1 /*
2 * G.723.1 compatible decoder
3 * Copyright (c) 2006 Benjamin Larsson
4 * Copyright (c) 2010 Mohamed Naufal Basheer
5 *
6 * This file is part of FFmpeg.
7 *
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 */
22
23 /**
24 * @file
25 * G.723.1 compatible decoder
26 */
27
28 #include "libavutil/channel_layout.h"
29 #include "libavutil/mem.h"
30 #include "libavutil/opt.h"
31
32 #define BITSTREAM_READER_LE
33 #include "acelp_vectors.h"
34 #include "avcodec.h"
35 #include "celp_filters.h"
36 #include "celp_math.h"
37 #include "get_bits.h"
38 #include "internal.h"
39 #include "g723_1.h"
40
41 #define CNG_RANDOM_SEED 12345
42
43 /**
44 * Postfilter gain weighting factors scaled by 2^15
45 */
46 static const int16_t ppf_gain_weight[2] = {0x1800, 0x2000};
47
48 static const int16_t pitch_contrib[340] = {
49 60, 0, 0, 2489, 60, 0, 0, 5217,
50 1, 6171, 0, 3953, 0, 10364, 1, 9357,
51 -1, 8843, 1, 9396, 0, 5794, -1, 10816,
52 2, 11606, -2, 12072, 0, 8616, 1, 12170,
53 0, 14440, 0, 7787, -1, 13721, 0, 18205,
54 0, 14471, 0, 15807, 1, 15275, 0, 13480,
55 -1, 18375, -1, 0, 1, 11194, -1, 13010,
56 1, 18836, -2, 20354, 1, 16233, -1, 0,
57 60, 0, 0, 12130, 0, 13385, 1, 17834,
58 1, 20875, 0, 21996, 1, 0, 1, 18277,
59 -1, 21321, 1, 13738, -1, 19094, -1, 20387,
60 -1, 0, 0, 21008, 60, 0, -2, 22807,
61 0, 15900, 1, 0, 0, 17989, -1, 22259,
62 1, 24395, 1, 23138, 0, 23948, 1, 22997,
63 2, 22604, -1, 25942, 0, 26246, 1, 25321,
64 0, 26423, 0, 24061, 0, 27247, 60, 0,
65 -1, 25572, 1, 23918, 1, 25930, 2, 26408,
66 -1, 19049, 1, 27357, -1, 24538, 60, 0,
67 -1, 25093, 0, 28549, 1, 0, 0, 22793,
68 -1, 25659, 0, 29377, 0, 30276, 0, 26198,
69 1, 22521, -1, 28919, 0, 27384, 1, 30162,
70 -1, 0, 0, 24237, -1, 30062, 0, 21763,
71 1, 30917, 60, 0, 0, 31284, 0, 29433,
72 1, 26821, 1, 28655, 0, 31327, 2, 30799,
73 1, 31389, 0, 32322, 1, 31760, -2, 31830,
74 0, 26936, -1, 31180, 1, 30875, 0, 27873,
75 -1, 30429, 1, 31050, 0, 0, 0, 31912,
76 1, 31611, 0, 31565, 0, 25557, 0, 31357,
77 60, 0, 1, 29536, 1, 28985, -1, 26984,
78 -1, 31587, 2, 30836, -2, 31133, 0, 30243,
79 -1, 30742, -1, 32090, 60, 0, 2, 30902,
80 60, 0, 0, 30027, 0, 29042, 60, 0,
81 0, 31756, 0, 24553, 0, 25636, -2, 30501,
82 60, 0, -1, 29617, 0, 30649, 60, 0,
83 0, 29274, 2, 30415, 0, 27480, 0, 31213,
84 -1, 28147, 0, 30600, 1, 31652, 2, 29068,
85 60, 0, 1, 28571, 1, 28730, 1, 31422,
86 0, 28257, 0, 24797, 60, 0, 0, 0,
87 60, 0, 0, 22105, 0, 27852, 60, 0,
88 60, 0, -1, 24214, 0, 24642, 0, 23305,
89 60, 0, 60, 0, 1, 22883, 0, 21601,
90 60, 0, 2, 25650, 60, 0, -2, 31253,
91 -2, 25144, 0, 17998
92 };
93
94 /**
95 * Size of the MP-MLQ fixed excitation codebooks
96 */
97 static const int32_t max_pos[4] = {593775, 142506, 593775, 142506};
98
99 /**
100 * 0.65^i (Zero part) and 0.75^i (Pole part) scaled by 2^15
101 */
102 static const int16_t postfilter_tbl[2][LPC_ORDER] = {
103 /* Zero */
104 {21299, 13844, 8999, 5849, 3802, 2471, 1606, 1044, 679, 441},
105 /* Pole */
106 {24576, 18432, 13824, 10368, 7776, 5832, 4374, 3281, 2460, 1845}
107 };
108
109 static const int cng_adaptive_cb_lag[4] = { 1, 0, 1, 3 };
110
111 static const int cng_filt[4] = { 273, 998, 499, 333 };
112
113 static const int cng_bseg[3] = { 2048, 18432, 231233 };
114
