sipr.c
Go to the documentation of this file.
1 /*
2  * SIPR / ACELP.NET decoder
3  *
4  * Copyright (c) 2008 Vladimir Voroshilov
5  * Copyright (c) 2009 Vitor Sessak
6  *
7  * This file is part of Libav.
8  *
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.
13  *
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.
18  *
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
22  */
23 
24 #include <math.h>
25 #include <stdint.h>
26 #include <string.h>
27 
28 #include "libavutil/mathematics.h"
29 #include "avcodec.h"
30 #define BITSTREAM_READER_LE
31 #include "get_bits.h"
32 #include "dsputil.h"
33 
34 #include "lsp.h"
35 #include "celp_math.h"
36 #include "acelp_vectors.h"
37 #include "acelp_pitch_delay.h"
38 #include "acelp_filters.h"
39 #include "celp_filters.h"
40 
41 #define MAX_SUBFRAME_COUNT 5
42 
43 #include "sipr.h"
44 #include "siprdata.h"
45 
46 typedef struct {
47  const char *mode_name;
48  uint16_t bits_per_frame;
49  uint8_t subframe_count;
52 
53  /* bitstream parameters */
56 
58  uint8_t vq_indexes_bits[5];
59 
61  uint8_t pitch_delay_bits[5];
62 
63  uint8_t gp_index_bits;
64  uint8_t fc_index_bits[10];
65  uint8_t gc_index_bits;
67 
68 static const SiprModeParam modes[MODE_COUNT] = {
69  [MODE_16k] = {
70  .mode_name = "16k",
71  .bits_per_frame = 160,
72  .subframe_count = SUBFRAME_COUNT_16k,
73  .frames_per_packet = 1,
74  .pitch_sharp_factor = 0.00,
75 
76  .number_of_fc_indexes = 10,
77  .ma_predictor_bits = 1,
78  .vq_indexes_bits = {7, 8, 7, 7, 7},
79  .pitch_delay_bits = {9, 6},
80  .gp_index_bits = 4,
81  .fc_index_bits = {4, 5, 4, 5, 4, 5, 4, 5, 4, 5},
82  .gc_index_bits = 5
83  },
84 
85  [MODE_8k5] = {
86  .mode_name = "8k5",
87  .bits_per_frame = 152,
88  .subframe_count = 3,
89  .frames_per_packet = 1,
90  .pitch_sharp_factor = 0.8,
91 
92  .number_of_fc_indexes = 3,
93  .ma_predictor_bits = 0,
94  .vq_indexes_bits = {6, 7, 7, 7, 5},
95  .pitch_delay_bits = {8, 5, 5},
96  .gp_index_bits = 0,
97  .fc_index_bits = {9, 9, 9},
98  .gc_index_bits = 7
99  },
100 
101  [MODE_6k5] = {
102  .mode_name = "6k5",
103  .bits_per_frame = 232,
104  .subframe_count = 3,
105  .frames_per_packet = 2,
106  .pitch_sharp_factor = 0.8,
107 
108  .number_of_fc_indexes = 3,
109  .ma_predictor_bits = 0,
110  .vq_indexes_bits = {6, 7, 7, 7, 5},
111  .pitch_delay_bits = {8, 5, 5},
112  .gp_index_bits = 0,
113  .fc_index_bits = {5, 5, 5},
114  .gc_index_bits = 7
115  },
116 
117  [MODE_5k0] = {
118  .mode_name = "5k0",
119  .bits_per_frame = 296,
120  .subframe_count = 5,
121  .frames_per_packet = 2,
122  .pitch_sharp_factor = 0.85,
123 
124  .number_of_fc_indexes = 1,
125  .ma_predictor_bits = 0,
126  .vq_indexes_bits = {6, 7, 7, 7, 5},
127  .pitch_delay_bits = {8, 5, 8, 5, 5},
128  .gp_index_bits = 0,
129  .fc_index_bits = {10},
