FFmpeg
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
ra144enc.c
Go to the documentation of this file.
1 /*
2  * Real Audio 1.0 (14.4K) encoder
3  * Copyright (c) 2010 Francesco Lavra <francescolavra@interfree.it>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21 
22 /**
23  * @file
24  * Real Audio 1.0 (14.4K) encoder
25  * @author Francesco Lavra <francescolavra@interfree.it>
26  */
27 
28 #include <float.h>
29 
30 #include "avcodec.h"
31 #include "audio_frame_queue.h"
32 #include "internal.h"
33 #include "put_bits.h"
34 #include "celp_filters.h"
35 #include "ra144.h"
36 
37 
39 {
40  RA144Context *ractx = avctx->priv_data;
41  ff_lpc_end(&ractx->lpc_ctx);
42  ff_af_queue_close(&ractx->afq);
43  return 0;
44 }
45 
46 
48 {
49  RA144Context *ractx;
50  int ret;
51 
52  if (avctx->channels != 1) {
53  av_log(avctx, AV_LOG_ERROR, "invalid number of channels: %d\n",
54  avctx->channels);
55  return -1;
56  }
57  avctx->frame_size = NBLOCKS * BLOCKSIZE;
58  avctx->delay = avctx->frame_size;
59  avctx->bit_rate = 8000;
60  ractx = avctx->priv_data;
61  ractx->lpc_coef[0] = ractx->lpc_tables[0];
62  ractx->lpc_coef[1] = ractx->lpc_tables[1];
63  ractx->avctx = avctx;
64  ff_dsputil_init(&ractx->dsp, avctx);
65  ret = ff_lpc_init(&ractx->lpc_ctx, avctx->frame_size, LPC_ORDER,
67  if (ret < 0)
68  goto error;
69 
70  ff_af_queue_init(avctx, &ractx->afq);
71 
72  return 0;
73 error:
74  ra144_encode_close(avctx);
75  return ret;
76 }
77 
78 
79 /**
80  * Quantize a value by searching a sorted table for the element with the
81  * nearest value
82  *
83  * @param value value to quantize
84  * @param table array containing the quantization table
85  * @param size size of the quantization table
86  * @return index of the quantization table corresponding to the element with the
87  * nearest value
88  */
89 static int quantize(int value, const int16_t *table, unsigned int size)
90 {
91  unsigned int low = 0, high = size - 1;
92 
93  while (1) {
94  int index = (low + high) >> 1;
95  int error = table[index] - value;
96 
97  if (index == low)
98  return table[high] + error > value ? low : high;
99  if (error > 0) {
100  high = index;
101  } else {
102  low = index;
103  }
104  }
105 }
106 
107 
108 /**
109  * Orthogonalize a vector to another vector
110  *
111  * @param v vector to orthogonalize
112  * @param u vector against which orthogonalization is performed
113  */
114 static void orthogonalize(float *v, const float *u)
115 {
116  int i;
117  float num = 0, den = 0;
118 
119  for (i = 0; i < BLOCKSIZE; i++) {
120  num += v[i] * u[i];
121  den += u[i] * u[i];
122  }
123  num /= den;
124  for (i = 0; i < BLOCKSIZE; i++)
125  v[i] -= num * u[i];
126 }
127 
128 
129 /**
130  * Calculate match score and gain of an LPC-filtered vector with respect to
131  * input data, possibly othogonalizing it to up to 2 other vectors
132  *
133  * @param work array used to calculate the filtered vector
134  * @param coefs coefficients of the LPC filter
135  * @param vect original vector
136  * @param ortho1 first vector against which orthogonalization is performed
137  * @param ortho2 second vector against which orthogonalization is performed
138  * @param data input data
139  * @param score pointer to variable where match score is returned
140  * @param gain pointer to variable where gain is returned
141  */
142 static void get_match_score(float *work, const float *coefs, float *vect,
143  const float *ortho1, const float *ortho2,
144  const float *data, float *score, float *gain)
145 {
146  float c, g;
147  int i;
148 
149  ff_celp_lp_synthesis_filterf(work, coefs, vect, BLOCKSIZE, LPC_ORDER);
150  if (ortho1)
151  orthogonalize(work, ortho1);
152  if (ortho2)
153  orthogonalize(work, ortho2);
154  c = g = 0;
155  for (i = 0; i < BLOCKSIZE; i++) {
156  g += work[i] * work[i];
157  c += data[i] * work[i];
158  }
159  if (c <= 0) {
160  *score = 0;
161  return;
162  }
163  *gain = c / g;
164  *score = *gain * c;
165 }
166 
167 
168 /**
169  * Create a vector from the adaptive codebook at a given lag value
170  *
