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
34 int AC3_NAME(allocate_sample_buffers)(AC3EncodeContext *s)
38 FF_ALLOC_OR_GOTO(s->avctx, s->windowed_samples, AC3_WINDOW_SIZE *
39 sizeof(*s->windowed_samples), alloc_fail);
40 FF_ALLOC_OR_GOTO(s->avctx, s->planar_samples, s->channels * sizeof(*s->planar_samples),
42 for (ch = 0; ch < s->channels; ch++) {
43 FF_ALLOCZ_OR_GOTO(s->avctx, s->planar_samples[ch],
44 (AC3_FRAME_SIZE+AC3_BLOCK_SIZE) * sizeof(**s->planar_samples),
50 return AVERROR(ENOMEM);
55 * Deinterleave input samples.
56 * Channels are reordered from Libav's default order to AC-3 order.
58 void AC3_NAME(deinterleave_input_samples)(AC3EncodeContext *s,
59 const SampleType *samples)
63 /* deinterleave and remap input samples */
64 for (ch = 0; ch < s->channels; ch++) {
65 const SampleType *sptr;
68 /* copy last 256 samples of previous frame to the start of the current frame */
69 memcpy(&s->planar_samples[ch][0], &s->planar_samples[ch][AC3_FRAME_SIZE],
70 AC3_BLOCK_SIZE * sizeof(s->planar_samples[0][0]));
74 sptr = samples + s->channel_map[ch];
75 for (i = AC3_BLOCK_SIZE; i < AC3_FRAME_SIZE+AC3_BLOCK_SIZE; i++) {
76 s->planar_samples[ch][i] = *sptr;
84 * Apply the MDCT to input samples to generate frequency coefficients.
85 * This applies the KBD window and normalizes the input to reduce precision
86 * loss due to fixed-point calculations.
88 void AC3_NAME(apply_mdct)(AC3EncodeContext *s)
92 for (ch = 0; ch < s->channels; ch++) {
93 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
94 AC3Block *block = &s->blocks[blk];
95 const SampleType *input_samples = &s->planar_samples[ch][blk * AC3_BLOCK_SIZE];
97 s->apply_window(&s->dsp, s->windowed_samples, input_samples,
98 s->mdct->window, AC3_WINDOW_SIZE);
101 block->coeff_shift[ch+1] = s->normalize_samples(s);
103 s->mdct->fft.mdct_calcw(&s->mdct->fft, block->mdct_coef[ch+1],
104 s->windowed_samples);
111 * Calculate a single coupling coordinate.
113 static inline float calc_cpl_coord(float energy_ch, float energy_cpl)
117 coord *= sqrtf(energy_ch / energy_cpl);
123 * Calculate coupling channel and coupling coordinates.
124 * TODO: Currently this is only used for the floating-point encoder. I was
125 * able to make it work for the fixed-point encoder, but quality was
126 * generally lower in most cases than not using coupling. If a more
127 * adaptive coupling strategy were to be implemented it might be useful
128 * at that time to use coupling for the fixed-point encoder as well.
130 void AC3_NAME(apply_channel_coupling)(AC3EncodeContext *s)
132 #if CONFIG_AC3ENC_FLOAT
133 LOCAL_ALIGNED_16(float, cpl_coords, [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
134 LOCAL_ALIGNED_16(int32_t, fixed_cpl_coords, [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
135 int blk, ch, bnd, i, j;
136 CoefSumType energy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][16] = {{{0}}};
137 int cpl_start, num_cpl_coefs;
139 memset(cpl_coords, 0, AC3_MAX_BLOCKS * sizeof(*cpl_coords));
140 memset(fixed_cpl_coords, 0, AC3_MAX_BLOCKS * sizeof(*fixed_cpl_coords));
142 /* align start to 16-byte boundary. align length to multiple of 32.