130  .gc_index_bits = 7
131  }
132 };
133 
134 const float ff_pow_0_5[] = {
135  1.0/(1 << 1), 1.0/(1 << 2), 1.0/(1 << 3), 1.0/(1 << 4),
136  1.0/(1 << 5), 1.0/(1 << 6), 1.0/(1 << 7), 1.0/(1 << 8),
137  1.0/(1 << 9), 1.0/(1 << 10), 1.0/(1 << 11), 1.0/(1 << 12),
138  1.0/(1 << 13), 1.0/(1 << 14), 1.0/(1 << 15), 1.0/(1 << 16)
139 };
140 
141 static void dequant(float *out, const int *idx, const float *cbs[])
142 {
143  int i;
144  int stride = 2;
145  int num_vec = 5;
146 
147  for (i = 0; i < num_vec; i++)
148  memcpy(out + stride*i, cbs[i] + stride*idx[i], stride*sizeof(float));
149 
150 }
151 
152 static void lsf_decode_fp(float *lsfnew, float *lsf_history,
153  const SiprParameters *parm)
154 {
155  int i;
156  float lsf_tmp[LP_FILTER_ORDER];
157 
158  dequant(lsf_tmp, parm->vq_indexes, lsf_codebooks);
159 
160  for (i = 0; i < LP_FILTER_ORDER; i++)
161  lsfnew[i] = lsf_history[i] * 0.33 + lsf_tmp[i] + mean_lsf[i];
162 
163  ff_sort_nearly_sorted_floats(lsfnew, LP_FILTER_ORDER - 1);
164 
165  /* Note that a minimum distance is not enforced between the last value and
166  the previous one, contrary to what is done in ff_acelp_reorder_lsf() */
167  ff_set_min_dist_lsf(lsfnew, LSFQ_DIFF_MIN, LP_FILTER_ORDER - 1);
168  lsfnew[9] = FFMIN(lsfnew[LP_FILTER_ORDER - 1], 1.3 * M_PI);
169 
170  memcpy(lsf_history, lsf_tmp, LP_FILTER_ORDER * sizeof(*lsf_history));
171 
172  for (i = 0; i < LP_FILTER_ORDER - 1; i++)
173  lsfnew[i] = cos(lsfnew[i]);
174  lsfnew[LP_FILTER_ORDER - 1] *= 6.153848 / M_PI;
175 }
176 
178 static void pitch_sharpening(int pitch_lag_int, float beta,
179  float *fixed_vector)
180 {
181  int i;
182 
183  for (i = pitch_lag_int; i < SUBFR_SIZE; i++)
184  fixed_vector[i] += beta * fixed_vector[i - pitch_lag_int];
185 }
186 
193  const SiprModeParam *p)
194 {
195  int i, j;
196 
197  if (p->ma_predictor_bits)
198  parms->ma_pred_switch = get_bits(pgb, p->ma_predictor_bits);
199 
200  for (i = 0; i < 5; i++)
201  parms->vq_indexes[i] = get_bits(pgb, p->vq_indexes_bits[i]);
202 
203  for (i = 0; i < p->subframe_count; i++) {
204  parms->pitch_delay[i] = get_bits(pgb, p->pitch_delay_bits[i]);
205  if (p->gp_index_bits)
206  parms->gp_index[i] = get_bits(pgb, p->gp_index_bits);
207 
208  for (j = 0; j < p->number_of_fc_indexes; j++)
209  parms->fc_indexes[i][j] = get_bits(pgb, p->fc_index_bits[j]);
210 
211  parms->gc_index[i] = get_bits(pgb, p->gc_index_bits);
212  }
213 }
214 
215 static void sipr_decode_lp(float *lsfnew, const float *lsfold, float *Az,
216  int num_subfr)
217 {
218  double lsfint[LP_FILTER_ORDER];
219  int i,j;
220  float t, t0 = 1.0 / num_subfr;
221 
222  t = t0 * 0.5;
223  for (i = 0; i < num_subfr; i++) {
224  for (j = 0; j < LP_FILTER_ORDER; j++)
225  lsfint[j] = lsfold[j] * (1 - t) + t * lsfnew[j];
226 
227  ff_amrwb_lsp2lpc(lsfint, Az, LP_FILTER_ORDER);
