333 if (CONFIG_EAC3_ENCODER && s->eac3)
334 ff_eac3_set_cpl_states(s);
335 #endif /* CONFIG_AC3ENC_FLOAT */
340 * Determine rematrixing flags for each block and band.
342 static void compute_rematrixing_strategy(AC3EncodeContext *s)
346 AC3Block *block, *av_uninit(block0);
348 if (s->channel_mode != AC3_CHMODE_STEREO)
351 for (blk = 0; blk < s->num_blocks; blk++) {
352 block = &s->blocks[blk];
353 block->new_rematrixing_strategy = !blk;
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;
362 nb_coefs = FFMIN(block->end_freq[1], block->end_freq[2]);
364 if (!s->rematrixing_enabled) {
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] = {0,};
374 for (i = start; i < end; i++) {
375 CoefType lt = block->mdct_coef[1][i];
376 CoefType rt = block->mdct_coef[2][i];
377 CoefType md = lt + rt;
378 CoefType sd = lt - rt;
379 MAC_COEF(sum[0], lt, lt);
380 MAC_COEF(sum[1], rt, rt);
381 MAC_COEF(sum[2], md, md);
382 MAC_COEF(sum[3], sd, sd);
385 /* compare sums to determine if rematrixing will be used for this band */
386 if (FFMIN(sum[2], sum[3]) < FFMIN(sum[0], sum[1]))
387 block->rematrixing_flags[bnd] = 1;
389 block->rematrixing_flags[bnd] = 0;
391 /* determine if new rematrixing flags will be sent */
393 block->rematrixing_flags[bnd] != block0->rematrixing_flags[bnd]) {
394 block->new_rematrixing_strategy = 1;
403 * Encode a single AC-3 frame.
405 int AC3_NAME(encode_frame)(AVCodecContext *avctx, unsigned char *frame,
406 int buf_size, void *data)
408 AC3EncodeContext *s = avctx->priv_data;
409 const SampleType *samples = data;
412 if (s->options.allow_per_frame_metadata) {
413 ret = ff_ac3_validate_metadata(s);
418 if (s->bit_alloc.sr_code == 1 || s->eac3)
419 ff_ac3_adjust_frame_size(s);
421 deinterleave_input_samples(s, samples);
426 scale_coefficients(s);
428 clip_coefficients(&s->dsp, s->blocks[0].mdct_coef[1],
429 AC3_MAX_COEFS * s->num_blocks * s->channels);
431 s->cpl_on = s->cpl_enabled;
432 ff_ac3_compute_coupling_strategy(s);
435 apply_channel_coupling(s);
437 compute_rematrixing_strategy(s);
440 scale_coefficients(s);
442 ff_ac3_apply_rematrixing(s);
444 ff_ac3_process_exponents(s);
446 ret = ff_ac3_compute_bit_allocation(s);
448 av_log(avctx, AV_LOG_ERROR, "Bit allocation failed. Try increasing the bitrate.\n");
452 ff_ac3_group_exponents(s);
454 ff_ac3_quantize_mantissas(s);
456 ff_ac3_output_frame(s, frame);
458 return s->frame_size;