143 note: coupling start bin % 4 will always be 1 */
144 cpl_start = s->start_freq[CPL_CH] - 1;
145 num_cpl_coefs = FFALIGN(s->num_cpl_subbands * 12 + 1, 32);
146 cpl_start = FFMIN(256, cpl_start + num_cpl_coefs) - num_cpl_coefs;
148 /* calculate coupling channel from fbw channels */
149 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
150 AC3Block *block = &s->blocks[blk];
151 CoefType *cpl_coef = &block->mdct_coef[CPL_CH][cpl_start];
152 if (!block->cpl_in_use)
154 memset(cpl_coef, 0, num_cpl_coefs * sizeof(*cpl_coef));
155 for (ch = 1; ch <= s->fbw_channels; ch++) {
156 CoefType *ch_coef = &block->mdct_coef[ch][cpl_start];
157 if (!block->channel_in_cpl[ch])
159 for (i = 0; i < num_cpl_coefs; i++)
160 cpl_coef[i] += ch_coef[i];
163 /* coefficients must be clipped to +/- 1.0 in order to be encoded */
164 s->dsp.vector_clipf(cpl_coef, cpl_coef, -1.0f, 1.0f, num_cpl_coefs);
166 /* scale coupling coefficients from float to 24-bit fixed-point */
167 s->ac3dsp.float_to_fixed24(&block->fixed_coef[CPL_CH][cpl_start],
168 cpl_coef, num_cpl_coefs);
171 /* calculate energy in each band in coupling channel and each fbw channel */
172 /* TODO: possibly use SIMD to speed up energy calculation */
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 < AC3_MAX_BLOCKS; blk++) {
179 AC3Block *block = &s->blocks[blk];
180 if (!block->cpl_in_use || (ch > CPL_CH && !block->channel_in_cpl[ch]))
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);
192 /* determine which blocks to send new coupling coordinates for */
193 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
194 AC3Block *block = &s->blocks[blk];
195 AC3Block *block0 = blk ? &s->blocks[blk-1] : NULL;
197 CoefSumType coord_diff[AC3_MAX_CHANNELS] = {0,};
199 if (block->cpl_in_use) {
200 /* calculate coupling coordinates for all blocks and calculate the
201 average difference between coordinates in successive blocks */
202 for (ch = 1; ch <= s->fbw_channels; ch++) {
203 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]);
209 if (blk > 0 && block0->cpl_in_use &&
210 block0->channel_in_cpl[ch]) {
211 coord_diff[ch] += fabs(cpl_coords[blk-1][ch][bnd] -
212 cpl_coords[blk ][ch][bnd]);
215 coord_diff[ch] /= s->num_cpl_bands;
218 /* send new coordinates if this is the first block, if previous
219 * block did not use coupling but this block does, the channels
220 * using coupling has changed from the previous block, or the
221 * coordinate difference from the last block for any channel is
222 * greater than a threshold value. */
225 } else if (!block0->cpl_in_use) {
228 for (ch = 1; ch <= s->fbw_channels; ch++) {
229 if (block->channel_in_cpl[ch] && !block0->channel_in_cpl[ch]) {
235 for (ch = 1; ch <= s->fbw_channels; ch++) {
236 if (block->channel_in_cpl[ch] && coord_diff[ch] > 0.04) {
244 block->new_cpl_coords = new_coords;
247 /* calculate final coupling coordinates, taking into account reusing of
248 coordinates in successive blocks */
249 for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
251 while (blk < AC3_MAX_BLOCKS) {
253 CoefSumType energy_cpl;
254 AC3Block *block = &s->blocks[blk];
256 if (!block->cpl_in_use) {
261 energy_cpl = energy[blk][CPL_CH][bnd];
263 while (!s->blocks[blk1].new_cpl_coords && blk1 < AC3_MAX_BLOCKS) {
264 if (s->blocks[blk1].cpl_in_use)
265 energy_cpl += energy[blk1][CPL_CH][bnd];
269 for (ch = 1; ch <= s->fbw_channels; ch++) {
271 if (!block->channel_in_cpl[ch])
273 energy_ch = energy[blk][ch][bnd];
275 while (!s->blocks[blk1].new_cpl_coords && blk1 < AC3_MAX_BLOCKS) {
276 if (s->blocks[blk1].cpl_in_use)
277 energy_ch += energy[blk1][ch][bnd];
280 cpl_coords[blk][ch][bnd] = calc_cpl_coord(energy_ch, energy_cpl);
286 /* calculate exponents/mantissas for coupling coordinates */
287 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
288 AC3Block *block = &s->blocks[blk];
289 if (!block->cpl_in_use || !block->new_cpl_coords)
292 s->ac3dsp.float_to_fixed24(fixed_cpl_coords[blk][1],
294 s->fbw_channels * 16);
295 s->ac3dsp.extract_exponents(block->cpl_coord_exp[1],
296 fixed_cpl_coords[blk][1],
297 s->fbw_channels * 16);
299 for (ch = 1; ch <= s->fbw_channels; ch++) {
300 int bnd, min_exp, max_exp, master_exp;
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 void AC3_NAME(compute_rematrixing_strategy)(AC3EncodeContext *s)
346 AC3Block *block, *av_uninit(block0);
348 if (s->channel_mode != AC3_CHMODE_STEREO)
351 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
352 block = &s->blocks[blk];
353 block->new_rematrixing_strategy = !blk;
355 if (!s->rematrixing_enabled) {
360 block->num_rematrixing_bands = 4;
361 if (block->cpl_in_use) {
362 block->num_rematrixing_bands -= (s->start_freq[CPL_CH] <= 61);
363 block->num_rematrixing_bands -= (s->start_freq[CPL_CH] == 37);
364 if (blk && block->num_rematrixing_bands != block0->num_rematrixing_bands)
365 block->new_rematrixing_strategy = 1;
367 nb_coefs = FFMIN(block->end_freq[1], block->end_freq[2]);
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;