git.sesse.net Git - ffmpeg/blob - 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/attributes.h"
  32 #include "libavutil/internal.h"
  33 #include "libavutil/mem_internal.h"
  34
  35 #include "audiodsp.h"
  36 #include "internal.h"
  37 #include "ac3enc.h"
  38 #include "eac3enc.h"
  39
  40
  41 static int allocate_sample_buffers(AC3EncodeContext *s)
  42 {
  43     int ch;
  44
  45     if (!FF_ALLOC_TYPED_ARRAY(s->windowed_samples, AC3_WINDOW_SIZE) ||
  46         !FF_ALLOCZ_TYPED_ARRAY(s->planar_samples,  s->channels))
  47         return AVERROR(ENOMEM);
  48
  49     for (ch = 0; ch < s->channels; ch++) {
  50         if (!(s->planar_samples[ch] = av_mallocz((AC3_FRAME_SIZE + AC3_BLOCK_SIZE) *
  51                                                   sizeof(**s->planar_samples))))
  52             return AVERROR(ENOMEM);
  53     }
  54     return 0;
  55 }
  56
  57
  58 /*
  59  * Copy input samples.
  60  * Channels are reordered from FFmpeg's default order to AC-3 order.
  61  */
  62 static void copy_input_samples(AC3EncodeContext *s, SampleType **samples)
  63 {
  64     int ch;
  65
  66     /* copy and remap input samples */
  67     for (ch = 0; ch < s->channels; ch++) {
  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_BLOCK_SIZE * s->num_blocks],
  70                AC3_BLOCK_SIZE * sizeof(s->planar_samples[0][0]));
  71
  72         /* copy new samples for current frame */
  73         memcpy(&s->planar_samples[ch][AC3_BLOCK_SIZE],
  74                samples[s->channel_map[ch]],
  75                AC3_BLOCK_SIZE * s->num_blocks * sizeof(s->planar_samples[0][0]));
  76     }
  77 }
  78
  79
  80 /*
  81  * Apply the MDCT to input samples to generate frequency coefficients.
  82  * This applies the KBD window and normalizes the input to reduce precision
  83  * loss due to fixed-point calculations.
  84  */
  85 static void apply_mdct(AC3EncodeContext *s)
  86 {
  87     int blk, ch;
  88
  89     for (ch = 0; ch < s->channels; ch++) {
  90         for (blk = 0; blk < s->num_blocks; blk++) {
  91             AC3Block *block = &s->blocks[blk];
  92             const SampleType *input_samples = &s->planar_samples[ch][blk * AC3_BLOCK_SIZE];
  93
  94             s->fdsp->vector_fmul(s->windowed_samples, input_samples,
  95                                  s->mdct_window, AC3_WINDOW_SIZE);
  96
  97             s->mdct.mdct_calc(&s->mdct, block->mdct_coef[ch+1],
  98                               s->windowed_samples);
  99         }
 100     }
 101 }
 102
 103
 104 /*
 105  * Calculate coupling channel and coupling coordinates.
 106  */
 107 static void apply_channel_coupling(AC3EncodeContext *s)
 108 {
 109     LOCAL_ALIGNED_16(CoefType, cpl_coords,      [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
 110 #if AC3ENC_FLOAT
 111     LOCAL_ALIGNED_16(int32_t, fixed_cpl_coords, [AC3_MAX_BLOCKS], [AC3_MAX_CHANNELS][16]);
 112 #else
 113     int32_t (*fixed_cpl_coords)[AC3_MAX_CHANNELS][16] = cpl_coords;
 114 #endif
 115     int av_uninit(blk), ch, bnd, i, j;
 116     CoefSumType energy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][16] = {{{0}}};
 117     int cpl_start, num_cpl_coefs;
 118
 119     memset(cpl_coords,       0, AC3_MAX_BLOCKS * sizeof(*cpl_coords));
 120 #if AC3ENC_FLOAT
 121     memset(fixed_cpl_coords, 0, AC3_MAX_BLOCKS * sizeof(*cpl_coords));
 122 #endif
 123
 124     /* align start to 16-byte boundary. align length to multiple of 32.
