68b9d2c04a018a6587cb1074c8c9c1ba51c16cbf
[libav.git] / libavcodec / qcelpdec.c
1 /*
2 * QCELP decoder
3 * Copyright (c) 2007 Reynaldo H. Verdejo Pinochet
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 libavcodec/qcelpdec.c
24 * QCELP decoder
25 * @author Reynaldo H. Verdejo Pinochet
26 * @remark FFmpeg merging spearheaded by Kenan Gillet
27 * @remark Development mentored by Benjamin Larson
28 */
29
30 #include <stddef.h>
31
32 #include "avcodec.h"
33 #include "internal.h"
34 #include "get_bits.h"
35
36 #include "qcelpdata.h"
37
38 #include "celp_math.h"
39 #include "celp_filters.h"
40 #include "acelp_vectors.h"
41
42 #undef NDEBUG
43 #include <assert.h>
44
45 typedef enum
46 {
47 I_F_Q = -1, /*!< insufficient frame quality */
48 SILENCE,
49 RATE_OCTAVE,
50 RATE_QUARTER,
51 RATE_HALF,
52 RATE_FULL
53 } qcelp_packet_rate;
54
55 typedef struct
56 {
57 GetBitContext gb;
58 qcelp_packet_rate bitrate;
59 QCELPFrame frame; /*!< unpacked data frame */
60
61 uint8_t erasure_count;
62 uint8_t octave_count; /*!< count the consecutive RATE_OCTAVE frames */
63 float prev_lspf[10];
64 float predictor_lspf[10];/*!< LSP predictor for RATE_OCTAVE and I_F_Q */
65 float pitch_synthesis_filter_mem[303];
66 float pitch_pre_filter_mem[303];
67 float rnd_fir_filter_mem[180];
68 float formant_mem[170];
69 float last_codebook_gain;
70 int prev_g1[2];
71 int prev_bitrate;
72 float pitch_gain[4];
73 uint8_t pitch_lag[4];
74 uint16_t first16bits;
75 uint8_t warned_buf_mismatch_bitrate;
76 } QCELPContext;
77
78 /**
79 * Reconstructs LPC coefficients from the line spectral pair frequencies.
80 *
81 * TIA/EIA/IS-733 2.4.3.3.5
82 */
83 void ff_celp_lspf2lpc(const double *lspf, float *lpc);
84
85 /**
86 * Initialize the speech codec according to the specification.
87 *
88 * TIA/EIA/IS-733 2.4.9
89 */
90 static av_cold int qcelp_decode_init(AVCodecContext *avctx)
91 {
92 QCELPContext *q = avctx->priv_data;
93 int i;
94
95 avctx->sample_fmt = SAMPLE_FMT_FLT;
96
97 for(i=0; i<10; i++)
98 q->prev_lspf[i] = (i+1)/11.;
99
100 return 0;
101 }
102
103 /**
104 * Decodes the 10 quantized LSP frequencies from the LSPV/LSP
105 * transmission codes of any bitrate and checks for badly received packets.
106 *
107 * @param q the context
108 * @param lspf line spectral pair frequencies
109 *
110 * @return 0 on success, -1 if the packet is badly received
111 *
112 * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3
113 */
114 static int decode_lspf(QCELPContext *q, float *lspf)
115 {
116 int i;
117 float tmp_lspf, smooth, erasure_coeff;
118 const float *predictors;
119
120 if(q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q)
121 {
122 predictors = (q->prev_bitrate != RATE_OCTAVE &&
123 q->prev_bitrate != I_F_Q ?