228  Az += LP_FILTER_ORDER;
229  t += t0;
230  }
231 }
232 
236 static void eval_ir(const float *Az, int pitch_lag, float *freq,
237  float pitch_sharp_factor)
238 {
239  float tmp1[SUBFR_SIZE+1], tmp2[LP_FILTER_ORDER+1];
240  int i;
241 
242  tmp1[0] = 1.;
243  for (i = 0; i < LP_FILTER_ORDER; i++) {
244  tmp1[i+1] = Az[i] * ff_pow_0_55[i];
245  tmp2[i ] = Az[i] * ff_pow_0_7 [i];
246  }
247  memset(tmp1 + 11, 0, 37 * sizeof(float));
248 
249  ff_celp_lp_synthesis_filterf(freq, tmp2, tmp1, SUBFR_SIZE,
250  LP_FILTER_ORDER);
251 
252  pitch_sharpening(pitch_lag, pitch_sharp_factor, freq);
253 }
254 
258 static void convolute_with_sparse(float *out, const AMRFixed *pulses,
259  const float *shape, int length)
260 {
261  int i, j;
262 
263  memset(out, 0, length*sizeof(float));
264  for (i = 0; i < pulses->n; i++)
265  for (j = pulses->x[i]; j < length; j++)
266  out[j] += pulses->y[i] * shape[j - pulses->x[i]];
267 }
268 
272 static void postfilter_5k0(SiprContext *ctx, const float *lpc, float *samples)
273 {
274  float buf[SUBFR_SIZE + LP_FILTER_ORDER];
275  float *pole_out = buf + LP_FILTER_ORDER;
276  float lpc_n[LP_FILTER_ORDER];
277  float lpc_d[LP_FILTER_ORDER];
278  int i;
279 
280  for (i = 0; i < LP_FILTER_ORDER; i++) {
281  lpc_d[i] = lpc[i] * ff_pow_0_75[i];
282  lpc_n[i] = lpc[i] * ff_pow_0_5 [i];
283  };
284 
285  memcpy(pole_out - LP_FILTER_ORDER, ctx->postfilter_mem,
286  LP_FILTER_ORDER*sizeof(float));
287 
288  ff_celp_lp_synthesis_filterf(pole_out, lpc_d, samples, SUBFR_SIZE,
289  LP_FILTER_ORDER);
290 
291  memcpy(ctx->postfilter_mem, pole_out + SUBFR_SIZE - LP_FILTER_ORDER,
292  LP_FILTER_ORDER*sizeof(float));
293 
294  ff_tilt_compensation(&ctx->tilt_mem, 0.4, pole_out, SUBFR_SIZE);
295 
296  memcpy(pole_out - LP_FILTER_ORDER, ctx->postfilter_mem5k0,
297  LP_FILTER_ORDER*sizeof(*pole_out));
298 
299  memcpy(ctx->postfilter_mem5k0, pole_out + SUBFR_SIZE - LP_FILTER_ORDER,
300  LP_FILTER_ORDER*sizeof(*pole_out));
301 
302  ff_celp_lp_zero_synthesis_filterf(samples, lpc_n, pole_out, SUBFR_SIZE,
303  LP_FILTER_ORDER);
304 
305 }
306 
307 static void decode_fixed_sparse(AMRFixed *fixed_sparse, const int16_t *pulses,
308  SiprMode mode, int low_gain)
309 {
310  int i;
311 
312  switch (mode) {
313  case MODE_6k5:
314  for (i = 0; i < 3; i++) {
315  fixed_sparse->x[i] = 3 * (pulses[i] & 0xf) + i;
316  fixed_sparse->y[i] = pulses[i] & 0x10 ? -1 : 1;
317  }
318  fixed_sparse->n = 3;
319  break;
320  case MODE_8k5:
321  for (i = 0; i < 3; i++) {
322  fixed_sparse->x[2*i ] = 3 * ((pulses[i] >> 4) & 0xf) + i;
323  fixed_sparse->x[2*i + 1] = 3 * ( pulses[i] & 0xf) + i;
324 
325  fixed_sparse->y[2*i ] = (pulses[i] & 0x100) ? -1.0: 1.0;
326 
327  fixed_sparse->y[2*i + 1] =
328  (fixed_sparse->x[2*i + 1] < fixed_sparse->x[2*i]) ?