171  * @param vect array where vector is stored
172  * @param cb adaptive codebook
173  * @param lag lag value
174  */
175 static void create_adapt_vect(float *vect, const int16_t *cb, int lag)
176 {
177  int i;
178 
179  cb += BUFFERSIZE - lag;
180  for (i = 0; i < FFMIN(BLOCKSIZE, lag); i++)
181  vect[i] = cb[i];
182  if (lag < BLOCKSIZE)
183  for (i = 0; i < BLOCKSIZE - lag; i++)
184  vect[lag + i] = cb[i];
185 }
186 
187 
188 /**
189  * Search the adaptive codebook for the best entry and gain and remove its
190  * contribution from input data
191  *
192  * @param adapt_cb array from which the adaptive codebook is extracted
193  * @param work array used to calculate LPC-filtered vectors
194  * @param coefs coefficients of the LPC filter
195  * @param data input data
196  * @return index of the best entry of the adaptive codebook
197  */
198 static int adaptive_cb_search(const int16_t *adapt_cb, float *work,
199  const float *coefs, float *data)
200 {
201  int i, av_uninit(best_vect);
202  float score, gain, best_score, av_uninit(best_gain);
203  float exc[BLOCKSIZE];
204 
205  gain = best_score = 0;
206  for (i = BLOCKSIZE / 2; i <= BUFFERSIZE; i++) {
207  create_adapt_vect(exc, adapt_cb, i);
208  get_match_score(work, coefs, exc, NULL, NULL, data, &score, &gain);
209  if (score > best_score) {
210  best_score = score;
211  best_vect = i;
212  best_gain = gain;
213  }
214  }
215  if (!best_score)
216  return 0;
217 
218  /**
219  * Re-calculate the filtered vector from the vector with maximum match score
220  * and remove its contribution from input data.
221  */
222  create_adapt_vect(exc, adapt_cb, best_vect);
224  for (i = 0; i < BLOCKSIZE; i++)
225  data[i] -= best_gain * work[i];
226  return best_vect - BLOCKSIZE / 2 + 1;
227 }
228 
229 
230 /**
231  * Find the best vector of a fixed codebook by applying an LPC filter to
232  * codebook entries, possibly othogonalizing them to up to 2 other vectors and
233  * matching the results with input data
234  *
235  * @param work array used to calculate the filtered vectors
236  * @param coefs coefficients of the LPC filter
237  * @param cb fixed codebook
238  * @param ortho1 first vector against which orthogonalization is performed
239  * @param ortho2 second vector against which orthogonalization is performed
240  * @param data input data
241  * @param idx pointer to variable where the index of the best codebook entry is
242  * returned
243  * @param gain pointer to variable where the gain of the best codebook entry is
244  * returned
245  */
246 static void find_best_vect(float *work, const float *coefs,
247  const int8_t cb[][BLOCKSIZE], const float *ortho1,
248  const float *ortho2, float *data, int *idx,
249  float *gain)
250 {
251  int i, j;
252  float g, score, best_score;
253  float vect[BLOCKSIZE];
254 
255  *idx = *gain = best_score = 0;
256  for (i = 0; i < FIXED_CB_SIZE; i++) {
257  for (j = 0; j < BLOCKSIZE; j++)
258  vect[j] = cb[i][j];
259  get_match_score(work, coefs, vect, ortho1, ortho2, data, &score, &g);
260  if (score > best_score) {
261  best_score = score;
262  *idx = i;
263  *gain = g;
264  }
265  }
266 }
267 
268 
269 /**
270  * Search the two fixed codebooks for the best entry and gain
271  *
272  * @param work array used to calculate LPC-filtered vectors
273  * @param coefs coefficients of the LPC filter
274  * @param data input data
275  * @param cba_idx index of the best entry of the adaptive codebook
276  * @param cb1_idx pointer to variable where the index of the best entry of the
277  * first fixed codebook is returned
278  * @param cb2_idx pointer to variable where the index of the best entry of the
279  * second fixed codebook is returned
280  */
281 static void fixed_cb_search(float *work, const float *coefs, float *data,
282  int cba_idx, int *cb1_idx, int *cb2_idx)
283 {
284  int i, ortho_cb1;
285  float gain;
286  float cba_vect[BLOCKSIZE], cb1_vect[BLOCKSIZE];
287  float vect[BLOCKSIZE];
288 
289  /**
290  * The filtered vector from the adaptive codebook can be retrieved from
291  * work, because this function is called just after adaptive_cb_search().