124 q->prev_lspf : q->predictor_lspf);
125
126 if(q->bitrate == RATE_OCTAVE)
127 {
128 q->octave_count++;
129
130 for(i=0; i<10; i++)
131 {
132 q->predictor_lspf[i] =
133 lspf[i] = (q->frame.lspv[i] ? QCELP_LSP_SPREAD_FACTOR
134 : -QCELP_LSP_SPREAD_FACTOR)
135 + predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR
136 + (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR)/11);
137 }
138 smooth = (q->octave_count < 10 ? .875 : 0.1);
139 }else
140 {
141 erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR;
142
143 assert(q->bitrate == I_F_Q);
144
145 if(q->erasure_count > 1)
146 erasure_coeff *= (q->erasure_count < 4 ? 0.9 : 0.7);
147
148 for(i=0; i<10; i++)
149 {
150 q->predictor_lspf[i] =
151 lspf[i] = (i + 1) * ( 1 - erasure_coeff)/11
152 + erasure_coeff * predictors[i];
153 }
154 smooth = 0.125;
155 }
156
157 // Check the stability of the LSP frequencies.
158 lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR);
159 for(i=1; i<10; i++)
160 lspf[i] = FFMAX(lspf[i], (lspf[i-1] + QCELP_LSP_SPREAD_FACTOR));
161
162 lspf[9] = FFMIN(lspf[9], (1.0 - QCELP_LSP_SPREAD_FACTOR));
163 for(i=9; i>0; i--)
164 lspf[i-1] = FFMIN(lspf[i-1], (lspf[i] - QCELP_LSP_SPREAD_FACTOR));
165
166 // Low-pass filter the LSP frequencies.
167 ff_weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0-smooth, 10);
168 }else
169 {
170 q->octave_count = 0;
171
172 tmp_lspf = 0.;
173 for(i=0; i<5 ; i++)
174 {
175 lspf[2*i+0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001;
176 lspf[2*i+1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001;
177 }
178
179 // Check for badly received packets.
180 if(q->bitrate == RATE_QUARTER)
181 {
182 if(lspf[9] <= .70 || lspf[9] >= .97)
183 return -1;
184 for(i=3; i<10; i++)
185 if(fabs(lspf[i] - lspf[i-2]) < .08)
186 return -1;
187 }else
188 {
189 if(lspf[9] <= .66 || lspf[9] >= .985)
190 return -1;
191 for(i=4; i<10; i++)
192 if (fabs(lspf[i] - lspf[i-4]) < .0931)
193 return -1;
194 }
195 }
196 return 0;
197 }
198
199 /**
200 * Converts codebook transmission codes to GAIN and INDEX.
201 *
202 * @param q the context
203 * @param gain array holding the decoded gain
204 *
205 * TIA/EIA/IS-733 2.4.6.2
206 */
207 static void decode_gain_and_index(QCELPContext *q,
208 float *gain) {
209 int i, subframes_count, g1[16];
210 float slope;
211
212 if(q->bitrate >= RATE_QUARTER)
213 {
214 switch(q->bitrate)
215 {
216 case RATE_FULL: subframes_count = 16; break;
217 case RATE_HALF: subframes_count = 4; break;
218 default: subframes_count = 5;
219 }
220 for(i=0; i<subframes_count; i++)
221 {
222 g1[i] = 4 * q->frame.cbgain[i];
223 if(q->bitrate == RATE_FULL && !((i+1) & 3))
224 {
225 g1[i] += av_clip((g1[i-1] + g1[i-2] + g1[i-3]) / 3 - 6, 0, 32);
226 }
227
228 gain[i] = qcelp_g12ga[g1[i]];
229
230 if(q->frame.cbsign[i])
231 {
232 gain[i] = -gain[i];
233 q->frame.cindex[i] = (q->frame.cindex[i]-89) & 127;
234 }
235 }
236
237 q->prev_g1[0] = g1[i-2];
238 q->prev_g1[1] = g1[i-1];
239 q->last_codebook_gain = qcelp_g12ga[g1[i-1]];
240
241 if(q->bitrate == RATE_QUARTER)
242 {
243 // Provide smoothing of the unvoiced excitation energy.