g723_1_decode_init(AVCodecContext * avctx)115 static av_cold int g723_1_decode_init(AVCodecContext *avctx)
116 {
117 G723_1_Context *s = avctx->priv_data;
118
119 avctx->sample_fmt = AV_SAMPLE_FMT_S16P;
120 if (avctx->channels < 1 || avctx->channels > 2) {
121 av_log(avctx, AV_LOG_ERROR, "Only mono and stereo are supported (requested channels: %d).\n", avctx->channels);
122 return AVERROR(EINVAL);
123 }
124 avctx->channel_layout = avctx->channels == 1 ? AV_CH_LAYOUT_MONO : AV_CH_LAYOUT_STEREO;
125 for (int ch = 0; ch < avctx->channels; ch++) {
126 G723_1_ChannelContext *p = &s->ch[ch];
127
128 p->pf_gain = 1 << 12;
129
130 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
131 memcpy(p->sid_lsp, dc_lsp, LPC_ORDER * sizeof(*p->sid_lsp));
132
133 p->cng_random_seed = CNG_RANDOM_SEED;
134 p->past_frame_type = SID_FRAME;
135 }
136
137 return 0;
138 }
139
140 /**
141 * Unpack the frame into parameters.
142 *
143 * @param p the context
144 * @param buf pointer to the input buffer
145 * @param buf_size size of the input buffer
146 */
unpack_bitstream(G723_1_ChannelContext * p,const uint8_t * buf,int buf_size)147 static int unpack_bitstream(G723_1_ChannelContext *p, const uint8_t *buf,
148 int buf_size)
149 {
150 GetBitContext gb;
151 int ad_cb_len;
152 int temp, info_bits, i;
153 int ret;
154
155 ret = init_get_bits8(&gb, buf, buf_size);
156 if (ret < 0)
157 return ret;
158
159 /* Extract frame type and rate info */
160 info_bits = get_bits(&gb, 2);
161
162 if (info_bits == 3) {
163 p->cur_frame_type = UNTRANSMITTED_FRAME;
164 return 0;
165 }
166
167 /* Extract 24 bit lsp indices, 8 bit for each band */
168 p->lsp_index[2] = get_bits(&gb, 8);
169 p->lsp_index[1] = get_bits(&gb, 8);
170 p->lsp_index[0] = get_bits(&gb, 8);
171
172 if (info_bits == 2) {
173 p->cur_frame_type = SID_FRAME;
174 p->subframe[0].amp_index = get_bits(&gb, 6);
175 return 0;
176 }
177
178 /* Extract the info common to both rates */
179 p->cur_rate = info_bits ? RATE_5300 : RATE_6300;
180 p->cur_frame_type = ACTIVE_FRAME;
181
182 p->pitch_lag[0] = get_bits(&gb, 7);
183 if (p->pitch_lag[0] > 123) /* test if forbidden code */
184 return -1;
185 p->pitch_lag[0] += PITCH_MIN;
186 p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
187
188 p->pitch_lag[1] = get_bits(&gb, 7);
189 if (p->pitch_lag[1] > 123)
190 return -1;
191 p->pitch_lag[1] += PITCH_MIN;
192 p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
193 p->subframe[0].ad_cb_lag = 1;
194 p->subframe[2].ad_cb_lag = 1;
195
196 for (i = 0; i < SUBFRAMES; i++) {
197 /* Extract combined gain */
198 temp = get_bits(&gb, 12);
199 ad_cb_len = 170;
200 p->subframe[i].dirac_train = 0;
201 if (p->cur_rate == RATE_6300 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
202 p->subframe[i].dirac_train = temp >> 11;
203 temp &= 0x7FF;
204 ad_cb_len = 85;
205 }
206 p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
207 if (p->subframe[i].ad_cb_gain < ad_cb_len) {
208 p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
209 GAIN_LEVELS;
210 } else {
211 return -1;
212 }
213 }
214
215 p->subframe[0].grid_index = get_bits1(&gb);
216 p->subframe[1].grid_index = get_bits1(&gb);
217 p->subframe[2].grid_index = get_bits1(&gb);
218 p->subframe[3].grid_index = get_bits1(&gb);
219
220 if (p->cur_rate == RATE_6300) {
221 skip_bits1(&gb); /* skip reserved bit */