292  */
293  if (cba_idx)
294  memcpy(cba_vect, work, sizeof(cba_vect));
295 
296  find_best_vect(work, coefs, ff_cb1_vects, cba_idx ? cba_vect : NULL, NULL,
297  data, cb1_idx, &gain);
298 
299  /**
300  * Re-calculate the filtered vector from the vector with maximum match score
301  * and remove its contribution from input data.
302  */
303  if (gain) {
304  for (i = 0; i < BLOCKSIZE; i++)
305  vect[i] = ff_cb1_vects[*cb1_idx][i];
306  ff_celp_lp_synthesis_filterf(work, coefs, vect, BLOCKSIZE, LPC_ORDER);
307  if (cba_idx)
308  orthogonalize(work, cba_vect);
309  for (i = 0; i < BLOCKSIZE; i++)
310  data[i] -= gain * work[i];
311  memcpy(cb1_vect, work, sizeof(cb1_vect));
312  ortho_cb1 = 1;
313  } else
314  ortho_cb1 = 0;
315 
316  find_best_vect(work, coefs, ff_cb2_vects, cba_idx ? cba_vect : NULL,
317  ortho_cb1 ? cb1_vect : NULL, data, cb2_idx, &gain);
318 }
319 
320 
321 /**
322  * Encode a subblock of the current frame
323  *
324  * @param ractx encoder context
325  * @param sblock_data input data of the subblock
326  * @param lpc_coefs coefficients of the LPC filter
327  * @param rms RMS of the reflection coefficients
328  * @param pb pointer to PutBitContext of the current frame
329  */
331  const int16_t *sblock_data,
332  const int16_t *lpc_coefs, unsigned int rms,
333  PutBitContext *pb)
334 {
335  float data[BLOCKSIZE] = { 0 }, work[LPC_ORDER + BLOCKSIZE];
336  float coefs[LPC_ORDER];
337  float zero[BLOCKSIZE], cba[BLOCKSIZE], cb1[BLOCKSIZE], cb2[BLOCKSIZE];
338  int cba_idx, cb1_idx, cb2_idx, gain;
339  int i, n;
340  unsigned m[3];
341  float g[3];
342  float error, best_error;
343 
344  for (i = 0; i < LPC_ORDER; i++) {
345  work[i] = ractx->curr_sblock[BLOCKSIZE + i];
346  coefs[i] = lpc_coefs[i] * (1/4096.0);
347  }
348 
349  /**
350  * Calculate the zero-input response of the LPC filter and subtract it from
351  * input data.
352  */
353  ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, data, BLOCKSIZE,
354  LPC_ORDER);
355  for (i = 0; i < BLOCKSIZE; i++) {
356  zero[i] = work[LPC_ORDER + i];
357  data[i] = sblock_data[i] - zero[i];
358  }
359 
360  /**
361  * Codebook search is performed without taking into account the contribution
362  * of the previous subblock, since it has been just subtracted from input
363  * data.
364  */
365  memset(work, 0, LPC_ORDER * sizeof(*work));
366 
367  cba_idx = adaptive_cb_search(ractx->adapt_cb, work + LPC_ORDER, coefs,
368  data);
369  if (cba_idx) {
370  /**
371  * The filtered vector from the adaptive codebook can be retrieved from
372  * work, see implementation of adaptive_cb_search().