244 gain[7] = gain[4];
245 gain[6] = 0.4*gain[3] + 0.6*gain[4];
246 gain[5] = gain[3];
247 gain[4] = 0.8*gain[2] + 0.2*gain[3];
248 gain[3] = 0.2*gain[1] + 0.8*gain[2];
249 gain[2] = gain[1];
250 gain[1] = 0.6*gain[0] + 0.4*gain[1];
251 }
252 }else if (q->bitrate != SILENCE)
253 {
254 if(q->bitrate == RATE_OCTAVE)
255 {
256 g1[0] = 2 * q->frame.cbgain[0]
257 + av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54);
258 subframes_count = 8;
259 }else
260 {
261 assert(q->bitrate == I_F_Q);
262
263 g1[0] = q->prev_g1[1];
264 switch(q->erasure_count)
265 {
266 case 1 : break;
267 case 2 : g1[0] -= 1; break;
268 case 3 : g1[0] -= 2; break;
269 default: g1[0] -= 6;
270 }
271 if(g1[0] < 0)
272 g1[0] = 0;
273 subframes_count = 4;
274 }
275 // This interpolation is done to produce smoother background noise.
276 slope = 0.5*(qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count;
277 for(i=1; i<=subframes_count; i++)
278 gain[i-1] = q->last_codebook_gain + slope * i;
279
280 q->last_codebook_gain = gain[i-2];
281 q->prev_g1[0] = q->prev_g1[1];
282 q->prev_g1[1] = g1[0];
283 }
284 }
285
286 /**
287 * If the received packet is Rate 1/4 a further sanity check is made of the
288 * codebook gain.
289 *
290 * @param cbgain the unpacked cbgain array
291 * @return -1 if the sanity check fails, 0 otherwise
292 *
293 * TIA/EIA/IS-733 2.4.8.7.3
294 */
295 static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)
296 {
297 int i, diff, prev_diff=0;
298
299 for(i=1; i<5; i++)
300 {
301 diff = cbgain[i] - cbgain[i-1];
302 if(FFABS(diff) > 10)
303 return -1;
304 else if(FFABS(diff - prev_diff) > 12)
305 return -1;
306 prev_diff = diff;
307 }
308 return 0;
309 }
310
311 /**
312 * Computes the scaled codebook vector Cdn From INDEX and GAIN
313 * for all rates.
314 *
315 * The specification lacks some information here.
316 *
317 * TIA/EIA/IS-733 has an omission on the codebook index determination
318 * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says
319 * you have to subtract the decoded index parameter from the given scaled
320 * codebook vector index 'n' to get the desired circular codebook index, but
321 * it does not mention that you have to clamp 'n' to [0-9] in order to get
322 * RI-compliant results.
323 *
324 * The reason for this mistake seems to be the fact they forgot to mention you
325 * have to do these calculations per codebook subframe and adjust given
326 * equation values accordingly.
327 *
328 * @param q the context
329 * @param gain array holding the 4 pitch subframe gain values
330 * @param cdn_vector array for the generated scaled codebook vector
331 */
332 static void compute_svector(QCELPContext *q, const float *gain,
333 float *cdn_vector)
334 {
335 int i, j, k;
336 uint16_t cbseed, cindex;
337 float *rnd, tmp_gain, fir_filter_value;
338
339 switch(q->bitrate)
340 {
341 case RATE_FULL:
342 for(i=0; i<16; i++)
343 {
344 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
345 cindex = -q->frame.cindex[i];
346 for(j=0; j<10; j++)
347 *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cindex++ & 127];
348 }
349 break;
350 case RATE_HALF:
351 for(i=0; i<4; i++)
352 {
353 tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO;
354 cindex = -q->frame.cindex[i];
355 for (j = 0; j < 40; j++)
356 *cdn_vector++ = tmp_gain * qcelp_rate_half_codebook[cindex++ & 127];
357 }
358 break;
359 case RATE_QUARTER:
360 cbseed = (0x0003 & q->frame.lspv[4])<<14 |
361 (0x003F & q->frame.lspv[3])<< 8 |
362 (0x0060 & q->frame.lspv[2])<< 1 |
363 (0x0007 & q->frame.lspv[1])<< 3 |
364 (0x0038 & q->frame.lspv[0])>> 3 ;
365 rnd = q->rnd_fir_filter_mem + 20;