 125         note: coupling start bin % 4 will always be 1 */
 126     cpl_start     = s->start_freq[CPL_CH] - 1;
 127     num_cpl_coefs = FFALIGN(s->num_cpl_subbands * 12 + 1, 32);
 128     cpl_start     = FFMIN(256, cpl_start + num_cpl_coefs) - num_cpl_coefs;
 129
 130     /* calculate coupling channel from fbw channels */
 131     for (blk = 0; blk < s->num_blocks; blk++) {
 132         AC3Block *block = &s->blocks[blk];
 133         CoefType *cpl_coef = &block->mdct_coef[CPL_CH][cpl_start];
 134         if (!block->cpl_in_use)
 135             continue;
 136         memset(cpl_coef, 0, num_cpl_coefs * sizeof(*cpl_coef));
 137         for (ch = 1; ch <= s->fbw_channels; ch++) {
 138             CoefType *ch_coef = &block->mdct_coef[ch][cpl_start];
 139             if (!block->channel_in_cpl[ch])
 140                 continue;
 141             for (i = 0; i < num_cpl_coefs; i++)
 142                 cpl_coef[i] += ch_coef[i];
 143         }
 144
 145         /* coefficients must be clipped in order to be encoded */
 146         clip_coefficients(&s->adsp, cpl_coef, num_cpl_coefs);
 147     }
 148
 149     /* calculate energy in each band in coupling channel and each fbw channel */
 150     /* TODO: possibly use SIMD to speed up energy calculation */
 151     bnd = 0;
 152     i = s->start_freq[CPL_CH];
 153     while (i < s->cpl_end_freq) {
 154         int band_size = s->cpl_band_sizes[bnd];
 155         for (ch = CPL_CH; ch <= s->fbw_channels; ch++) {
 156             for (blk = 0; blk < s->num_blocks; blk++) {
 157                 AC3Block *block = &s->blocks[blk];
 158                 if (!block->cpl_in_use || (ch > CPL_CH && !block->channel_in_cpl[ch]))
 159                     continue;
 160                 for (j = 0; j < band_size; j++) {
 161                     CoefType v = block->mdct_coef[ch][i+j];
 162                     MAC_COEF(energy[blk][ch][bnd], v, v);
 163                 }
 164             }
 165         }
 166         i += band_size;
 167         bnd++;
 168     }
 169
 170     /* calculate coupling coordinates for all blocks for all channels */
 171     for (blk = 0; blk < s->num_blocks; blk++) {
 172         AC3Block *block  = &s->blocks[blk];
 173         if (!block->cpl_in_use)
 174             continue;
 175         for (ch = 1; ch <= s->fbw_channels; ch++) {
 176             if (!block->channel_in_cpl[ch])
 177                 continue;
 178             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
 179                 cpl_coords[blk][ch][bnd] = calc_cpl_coord(energy[blk][ch][bnd],
 180                                                           energy[blk][CPL_CH][bnd]);
 181             }
 182         }
 183     }
 184
 185     /* determine which blocks to send new coupling coordinates for */
 186     for (blk = 0; blk < s->num_blocks; blk++) {
 187         AC3Block *block  = &s->blocks[blk];
 188         AC3Block *block0 = blk ? &s->blocks[blk-1] : NULL;
 189
 190         memset(block->new_cpl_coords, 0, sizeof(block->new_cpl_coords));
 191
 192         if (block->cpl_in_use) {
 193             /* send new coordinates if this is the first block, if previous
 194              * block did not use coupling but this block does, the channels
 195              * using coupling has changed from the previous block, or the
 196              * coordinate difference from the last block for any channel is
 197              * greater than a threshold value. */
 198             if (blk == 0 || !block0->cpl_in_use) {
 199                 for (ch = 1; ch <= s->fbw_channels; ch++)
 200                     block->new_cpl_coords[ch] = 1;
 201             } else {
 202                 for (ch = 1; ch <= s->fbw_channels; ch++) {
 203                     if (!block->channel_in_cpl[ch])
 204                         continue;
 205                     if (!block0->channel_in_cpl[ch]) {
 206                         block->new_cpl_coords[ch] = 1;
 207                     } else {
 208                         CoefSumType coord_diff = 0;
 209                         for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
 210                             coord_diff += FFABS(cpl_coords[blk-1][ch][bnd] -
 211                                                 cpl_coords[blk  ][ch][bnd]);
