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
31 #include "libavutil/internal.h"
33 /* prototypes for static functions in ac3enc_fixed.c and ac3enc_float.c */
35 static void scale_coefficients(AC3EncodeContext *s);
37 static int normalize_samples(AC3EncodeContext *s);
39 static void clip_coefficients(DSPContext *dsp, CoefType *coef, unsigned int len);
41 static CoefType calc_cpl_coord(CoefSumType energy_ch, CoefSumType energy_cpl);
44 int AC3_NAME(allocate_sample_buffers)(AC3EncodeContext *s)
48 FF_ALLOC_OR_GOTO(s->avctx, s->windowed_samples, AC3_WINDOW_SIZE *
49 sizeof(*s->windowed_samples), alloc_fail);
50 FF_ALLOC_OR_GOTO(s->avctx, s->planar_samples, s->channels * sizeof(*s->planar_samples),
52 for (ch = 0; ch < s->channels; ch++) {
53 FF_ALLOCZ_OR_GOTO(s->avctx, s->planar_samples[ch],
54 (AC3_FRAME_SIZE+AC3_BLOCK_SIZE) * sizeof(**s->planar_samples),
60 return AVERROR(ENOMEM);
66 * Channels are reordered from Libav's default order to AC-3 order.
68 static void copy_input_samples(AC3EncodeContext *s, SampleType **samples)
72 /* copy and remap input samples */
73 for (ch = 0; ch < s->channels; ch++) {
74 /* copy last 256 samples of previous frame to the start of the current frame */
75 memcpy(&s->planar_samples[ch][0], &s->planar_samples[ch][AC3_BLOCK_SIZE * s->num_blocks],
76 AC3_BLOCK_SIZE * sizeof(s->planar_samples[0][0]));
78 /* copy new samples for current frame */
79 memcpy(&s->planar_samples[ch][AC3_BLOCK_SIZE],
80 samples[s->channel_map[ch]],
81 AC3_BLOCK_SIZE * s->num_blocks * sizeof(s->planar_samples[0][0]));
87 * Apply the MDCT to input samples to generate frequency coefficients.
88 * This applies the KBD window and normalizes the input to reduce precision
89 * loss due to fixed-point calculations.
91 static void apply_mdct(AC3EncodeContext *s)
95 for (ch = 0; ch < s->channels; ch++) {
96 for (blk = 0; blk < s->num_blocks; blk++) {
97 AC3Block *block = &s->blocks[blk];
98 const SampleType *input_samples = &s->planar_samples[ch][blk * AC3_BLOCK_SIZE];
100 #if CONFIG_AC3ENC_FLOAT
101 s->fdsp.vector_fmul(s->windowed_samples, input_samples,
102 s->mdct_window, AC3_WINDOW_SIZE);
104 s->ac3dsp.apply_window_int16(s->windowed_samples, input_samples,
105 s->mdct_window, AC3_WINDOW_SIZE);
109 block->coeff_shift[ch+1] = normalize_samples(s);
111 s->mdct.mdct_calcw(&s->mdct, block->mdct_coef[ch+1],
112 s->windowed_samples);
119 * Calculate coupling channel and coupling coordinates.
121 static void apply_channel_coupling(AC3EncodeContext *s)
123 LOCAL_ALIGNED_16(CoefType, cpl_coords, [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
124 #if CONFIG_AC3ENC_FLOAT
125 LOCAL_ALIGNED_16(int32_t, fixed_cpl_coords, [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
127 int32_t (*fixed_cpl_coords)[AC3_MAX_CHANNELS][16] = cpl_coords;
129 int blk, ch, bnd, i, j;
130 CoefSumType energy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][16] = {{{0}}};
131 int cpl_start, num_cpl_coefs;
133 memset(cpl_coords, 0, AC3_MAX_BLOCKS * sizeof(*cpl_coords));
134 #if CONFIG_AC3ENC_FLOAT
135 memset(fixed_cpl_coords, 0, AC3_MAX_BLOCKS * sizeof(*cpl_coords));
138 /* align start to 16-byte boundary. align length to multiple of 32.