373  */
374  memcpy(cba, work + LPC_ORDER, sizeof(cba));
375 
376  ff_copy_and_dup(ractx->buffer_a, ractx->adapt_cb, cba_idx + BLOCKSIZE / 2 - 1);
377  m[0] = (ff_irms(&ractx->dsp, ractx->buffer_a) * rms) >> 12;
378  }
379  fixed_cb_search(work + LPC_ORDER, coefs, data, cba_idx, &cb1_idx, &cb2_idx);
380  for (i = 0; i < BLOCKSIZE; i++) {
381  cb1[i] = ff_cb1_vects[cb1_idx][i];
382  cb2[i] = ff_cb2_vects[cb2_idx][i];
383  }
384  ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, cb1, BLOCKSIZE,
385  LPC_ORDER);
386  memcpy(cb1, work + LPC_ORDER, sizeof(cb1));
387  m[1] = (ff_cb1_base[cb1_idx] * rms) >> 8;
388  ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, cb2, BLOCKSIZE,
389  LPC_ORDER);
390  memcpy(cb2, work + LPC_ORDER, sizeof(cb2));
391  m[2] = (ff_cb2_base[cb2_idx] * rms) >> 8;
392  best_error = FLT_MAX;
393  gain = 0;
394  for (n = 0; n < 256; n++) {
395  g[1] = ((ff_gain_val_tab[n][1] * m[1]) >> ff_gain_exp_tab[n]) *
396  (1/4096.0);
397  g[2] = ((ff_gain_val_tab[n][2] * m[2]) >> ff_gain_exp_tab[n]) *
398  (1/4096.0);
399  error = 0;
400  if (cba_idx) {
401  g[0] = ((ff_gain_val_tab[n][0] * m[0]) >> ff_gain_exp_tab[n]) *
402  (1/4096.0);
403  for (i = 0; i < BLOCKSIZE; i++) {
404  data[i] = zero[i] + g[0] * cba[i] + g[1] * cb1[i] +
405  g[2] * cb2[i];
406  error += (data[i] - sblock_data[i]) *
407  (data[i] - sblock_data[i]);
408  }
409  } else {
410  for (i = 0; i < BLOCKSIZE; i++) {
411  data[i] = zero[i] + g[1] * cb1[i] + g[2] * cb2[i];
412  error += (data[i] - sblock_data[i]) *
413  (data[i] - sblock_data[i]);
414  }
415  }
416  if (error < best_error) {
417  best_error = error;
418  gain = n;
419  }
420  }
421  put_bits(pb, 7, cba_idx);
422  put_bits(pb, 8, gain);
423  put_bits(pb, 7, cb1_idx);
424  put_bits(pb, 7, cb2_idx);
425  ff_subblock_synthesis(ractx, lpc_coefs, cba_idx, cb1_idx, cb2_idx, rms,
426  gain);
427 }
428 
429 
430 static int ra144_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
431  const AVFrame *frame, int *got_packet_ptr)
432 {
433  static const uint8_t sizes[LPC_ORDER] = {64, 32, 32, 16, 16, 8, 8, 8, 8, 4};
434  static const uint8_t bit_sizes[LPC_ORDER] = {6, 5, 5, 4, 4, 3, 3, 3, 3, 2};
435  RA144Context *ractx = avctx->priv_data;
436  PutBitContext pb;
437  int32_t lpc_data[NBLOCKS * BLOCKSIZE];
438  int32_t lpc_coefs[LPC_ORDER][MAX_LPC_ORDER];
439  int shift[LPC_ORDER];
440  int16_t block_coefs[NBLOCKS][LPC_ORDER];
441  int lpc_refl[LPC_ORDER]; /**< reflection coefficients of the frame */
442  unsigned int refl_rms[NBLOCKS]; /**< RMS of the reflection coefficients */
443  const int16_t *samples = frame ? (const int16_t *)frame->data[0] : NULL;
444  int energy = 0;
445  int i, idx, ret;
446 
447  if (ractx->last_frame)
448  return 0;
449 
450  if ((ret = ff_alloc_packet2(avctx, avpkt, FRAME_SIZE)) < 0)
451  return ret;
452 
453  /**
454  * Since the LPC coefficients are calculated on a frame centered over the
455  * fourth subframe, to encode a given frame, data from the next frame is
456  * needed. In each call to this function, the previous frame (whose data are
457  * saved in the encoder context) is encoded, and data from the current frame
458  * are saved in the encoder context to be used in the next function call.
459  */
460  for (i = 0; i < (2 * BLOCKSIZE + BLOCKSIZE / 2); i++) {
461  lpc_data[i] = ractx->curr_block[BLOCKSIZE + BLOCKSIZE / 2 + i];
462  energy += (lpc_data[i] * lpc_data[i]) >> 4;
463  }
464  if (frame) {
465  int j;
466  for (j = 0; j < frame->nb_samples && i < NBLOCKS * BLOCKSIZE; i++, j++) {
467  lpc_data[i] = samples[j] >> 2;
468  energy += (lpc_data[i] * lpc_data[i]) >> 4;
469  }
470  }
471  if (i < NBLOCKS * BLOCKSIZE)
472  memset(&lpc_data[i], 0, (NBLOCKS * BLOCKSIZE - i) * sizeof(*lpc_data));
473  energy = ff_energy_tab[quantize(ff_t_sqrt(energy >> 5) >> 10, ff_energy_tab,
474  32)];
475 
476  ff_lpc_calc_coefs(&ractx->lpc_ctx, lpc_data, NBLOCKS * BLOCKSIZE, LPC_ORDER,
477  LPC_ORDER, 16, lpc_coefs, shift, FF_LPC_TYPE_LEVINSON,
478  0, ORDER_METHOD_EST, 12, 0);
479  for (i = 0; i < LPC_ORDER; i++)
480  block_coefs[NBLOCKS - 1][i] = -(lpc_coefs[LPC_ORDER - 1][i] <<
481  (12 - shift[LPC_ORDER - 1]));
482 
483  /**
484  * TODO: apply perceptual weighting of the input speech through bandwidth
485  * expansion of the LPC filter.