366 for(i=0; i<8; i++)
367 {
368 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
369 for(k=0; k<20; k++)
370 {
371 cbseed = 521 * cbseed + 259;
372 *rnd = (int16_t)cbseed;
373
374 // FIR filter
375 fir_filter_value = 0.0;
376 for(j=0; j<10; j++)
377 fir_filter_value += qcelp_rnd_fir_coefs[j ]
378 * (rnd[-j ] + rnd[-20+j]);
379
380 fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10];
381 *cdn_vector++ = tmp_gain * fir_filter_value;
382 rnd++;
383 }
384 }
385 memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160, 20 * sizeof(float));
386 break;
387 case RATE_OCTAVE:
388 cbseed = q->first16bits;
389 for(i=0; i<8; i++)
390 {
391 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
392 for(j=0; j<20; j++)
393 {
394 cbseed = 521 * cbseed + 259;
395 *cdn_vector++ = tmp_gain * (int16_t)cbseed;
396 }
397 }
398 break;
399 case I_F_Q:
400 cbseed = -44; // random codebook index
401 for(i=0; i<4; i++)
402 {
403 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
404 for(j=0; j<40; j++)
405 *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cbseed++ & 127];
406 }
407 break;
408 case SILENCE:
409 memset(cdn_vector, 0, 160 * sizeof(float));
410 break;
411 }
412 }
413
414 /**
415 * Compute the gain control
416 *
417 * @param v_in gain-controlled vector
418 * @param v_ref vector to control gain of
419 *
420 * @return gain control
421 *
422 * FIXME: If v_ref is a zero vector, it energy is zero
423 * and the behavior of the gain control is
424 * undefined in the specs.
425 *
426 * TIA/EIA/IS-733 2.4.8.3-2/3/4/5, 2.4.8.6
427 */
428 static float compute_gain_ctrl(const float *v_ref, const float *v_in, const int len)
429 {
430 float scalefactor = ff_dot_productf(v_in, v_in, len);
431
432 if(scalefactor)
433 scalefactor = sqrt(ff_dot_productf(v_ref, v_ref, len) / scalefactor);
434 else
435 ff_log_missing_feature(NULL, "Zero energy for gain control", 1);
436 return scalefactor;
437 }
438
439 /**
440 * Apply generic gain control.
441 *
442 * @param v_out output vector
443 * @param v_in gain-controlled vector
444 * @param v_ref vector to control gain of
445 *
446 * TIA/EIA/IS-733 2.4.8.3, 2.4.8.6
447 */
448 static void apply_gain_ctrl(float *v_out, const float *v_ref,
449 const float *v_in)
450 {
451 int i, j, len;
452 float scalefactor;
453
454 for(i=0, j=0; i<4; i++)
455 {
456 scalefactor = compute_gain_ctrl(v_ref + j, v_in + j, 40);
457 for(len=j+40; j<len; j++)
458 v_out[j] = scalefactor * v_in[j];
459 }
460 }
461
462 /**
463 * Apply filter in pitch-subframe steps.
464 *
465 * @param memory buffer for the previous state of the filter
466 * - must be able to contain 303 elements
467 * - the 143 first elements are from the previous state
468 * - the next 160 are for output
469 * @param v_in input filter vector
470 * @param gain per-subframe gain array, each element is between 0.0 and 2.0
471 * @param lag per-subframe lag array, each element is
472 * - between 16 and 143 if its corresponding pfrac is 0,
473 * - between 16 and 139 otherwise
474 * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0
475 * otherwise
476 *
477 * @return filter output vector
478 */
479 static const float *do_pitchfilter(float memory[303], const float v_in[160],
480 const float gain[4], const uint8_t *lag,
481 const uint8_t pfrac[4])
482 {
483 int i, j;
484 float *v_lag, *v_out;
485 const float *v_len;
486
487 v_out = memory + 143; // Output vector starts at memory[143].
488
489 for(i=0; i<4; i++)
490 {
491 if(gain[i])
492 {
493 v_lag = memory + 143 + 40 * i - lag[i];
494 for(v_len=v_in+40; v_in<v_len; v_in++)
495 {
496 if(pfrac[i]) // If it is a fractional lag...