139 note: coupling start bin % 4 will always be 1 */
140 cpl_start = s->start_freq[CPL_CH] - 1;
141 num_cpl_coefs = FFALIGN(s->num_cpl_subbands * 12 + 1, 32);
142 cpl_start = FFMIN(256, cpl_start + num_cpl_coefs) - num_cpl_coefs;
144 /* calculate coupling channel from fbw channels */
145 for (blk = 0; blk < s->num_blocks; blk++) {
146 AC3Block *block = &s->blocks[blk];
147 CoefType *cpl_coef = &block->mdct_coef[CPL_CH][cpl_start];
148 if (!block->cpl_in_use)
150 memset(cpl_coef, 0, num_cpl_coefs * sizeof(*cpl_coef));
151 for (ch = 1; ch <= s->fbw_channels; ch++) {
152 CoefType *ch_coef = &block->mdct_coef[ch][cpl_start];
153 if (!block->channel_in_cpl[ch])
155 for (i = 0; i < num_cpl_coefs; i++)
156 cpl_coef[i] += ch_coef[i];
159 /* coefficients must be clipped in order to be encoded */
160 clip_coefficients(&s->dsp, cpl_coef, num_cpl_coefs);
163 /* calculate energy in each band in coupling channel and each fbw channel */
164 /* TODO: possibly use SIMD to speed up energy calculation */
166 i = s->start_freq[CPL_CH];
167 while (i < s->cpl_end_freq) {
168 int band_size = s->cpl_band_sizes[bnd];
169 for (ch = CPL_CH; ch <= s->fbw_channels; ch++) {
170 for (blk = 0; blk < s->num_blocks; blk++) {
171 AC3Block *block = &s->blocks[blk];
172 if (!block->cpl_in_use || (ch > CPL_CH && !block->channel_in_cpl[ch]))
174 for (j = 0; j < band_size; j++) {
175 CoefType v = block->mdct_coef[ch][i+j];
176 MAC_COEF(energy[blk][ch][bnd], v, v);
184 /* calculate coupling coordinates for all blocks for all channels */
185 for (blk = 0; blk < s->num_blocks; blk++) {
186 AC3Block *block = &s->blocks[blk];
187 if (!block->cpl_in_use)
189 for (ch = 1; ch <= s->fbw_channels; ch++) {
190 if (!block->channel_in_cpl[ch])
192 for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
193 cpl_coords[blk][ch][bnd] = calc_cpl_coord(energy[blk][ch][bnd],
194 energy[blk][CPL_CH][bnd]);
199 /* determine which blocks to send new coupling coordinates for */
200 for (blk = 0; blk < s->num_blocks; blk++) {
201 AC3Block *block = &s->blocks[blk];
202 AC3Block *block0 = blk ? &s->blocks[blk-1] : NULL;
204 memset(block->new_cpl_coords, 0, sizeof(block->new_cpl_coords));
206 if (block->cpl_in_use) {
207 /* send new coordinates if this is the first block, if previous
208 * block did not use coupling but this block does, the channels
209 * using coupling has changed from the previous block, or the
210 * coordinate difference from the last block for any channel is
211 * greater than a threshold value. */
212 if (blk == 0 || !block0->cpl_in_use) {
213 for (ch = 1; ch <= s->fbw_channels; ch++)
214 block->new_cpl_coords[ch] = 1;
216 for (ch = 1; ch <= s->fbw_channels; ch++) {
217 if (!block->channel_in_cpl[ch])
219 if (!block0->channel_in_cpl[ch]) {
220 block->new_cpl_coords[ch] = 1;
222 CoefSumType coord_diff = 0;
223 for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
224 coord_diff += FFABS(cpl_coords[blk-1][ch][bnd] -
225 cpl_coords[blk ][ch][bnd]);
227 coord_diff /= s->num_cpl_bands;
228 if (coord_diff > NEW_CPL_COORD_THRESHOLD)
229 block->new_cpl_coords[ch] = 1;
236 /* calculate final coupling coordinates, taking into account reusing of
237 coordinates in successive blocks */
238 for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
240 while (blk < s->num_blocks) {
242 AC3Block *block = &s->blocks[blk];
244 if (!block->cpl_in_use) {
249 for (ch = 1; ch <= s->fbw_channels; ch++) {
250 CoefSumType energy_ch, energy_cpl;
251 if (!block->channel_in_cpl[ch])
253 energy_cpl = energy[blk][CPL_CH][bnd];
254 energy_ch = energy[blk][ch][bnd];
256 while (!s->blocks[blk1].new_cpl_coords[ch] && blk1 < s->num_blocks) {
257 if (s->blocks[blk1].cpl_in_use) {
258 energy_cpl += energy[blk1][CPL_CH][bnd];
259 energy_ch += energy[blk1][ch][bnd];
263 cpl_coords[blk][ch][bnd] = calc_cpl_coord(energy_ch, energy_cpl);
269 /* calculate exponents/mantissas for coupling coordinates */
270 for (blk = 0; blk < s->num_blocks; blk++) {
271 AC3Block *block = &s->blocks[blk];
272 if (!block->cpl_in_use)
275 #if CONFIG_AC3ENC_FLOAT
276 s->ac3dsp.float_to_fixed24(fixed_cpl_coords[blk][1],
278 s->fbw_channels * 16);
280 s->ac3dsp.extract_exponents(block->cpl_coord_exp[1],
281 fixed_cpl_coords[blk][1],
282 s->fbw_channels * 16);
284 for (ch = 1; ch <= s->fbw_channels; ch++) {
285 int bnd, min_exp, max_exp, master_exp;
287 if (!block->new_cpl_coords[ch])
290 /* determine master exponent */
291 min_exp = max_exp = block->cpl_coord_exp[ch][0];
292 for (bnd = 1; bnd < s->num_cpl_bands; bnd++) {
293 int exp = block->cpl_coord_exp[ch][bnd];
294 min_exp = FFMIN(exp, min_exp);
295 max_exp = FFMAX(exp, max_exp);
297 master_exp = ((max_exp - 15) + 2) / 3;
298 master_exp = FFMAX(master_exp, 0);
299 while (min_exp < master_exp * 3)
301 for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
302 block->cpl_coord_exp[ch][bnd] = av_clip(block->cpl_coord_exp[ch][bnd] -
303 master_exp * 3, 0, 15);
305 block->cpl_master_exp[ch] = master_exp;
307 /* quantize mantissas */
308 for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
309 int cpl_exp = block->cpl_coord_exp[ch][bnd];
310 int cpl_mant = (fixed_cpl_coords[blk][ch][bnd] << (5 + cpl_exp + master_exp * 3)) >> 24;
316 block->cpl_coord_mant[ch][bnd] = cpl_mant;
321 if (CONFIG_EAC3_ENCODER && s->eac3)
322 ff_eac3_set_cpl_states(s);
327 * Determine rematrixing flags for each block and band.