329  -fixed_sparse->y[2*i ] : fixed_sparse->y[2*i];
330  }
331 
332  fixed_sparse->n = 6;
333  break;
334  case MODE_5k0:
335  default:
336  if (low_gain) {
337  int offset = (pulses[0] & 0x200) ? 2 : 0;
338  int val = pulses[0];
339 
340  for (i = 0; i < 3; i++) {
341  int index = (val & 0x7) * 6 + 4 - i*2;
342 
343  fixed_sparse->y[i] = (offset + index) & 0x3 ? -1 : 1;
344  fixed_sparse->x[i] = index;
345 
346  val >>= 3;
347  }
348  fixed_sparse->n = 3;
349  } else {
350  int pulse_subset = (pulses[0] >> 8) & 1;
351 
352  fixed_sparse->x[0] = ((pulses[0] >> 4) & 15) * 3 + pulse_subset;
353  fixed_sparse->x[1] = ( pulses[0] & 15) * 3 + pulse_subset + 1;
354 
355  fixed_sparse->y[0] = pulses[0] & 0x200 ? -1 : 1;
356  fixed_sparse->y[1] = -fixed_sparse->y[0];
357  fixed_sparse->n = 2;
358  }
359  break;
360  }
361 }
362 
363 static void decode_frame(SiprContext *ctx, SiprParameters *params,
364  float *out_data)
365 {
366  int i, j;
367  int subframe_count = modes[ctx->mode].subframe_count;
368  int frame_size = subframe_count * SUBFR_SIZE;
370  float *excitation;
371  float ir_buf[SUBFR_SIZE + LP_FILTER_ORDER];
372  float lsf_new[LP_FILTER_ORDER];
373  float *impulse_response = ir_buf + LP_FILTER_ORDER;
374  float *synth = ctx->synth_buf + 16; // 16 instead of LP_FILTER_ORDER for
375  // memory alignment
376  int t0_first = 0;
377  AMRFixed fixed_cb;
378 
379  memset(ir_buf, 0, LP_FILTER_ORDER * sizeof(float));
380  lsf_decode_fp(lsf_new, ctx->lsf_history, params);
381 
382  sipr_decode_lp(lsf_new, ctx->lsp_history, Az, subframe_count);
383 
384  memcpy(ctx->lsp_history, lsf_new, LP_FILTER_ORDER * sizeof(float));
385 
386  excitation = ctx->excitation + PITCH_DELAY_MAX + L_INTERPOL;
387 
388  for (i = 0; i < subframe_count; i++) {
389  float *pAz = Az + i*LP_FILTER_ORDER;
390  float fixed_vector[SUBFR_SIZE];
391  int T0,T0_frac;
392  float pitch_gain, gain_code, avg_energy;
393 
394  ff_decode_pitch_lag(&T0, &T0_frac, params->pitch_delay[i], t0_first, i,
395  ctx->mode == MODE_5k0, 6);
396 
397  if (i == 0 || (i == 2 && ctx->mode == MODE_5k0))
398  t0_first = T0;
399 
400  ff_acelp_interpolatef(excitation, excitation - T0 + (T0_frac <= 0),
401  ff_b60_sinc, 6,
402  2 * ((2 + T0_frac)%3 + 1), LP_FILTER_ORDER,
403  SUBFR_SIZE);
404 
405  decode_fixed_sparse(&fixed_cb, params->fc_indexes[i], ctx->mode,
406  ctx->past_pitch_gain < 0.8);
407 
408  eval_ir(pAz, T0, impulse_response, modes[ctx->mode].pitch_sharp_factor);
409 
410  convolute_with_sparse(fixed_vector, &fixed_cb, impulse_response,
411  SUBFR_SIZE);
412 
413  avg_energy =
414  (0.01 + ff_dot_productf(fixed_vector, fixed_vector, SUBFR_SIZE))/
415  SUBFR_SIZE;
416 
417  ctx->past_pitch_gain = pitch_gain = gain_cb[params->gc_index[i]][0];
418 
419  gain_code = ff_amr_set_fixed_gain(gain_cb[params->gc_index[i]][1],
420  avg_energy, ctx->energy_history,
421  34 - 15.0/(0.05*M_LN10/M_LN2),
422  pred);
423 
424  ff_weighted_vector_sumf(excitation, excitation, fixed_vector,
425  pitch_gain, gain_code, SUBFR_SIZE);
426 
427  pitch_gain *= 0.5 * pitch_gain;
428  pitch_gain = FFMIN(pitch_gain, 0.4);
429 
430  ctx->gain_mem = 0.7 * ctx->gain_mem + 0.3 * pitch_gain;
431  ctx->gain_mem = FFMIN(ctx->gain_mem, pitch_gain);
432  gain_code *= ctx->gain_mem;
433 
434  for (j = 0; j < SUBFR_SIZE; j++)
435  fixed_vector[j] = excitation[j] - gain_code * fixed_vector[j];
436 
437  if (ctx->mode == MODE_5k0) {
438  postfilter_5k0(ctx, pAz, fixed_vector);
439 
440  ff_celp_lp_synthesis_filterf(ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i*SUBFR_SIZE,
441  pAz, excitation, SUBFR_SIZE,
442  LP_FILTER_ORDER);
443  }
444 
445  ff_celp_lp_synthesis_filterf(synth + i*SUBFR_SIZE, pAz, fixed_vector,
446  SUBFR_SIZE, LP_FILTER_ORDER);
447 
448  excitation += SUBFR_SIZE;
449  }
450 