497 {
498 for(j=0, *v_out=0.; j<4; j++)
499 *v_out += qcelp_hammsinc_table[j] * (v_lag[j-4] + v_lag[3-j]);
500 }else
501 *v_out = *v_lag;
502
503 *v_out = *v_in + gain[i] * *v_out;
504
505 v_lag++;
506 v_out++;
507 }
508 }else
509 {
510 memcpy(v_out, v_in, 40 * sizeof(float));
511 v_in += 40;
512 v_out += 40;
513 }
514 }
515
516 memmove(memory, memory + 160, 143 * sizeof(float));
517 return memory + 143;
518 }
519
520 /**
521 * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector.
522 * TIA/EIA/IS-733 2.4.5.2, 2.4.8.7.2
523 *
524 * @param q the context
525 * @param cdn_vector the scaled codebook vector
526 */
527 static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
528 {
529 int i;
530 const float *v_synthesis_filtered, *v_pre_filtered;
531
532 if(q->bitrate >= RATE_HALF ||
533 q->bitrate == SILENCE ||
534 (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF)))
535 {
536
537 if(q->bitrate >= RATE_HALF)
538 {
539
540 // Compute gain & lag for the whole frame.
541 for(i=0; i<4; i++)
542 {
543 q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0;
544
545 q->pitch_lag[i] = q->frame.plag[i] + 16;
546 }
547 }else
548 {
549 float max_pitch_gain;
550
551 if (q->bitrate == I_F_Q)
552 {
553 if (q->erasure_count < 3)
554 max_pitch_gain = 0.9 - 0.3 * (q->erasure_count - 1);
555 else
556 max_pitch_gain = 0.0;
557 }else
558 {
559 assert(q->bitrate == SILENCE);
560 max_pitch_gain = 1.0;
561 }
562 for(i=0; i<4; i++)
563 q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain);
564
565 memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac));
566 }
567
568 // pitch synthesis filter
569 v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem,
570 cdn_vector, q->pitch_gain,
571 q->pitch_lag, q->frame.pfrac);
572
573 // pitch prefilter update
574 for(i=0; i<4; i++)
575 q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0);
576
577 v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem,
578 v_synthesis_filtered,
579 q->pitch_gain, q->pitch_lag,
580 q->frame.pfrac);
581
582 apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered);
583 }else
584 {
585 memcpy(q->pitch_synthesis_filter_mem, cdn_vector + 17,
586 143 * sizeof(float));
587 memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float));
588 memset(q->pitch_gain, 0, sizeof(q->pitch_gain));
589 memset(q->pitch_lag, 0, sizeof(q->pitch_lag));
590 }
591 }
592
593 /**
594 * Reconstructs LPC coefficients from the line spectral pair frequencies
595 * and performs bandwidth expansion.
596 *
597 * @param lspf line spectral pair frequencies
598 * @param lpc linear predictive coding coefficients
599 *
600 * @note: bandwith_expansion_coeff could be precalculated into a table
601 * but it seems to be slower on x86
602 *
603 * TIA/EIA/IS-733 2.4.3.3.5
604 */
605 static void lspf2lpc(const float *lspf, float *lpc)
606 {
607 double lsf[10];
608 double bandwith_expansion_coeff = QCELP_BANDWITH_EXPANSION_COEFF;
609 int i;
610
611 for (i=0; i<10; i++)
612 lsf[i] = cos(M_PI * lspf[i]);
613
614 ff_celp_lspf2lpc(lsf, lpc);
615
616 for (i=0; i<10; i++)
617 {
618 lpc[i] *= bandwith_expansion_coeff;
619 bandwith_expansion_coeff *= QCELP_BANDWITH_EXPANSION_COEFF;
620 }
621 }
622
623 /**
624 * Interpolates LSP frequencies and computes LPC coefficients
625 * for a given bitrate & pitch subframe.