329 static void compute_rematrixing_strategy(AC3EncodeContext *s)
333 AC3Block *block, *block0;
335 if (s->channel_mode != AC3_CHMODE_STEREO)
338 for (blk = 0; blk < s->num_blocks; blk++) {
339 block = &s->blocks[blk];
340 block->new_rematrixing_strategy = !blk;
342 block->num_rematrixing_bands = 4;
343 if (block->cpl_in_use) {
344 block->num_rematrixing_bands -= (s->start_freq[CPL_CH] <= 61);
345 block->num_rematrixing_bands -= (s->start_freq[CPL_CH] == 37);
346 if (blk && block->num_rematrixing_bands != block0->num_rematrixing_bands)
347 block->new_rematrixing_strategy = 1;
349 nb_coefs = FFMIN(block->end_freq[1], block->end_freq[2]);
351 if (!s->rematrixing_enabled) {
356 for (bnd = 0; bnd < block->num_rematrixing_bands; bnd++) {
357 /* calculate calculate sum of squared coeffs for one band in one block */
358 int start = ff_ac3_rematrix_band_tab[bnd];
359 int end = FFMIN(nb_coefs, ff_ac3_rematrix_band_tab[bnd+1]);
360 CoefSumType sum[4] = {0,};
361 for (i = start; i < end; i++) {
362 CoefType lt = block->mdct_coef[1][i];
363 CoefType rt = block->mdct_coef[2][i];
364 CoefType md = lt + rt;
365 CoefType sd = lt - rt;
366 MAC_COEF(sum[0], lt, lt);
367 MAC_COEF(sum[1], rt, rt);
368 MAC_COEF(sum[2], md, md);
369 MAC_COEF(sum[3], sd, sd);
372 /* compare sums to determine if rematrixing will be used for this band */
373 if (FFMIN(sum[2], sum[3]) < FFMIN(sum[0], sum[1]))
374 block->rematrixing_flags[bnd] = 1;
376 block->rematrixing_flags[bnd] = 0;
378 /* determine if new rematrixing flags will be sent */
380 block->rematrixing_flags[bnd] != block0->rematrixing_flags[bnd]) {
381 block->new_rematrixing_strategy = 1;
389 int AC3_NAME(encode_frame)(AVCodecContext *avctx, AVPacket *avpkt,
390 const AVFrame *frame, int *got_packet_ptr)
392 AC3EncodeContext *s = avctx->priv_data;
395 if (s->options.allow_per_frame_metadata) {
396 ret = ff_ac3_validate_metadata(s);
401 if (s->bit_alloc.sr_code == 1 || s->eac3)
402 ff_ac3_adjust_frame_size(s);
404 copy_input_samples(s, (SampleType **)frame->extended_data);
409 scale_coefficients(s);
411 clip_coefficients(&s->dsp, s->blocks[0].mdct_coef[1],
412 AC3_MAX_COEFS * s->num_blocks * s->channels);
414 s->cpl_on = s->cpl_enabled;
415 ff_ac3_compute_coupling_strategy(s);
418 apply_channel_coupling(s);
420 compute_rematrixing_strategy(s);
423 scale_coefficients(s);
425 ff_ac3_apply_rematrixing(s);
427 ff_ac3_process_exponents(s);
429 ret = ff_ac3_compute_bit_allocation(s);
431 av_log(avctx, AV_LOG_ERROR, "Bit allocation failed. Try increasing the bitrate.\n");
435 ff_ac3_group_exponents(s);
437 ff_ac3_quantize_mantissas(s);
439 if ((ret = ff_alloc_packet(avpkt, s->frame_size))) {
440 av_log(avctx, AV_LOG_ERROR, "Error getting output packet\n");
443 ff_ac3_output_frame(s, avpkt->data);
445 if (frame->pts != AV_NOPTS_VALUE)
446 avpkt->pts = frame->pts - ff_samples_to_time_base(avctx, avctx->delay);