451  memcpy(synth - LP_FILTER_ORDER, synth + frame_size - LP_FILTER_ORDER,
452  LP_FILTER_ORDER * sizeof(float));
453 
454  if (ctx->mode == MODE_5k0) {
455  for (i = 0; i < subframe_count; i++) {
456  float energy = ff_dot_productf(ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i*SUBFR_SIZE,
457  ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i*SUBFR_SIZE,
458  SUBFR_SIZE);
459  ff_adaptive_gain_control(&synth[i * SUBFR_SIZE],
460  &synth[i * SUBFR_SIZE], energy,
461  SUBFR_SIZE, 0.9, &ctx->postfilter_agc);
462  }
463 
464  memcpy(ctx->postfilter_syn5k0, ctx->postfilter_syn5k0 + frame_size,
465  LP_FILTER_ORDER*sizeof(float));
466  }
467  memmove(ctx->excitation, excitation - PITCH_DELAY_MAX - L_INTERPOL,
468  (PITCH_DELAY_MAX + L_INTERPOL) * sizeof(float));
469 
471  (const float[2]) {-1.99997 , 1.000000000},
472  (const float[2]) {-1.93307352, 0.935891986},
473  0.939805806,
474  ctx->highpass_filt_mem,
475  frame_size);
476 }
477 
479 {
480  SiprContext *ctx = avctx->priv_data;
481  int i;
482 
483  switch (avctx->block_align) {
484  case 20: ctx->mode = MODE_16k; break;
485  case 19: ctx->mode = MODE_8k5; break;
486  case 29: ctx->mode = MODE_6k5; break;
487  case 37: ctx->mode = MODE_5k0; break;
488  default:
489  if (avctx->bit_rate > 12200) ctx->mode = MODE_16k;
490  else if (avctx->bit_rate > 7500 ) ctx->mode = MODE_8k5;
491  else if (avctx->bit_rate > 5750 ) ctx->mode = MODE_6k5;
492  else ctx->mode = MODE_5k0;
493  av_log(avctx, AV_LOG_WARNING,
494  "Invalid block_align: %d. Mode %s guessed based on bitrate: %d\n",
495  avctx->block_align, modes[ctx->mode].mode_name, avctx->bit_rate);
496  }
497 
498  av_log(avctx, AV_LOG_DEBUG, "Mode: %s\n", modes[ctx->mode].mode_name);
499 
500  if (ctx->mode == MODE_16k) {
501  ff_sipr_init_16k(ctx);
503  } else {
504  ctx->decode_frame = decode_frame;
505  }
506 
507  for (i = 0; i < LP_FILTER_ORDER; i++)
508  ctx->lsp_history[i] = cos((i+1) * M_PI / (LP_FILTER_ORDER + 1));
509 
510  for (i = 0; i < 4; i++)
511  ctx->energy_history[i] = -14;
512 
513  avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
514 
516  avctx->coded_frame = &ctx->frame;
517 
518  return 0;
519 }
520 
521 static int sipr_decode_frame(AVCodecContext *avctx, void *data,
522  int *got_frame_ptr, AVPacket *avpkt)
523 {
524  SiprContext *ctx = avctx->priv_data;
525  const uint8_t *buf=avpkt->data;
526  SiprParameters parm;
527  const SiprModeParam *mode_par = &modes[ctx->mode];
528  GetBitContext gb;
529  float *samples;
530  int subframe_size = ctx->mode == MODE_16k ? L_SUBFR_16k : SUBFR_SIZE;
531  int i, ret;
532 
533  ctx->avctx = avctx;
534  if (avpkt->size < (mode_par->bits_per_frame >> 3)) {
535  av_log(avctx, AV_LOG_ERROR,
536  "Error processing packet: packet size (%d) too small\n",
537  avpkt->size);
538  return -1;
539  }
540 
541  /* get output buffer */
542  ctx->frame.nb_samples = mode_par->frames_per_packet * subframe_size *
543  mode_par->subframe_count;
544  if ((ret = avctx->get_buffer(avctx, &ctx->frame)) < 0) {
545  av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
546  return ret;
547  }
548  samples = (float *)ctx->frame.data[0];
549 
550  init_get_bits(&gb, buf, mode_par->bits_per_frame);
551 
552  for (i = 0; i < mode_par->frames_per_packet; i++) {
553  decode_parameters(&parm, &gb, mode_par);
554 
555  ctx->decode_frame(ctx, &parm, samples);
556 
557  samples += subframe_size * mode_par->subframe_count;
558  }
559 
560  *got_frame_ptr = 1;
561  *(AVFrame *)data = ctx->frame;
562 
563  return mode_par->bits_per_frame >> 3;
564 }
565 
567  .name = "sipr",
568  .type = AVMEDIA_TYPE_AUDIO,
569  .id = CODEC_ID_SIPR,
570  .priv_data_size = sizeof(SiprContext),
573  .capabilities = CODEC_CAP_DR1,
574  .long_name = NULL_IF_CONFIG_SMALL("RealAudio SIPR / ACELP.NET"),
575 };