626 *
627 * TIA/EIA/IS-733 2.4.3.3.4, 2.4.8.7.2
628 *
629 * @param q the context
630 * @param curr_lspf LSP frequencies vector of the current frame
631 * @param lpc float vector for the resulting LPC
632 * @param subframe_num frame number in decoded stream
633 */
634 void interpolate_lpc(QCELPContext *q, const float *curr_lspf, float *lpc,
635 const int subframe_num)
636 {
637 float interpolated_lspf[10];
638 float weight;
639
640 if(q->bitrate >= RATE_QUARTER)
641 weight = 0.25 * (subframe_num + 1);
642 else if(q->bitrate == RATE_OCTAVE && !subframe_num)
643 weight = 0.625;
644 else
645 weight = 1.0;
646
647 if(weight != 1.0)
648 {
649 ff_weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
650 weight, 1.0 - weight, 10);
651 lspf2lpc(interpolated_lspf, lpc);
652 }else if(q->bitrate >= RATE_QUARTER ||
653 (q->bitrate == I_F_Q && !subframe_num))
654 lspf2lpc(curr_lspf, lpc);
655 else if(q->bitrate == SILENCE && !subframe_num)
656 lspf2lpc(q->prev_lspf, lpc);
657 }
658
659 static qcelp_packet_rate buf_size2bitrate(const int buf_size)
660 {
661 switch(buf_size)
662 {
663 case 35: return RATE_FULL;
664 case 17: return RATE_HALF;
665 case 8: return RATE_QUARTER;
666 case 4: return RATE_OCTAVE;
667 case 1: return SILENCE;
668 }
669
670 return I_F_Q;
671 }
672
673 /**
674 * Determine the bitrate from the frame size and/or the first byte of the frame.
675 *
676 * @param avctx the AV codec context
677 * @param buf_size length of the buffer
678 * @param buf the bufffer
679 *
680 * @return the bitrate on success,
681 * I_F_Q if the bitrate cannot be satisfactorily determined
682 *
683 * TIA/EIA/IS-733 2.4.8.7.1
684 */
685 static qcelp_packet_rate determine_bitrate(AVCodecContext *avctx, const int buf_size,
686 const uint8_t **buf)
687 {
688 qcelp_packet_rate bitrate;
689
690 if((bitrate = buf_size2bitrate(buf_size)) >= 0)
691 {
692 if(bitrate > **buf)
693 {
694 QCELPContext *q = avctx->priv_data;
695 if (!q->warned_buf_mismatch_bitrate)
696 {
697 av_log(avctx, AV_LOG_WARNING,
698 "Claimed bitrate and buffer size mismatch.\n");
699 q->warned_buf_mismatch_bitrate = 1;
700 }
701 bitrate = **buf;
702 }else if(bitrate < **buf)
703 {
704 av_log(avctx, AV_LOG_ERROR,
705 "Buffer is too small for the claimed bitrate.\n");
706 return I_F_Q;
707 }
708 (*buf)++;
709 }else if((bitrate = buf_size2bitrate(buf_size + 1)) >= 0)
710 {
711 av_log(avctx, AV_LOG_WARNING,
712 "Bitrate byte is missing, guessing the bitrate from packet size.\n");
713 }else
714 return I_F_Q;
715
716 if(bitrate == SILENCE)
717 {
718 //FIXME: Remove experimental warning when tested with samples.