 212                         }
 213                         coord_diff /= s->num_cpl_bands;
 214                         if (coord_diff > NEW_CPL_COORD_THRESHOLD)
 215                             block->new_cpl_coords[ch] = 1;
 216                     }
 217                 }
 218             }
 219         }
 220     }
 221
 222     /* calculate final coupling coordinates, taking into account reusing of
 223        coordinates in successive blocks */
 224     for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
 225         blk = 0;
 226         while (blk < s->num_blocks) {
 227             int av_uninit(blk1);
 228             AC3Block *block  = &s->blocks[blk];
 229
 230             if (!block->cpl_in_use) {
 231                 blk++;
 232                 continue;
 233             }
 234
 235             for (ch = 1; ch <= s->fbw_channels; ch++) {
 236                 CoefSumType energy_ch, energy_cpl;
 237                 if (!block->channel_in_cpl[ch])
 238                     continue;
 239                 energy_cpl = energy[blk][CPL_CH][bnd];
 240                 energy_ch = energy[blk][ch][bnd];
 241                 blk1 = blk+1;
 242                 while (blk1 < s->num_blocks && !s->blocks[blk1].new_cpl_coords[ch]) {
 243                     if (s->blocks[blk1].cpl_in_use) {
 244                         energy_cpl += energy[blk1][CPL_CH][bnd];
 245                         energy_ch += energy[blk1][ch][bnd];
 246                     }
 247                     blk1++;
 248                 }
 249                 cpl_coords[blk][ch][bnd] = calc_cpl_coord(energy_ch, energy_cpl);
 250             }
 251             blk = blk1;
 252         }
 253     }
 254
 255     /* calculate exponents/mantissas for coupling coordinates */
 256     for (blk = 0; blk < s->num_blocks; blk++) {
 257         AC3Block *block = &s->blocks[blk];
 258         if (!block->cpl_in_use)
 259             continue;
 260
 261 #if AC3ENC_FLOAT
 262         s->ac3dsp.float_to_fixed24(fixed_cpl_coords[blk][1],
 263                                    cpl_coords[blk][1],
 264                                    s->fbw_channels * 16);
 265 #endif
 266         s->ac3dsp.extract_exponents(block->cpl_coord_exp[1],
 267                                     fixed_cpl_coords[blk][1],
 268                                     s->fbw_channels * 16);
 269
 270         for (ch = 1; ch <= s->fbw_channels; ch++) {
 271             int bnd, min_exp, max_exp, master_exp;
 272
 273             if (!block->new_cpl_coords[ch])
 274                 continue;
 275
 276             /* determine master exponent */
 277             min_exp = max_exp = block->cpl_coord_exp[ch][0];
 278             for (bnd = 1; bnd < s->num_cpl_bands; bnd++) {
 279                 int exp = block->cpl_coord_exp[ch][bnd];
 280                 min_exp = FFMIN(exp, min_exp);
 281                 max_exp = FFMAX(exp, max_exp);
 282             }
 283             master_exp = ((max_exp - 15) + 2) / 3;
 284             master_exp = FFMAX(master_exp, 0);
 285             while (min_exp < master_exp * 3)
 286                 master_exp--;
 287             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
 288                 block->cpl_coord_exp[ch][bnd] = av_clip(block->cpl_coord_exp[ch][bnd] -
 289                                                         master_exp * 3, 0, 15);
 290             }
 291             block->cpl_master_exp[ch] = master_exp;
 292
 293             /* quantize mantissas */
 294             for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
 295                 int cpl_exp  = block->cpl_coord_exp[ch][bnd];
 296                 int cpl_mant = (fixed_cpl_coords[blk][ch][bnd] << (5 + cpl_exp + master_exp * 3)) >> 24;
 297                 if (cpl_exp == 15)
 298                     cpl_mant >>= 1;
 299                 else
 300                     cpl_mant -= 16;
 301
 302                 block->cpl_coord_mant[ch][bnd] = cpl_mant;
 303             }
 304         }
 305     }
 306
 307     if (AC3ENC_FLOAT && CONFIG_EAC3_ENCODER && s->eac3)
 308         ff_eac3_set_cpl_states(s);
 309 }
 310
 311
 312 /*
 313  * Determine rematrixing flags for each block and band.