486  */
487 
488  if (ff_eval_refl(lpc_refl, block_coefs[NBLOCKS - 1], avctx)) {
489  /**
490  * The filter is unstable: use the coefficients of the previous frame.
491  */
492  ff_int_to_int16(block_coefs[NBLOCKS - 1], ractx->lpc_coef[1]);
493  if (ff_eval_refl(lpc_refl, block_coefs[NBLOCKS - 1], avctx)) {
494  /* the filter is still unstable. set reflection coeffs to zero. */
495  memset(lpc_refl, 0, sizeof(lpc_refl));
496  }
497  }
498  init_put_bits(&pb, avpkt->data, avpkt->size);
499  for (i = 0; i < LPC_ORDER; i++) {
500  idx = quantize(lpc_refl[i], ff_lpc_refl_cb[i], sizes[i]);
501  put_bits(&pb, bit_sizes[i], idx);
502  lpc_refl[i] = ff_lpc_refl_cb[i][idx];
503  }
504  ractx->lpc_refl_rms[0] = ff_rms(lpc_refl);
505  ff_eval_coefs(ractx->lpc_coef[0], lpc_refl);
506  refl_rms[0] = ff_interp(ractx, block_coefs[0], 1, 1, ractx->old_energy);
507  refl_rms[1] = ff_interp(ractx, block_coefs[1], 2,
508  energy <= ractx->old_energy,
509  ff_t_sqrt(energy * ractx->old_energy) >> 12);
510  refl_rms[2] = ff_interp(ractx, block_coefs[2], 3, 0, energy);
511  refl_rms[3] = ff_rescale_rms(ractx->lpc_refl_rms[0], energy);
512  ff_int_to_int16(block_coefs[NBLOCKS - 1], ractx->lpc_coef[0]);
513  put_bits(&pb, 5, quantize(energy, ff_energy_tab, 32));
514  for (i = 0; i < NBLOCKS; i++)
515  ra144_encode_subblock(ractx, ractx->curr_block + i * BLOCKSIZE,
516  block_coefs[i], refl_rms[i], &pb);
517  flush_put_bits(&pb);
518  ractx->old_energy = energy;
519  ractx->lpc_refl_rms[1] = ractx->lpc_refl_rms[0];
520  FFSWAP(unsigned int *, ractx->lpc_coef[0], ractx->lpc_coef[1]);
521 
522  /* copy input samples to current block for processing in next call */
523  i = 0;
524  if (frame) {
525  for (; i < frame->nb_samples; i++)
526  ractx->curr_block[i] = samples[i] >> 2;
527 
528  if ((ret = ff_af_queue_add(&ractx->afq, frame)) < 0)
529  return ret;
530  } else
531  ractx->last_frame = 1;
532  memset(&ractx->curr_block[i], 0,
533  (NBLOCKS * BLOCKSIZE - i) * sizeof(*ractx->curr_block));
534 
535  /* Get the next frame pts/duration */
536  ff_af_queue_remove(&ractx->afq, avctx->frame_size, &avpkt->pts,
537  &avpkt->duration);
538 
539  avpkt->size = FRAME_SIZE;
540  *got_packet_ptr = 1;
541  return 0;
542 }
543 
544 
546  .name = "real_144",
547  .long_name = NULL_IF_CONFIG_SMALL("RealAudio 1.0 (14.4K)"),
548  .type = AVMEDIA_TYPE_AUDIO,
549  .id = AV_CODEC_ID_RA_144,
550  .priv_data_size = sizeof(RA144Context),
552  .encode2 = ra144_encode_frame,
555  .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S16,
557  .supported_samplerates = (const int[]){ 8000, 0 },
558  .channel_layouts = (const uint64_t[]) { AV_CH_LAYOUT_MONO, 0 },
559 };