719 ff_log_ask_for_sample(avctx, "'Blank frame handling is experimental.");
720 }
721 return bitrate;
722 }
723
724 static void warn_insufficient_frame_quality(AVCodecContext *avctx,
725 const char *message)
726 {
727 av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n", avctx->frame_number,
728 message);
729 }
730
731 static int qcelp_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
732 AVPacket *avpkt)
733 {
734 const uint8_t *buf = avpkt->data;
735 int buf_size = avpkt->size;
736 QCELPContext *q = avctx->priv_data;
737 float *outbuffer = data;
738 int i;
739 float quantized_lspf[10], lpc[10];
740 float gain[16];
741 float *formant_mem;
742
743 if((q->bitrate = determine_bitrate(avctx, buf_size, &buf)) == I_F_Q)
744 {
745 warn_insufficient_frame_quality(avctx, "bitrate cannot be determined.");
746 goto erasure;
747 }
748
749 if(q->bitrate == RATE_OCTAVE &&
750 (q->first16bits = AV_RB16(buf)) == 0xFFFF)
751 {
752 warn_insufficient_frame_quality(avctx, "Bitrate is 1/8 and first 16 bits are on.");
753 goto erasure;
754 }
755
756 if(q->bitrate > SILENCE)
757 {
758 const QCELPBitmap *bitmaps = qcelp_unpacking_bitmaps_per_rate[q->bitrate];
759 const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate]
760 + qcelp_unpacking_bitmaps_lengths[q->bitrate];
761 uint8_t *unpacked_data = (uint8_t *)&q->frame;
762
763 init_get_bits(&q->gb, buf, 8*buf_size);
764
765 memset(&q->frame, 0, sizeof(QCELPFrame));
766
767 for(; bitmaps < bitmaps_end; bitmaps++)
768 unpacked_data[bitmaps->index] |= get_bits(&q->gb, bitmaps->bitlen) << bitmaps->bitpos;
769
770 // Check for erasures/blanks on rates 1, 1/4 and 1/8.
771 if(q->frame.reserved)
772 {
773 warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area.");
774 goto erasure;
775 }
776 if(q->bitrate == RATE_QUARTER &&
777 codebook_sanity_check_for_rate_quarter(q->frame.cbgain))
778 {
779 warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed.");
780 goto erasure;
781 }
782
783 if(q->bitrate >= RATE_HALF)
784 {
785 for(i=0; i<4; i++)
786 {
787 if(q->frame.pfrac[i] && q->frame.plag[i] >= 124)
788 {
789 warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter.");
790 goto erasure;
791 }
792 }
793 }
794 }
795
796 decode_gain_and_index(q, gain);
797 compute_svector(q, gain, outbuffer);
798
799 if(decode_lspf(q, quantized_lspf) < 0)
800 {
801 warn_insufficient_frame_quality(avctx, "Badly received packets in frame.");
802 goto erasure;
803 }
804
805
806 apply_pitch_filters(q, outbuffer);
807
808 if(q->bitrate == I_F_Q)
809 {
810 erasure:
811 q->bitrate = I_F_Q;
812 q->erasure_count++;
813 decode_gain_and_index(q, gain);
814 compute_svector(q, gain, outbuffer);
815 decode_lspf(q, quantized_lspf);
816 apply_pitch_filters(q, outbuffer);
817 }else
818 q->erasure_count = 0;
819
820 formant_mem = q->formant_mem + 10;
821 for(i=0; i<4; i++)
822 {
823 interpolate_lpc(q, quantized_lspf, lpc, i);
824 ff_celp_lp_synthesis_filterf(formant_mem, lpc, outbuffer + i * 40, 40,
825 10);
826 formant_mem += 40;
827 }
828 memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float));
829
830 // FIXME: postfilter and final gain control should be here.
831 // TIA/EIA/IS-733 2.4.8.6
832
833 formant_mem = q->formant_mem + 10;
834 for(i=0; i<160; i++)
835 *outbuffer++ = av_clipf(*formant_mem++, QCELP_CLIP_LOWER_BOUND,
836 QCELP_CLIP_UPPER_BOUND);
837
838 memcpy(q->prev_lspf, quantized_lspf, sizeof(q->prev_lspf));
839 q->prev_bitrate = q->bitrate;
840
841 *data_size = 160 * sizeof(*outbuffer);
842
843 return *data_size;
844 }
845
846 AVCodec qcelp_decoder =
847 {
848 .name = "qcelp",
849 .type = CODEC_TYPE_AUDIO,
850 .id = CODEC_ID_QCELP,
851 .init = qcelp_decode_init,
852 .decode = qcelp_decode_frame,
853 .priv_data_size = sizeof(QCELPContext),
854 .long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"),
855 };