 314  */
 315 static void compute_rematrixing_strategy(AC3EncodeContext *s)
 316 {
 317     int nb_coefs;
 318     int blk, bnd;
 319     AC3Block *block, *block0 = NULL;
 320
 321     if (s->channel_mode != AC3_CHMODE_STEREO)
 322         return;
 323
 324     for (blk = 0; blk < s->num_blocks; blk++) {
 325         block = &s->blocks[blk];
 326         block->new_rematrixing_strategy = !blk;
 327
 328         block->num_rematrixing_bands = 4;
 329         if (block->cpl_in_use) {
 330             block->num_rematrixing_bands -= (s->start_freq[CPL_CH] <= 61);
 331             block->num_rematrixing_bands -= (s->start_freq[CPL_CH] == 37);
 332             if (blk && block->num_rematrixing_bands != block0->num_rematrixing_bands)
 333                 block->new_rematrixing_strategy = 1;
 334         }
 335         nb_coefs = FFMIN(block->end_freq[1], block->end_freq[2]);
 336
 337         if (!s->rematrixing_enabled) {
 338             block0 = block;
 339             continue;
 340         }
 341
 342         for (bnd = 0; bnd < block->num_rematrixing_bands; bnd++) {
 343             /* calculate sum of squared coeffs for one band in one block */
 344             int start = ff_ac3_rematrix_band_tab[bnd];
 345             int end   = FFMIN(nb_coefs, ff_ac3_rematrix_band_tab[bnd+1]);
 346             CoefSumType sum[4];
 347             sum_square_butterfly(s, sum, block->mdct_coef[1] + start,
 348                                  block->mdct_coef[2] + start, end - start);
 349
 350             /* compare sums to determine if rematrixing will be used for this band */
 351             if (FFMIN(sum[2], sum[3]) < FFMIN(sum[0], sum[1]))
 352                 block->rematrixing_flags[bnd] = 1;
 353             else
 354                 block->rematrixing_flags[bnd] = 0;
 355
 356             /* determine if new rematrixing flags will be sent */
 357             if (blk &&
 358                 block->rematrixing_flags[bnd] != block0->rematrixing_flags[bnd]) {
 359                 block->new_rematrixing_strategy = 1;
 360             }
 361         }
 362         block0 = block;
 363     }
 364 }
 365
 366
 367 int AC3_NAME(encode_frame)(AVCodecContext *avctx, AVPacket *avpkt,
 368                            const AVFrame *frame, int *got_packet_ptr)
 369 {
 370     AC3EncodeContext *s = avctx->priv_data;
 371     int ret;
 372
 373     if (s->options.allow_per_frame_metadata) {
 374         ret = ff_ac3_validate_metadata(s);
 375         if (ret)
 376             return ret;
 377     }
 378
 379     if (s->bit_alloc.sr_code == 1 || (AC3ENC_FLOAT && s->eac3))
 380         ff_ac3_adjust_frame_size(s);
 381
 382     copy_input_samples(s, (SampleType **)frame->extended_data);
 383
 384     apply_mdct(s);
 385
 386     clip_coefficients(&s->adsp, s->blocks[0].mdct_coef[1],
 387                       AC3_MAX_COEFS * s->num_blocks * s->channels);
 388
 389     s->cpl_on = s->cpl_enabled;
 390     ff_ac3_compute_coupling_strategy(s);
 391
 392     if (s->cpl_on)
 393         apply_channel_coupling(s);
 394
 395     compute_rematrixing_strategy(s);
 396
 397 #if AC3ENC_FLOAT
 398     scale_coefficients(s);
 399 #endif
 400
 401     return ff_ac3_encode_frame_common_end(avctx, avpkt, frame, got_packet_ptr);
 402 }