222
223 /* Compute pulse_pos index using the 13-bit combined position index */
224 temp = get_bits(&gb, 13);
225 p->subframe[0].pulse_pos = temp / 810;
226
227 temp -= p->subframe[0].pulse_pos * 810;
228 p->subframe[1].pulse_pos = FASTDIV(temp, 90);
229
230 temp -= p->subframe[1].pulse_pos * 90;
231 p->subframe[2].pulse_pos = FASTDIV(temp, 9);
232 p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
233
234 p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
235 get_bits(&gb, 16);
236 p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
237 get_bits(&gb, 14);
238 p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
239 get_bits(&gb, 16);
240 p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
241 get_bits(&gb, 14);
242
243 p->subframe[0].pulse_sign = get_bits(&gb, 6);
244 p->subframe[1].pulse_sign = get_bits(&gb, 5);
245 p->subframe[2].pulse_sign = get_bits(&gb, 6);
246 p->subframe[3].pulse_sign = get_bits(&gb, 5);
247 } else { /* 5300 bps */
248 p->subframe[0].pulse_pos = get_bits(&gb, 12);
249 p->subframe[1].pulse_pos = get_bits(&gb, 12);
250 p->subframe[2].pulse_pos = get_bits(&gb, 12);
251 p->subframe[3].pulse_pos = get_bits(&gb, 12);
252
253 p->subframe[0].pulse_sign = get_bits(&gb, 4);
254 p->subframe[1].pulse_sign = get_bits(&gb, 4);
255 p->subframe[2].pulse_sign = get_bits(&gb, 4);
256 p->subframe[3].pulse_sign = get_bits(&gb, 4);
257 }
258
259 return 0;
260 }
261
262 /**
263 * Bitexact implementation of sqrt(val/2).
264 */
square_root(unsigned val)265 static int16_t square_root(unsigned val)
266 {
267 av_assert2(!(val & 0x80000000));
268
269 return (ff_sqrt(val << 1) >> 1) & (~1);
270 }
271
272 /**
273 * Generate fixed codebook excitation vector.
274 *
275 * @param vector decoded excitation vector
276 * @param subfrm current subframe
277 * @param cur_rate current bitrate
278 * @param pitch_lag closed loop pitch lag
279 * @param index current subframe index
280 */
gen_fcb_excitation(int16_t * vector,G723_1_Subframe * subfrm,enum Rate cur_rate,int pitch_lag,int index)281 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe *subfrm,
282 enum Rate cur_rate, int pitch_lag, int index)
283 {
284 int temp, i, j;
285
286 memset(vector, 0, SUBFRAME_LEN * sizeof(*vector));
287
288 if (cur_rate == RATE_6300) {
289 if (subfrm->pulse_pos >= max_pos[index])
290 return;
291
292 /* Decode amplitudes and positions */
293 j = PULSE_MAX - pulses[index];
294 temp = subfrm->pulse_pos;
295 for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
296 temp -= ff_g723_1_combinatorial_table[j][i];
297 if (temp >= 0)
298 continue;
299 temp += ff_g723_1_combinatorial_table[j++][i];
300 if (subfrm->pulse_sign & (1 << (PULSE_MAX - j))) {
301 vector[subfrm->grid_index + GRID_SIZE * i] =
302 -ff_g723_1_fixed_cb_gain[subfrm->amp_index];
303 } else {
304 vector[subfrm->grid_index + GRID_SIZE * i] =
305 ff_g723_1_fixed_cb_gain[subfrm->amp_index];
306 }
307 if (j == PULSE_MAX)
308 break;
309 }
310 if (subfrm->dirac_train == 1)
311 ff_g723_1_gen_dirac_train(vector, pitch_lag);
312 } else { /* 5300 bps */
313 int cb_gain = ff_g723_1_fixed_cb_gain[subfrm->amp_index];
314 int cb_shift = subfrm->grid_index;
315 int cb_sign = subfrm->pulse_sign;
316 int cb_pos = subfrm->pulse_pos;
317 int offset, beta, lag;
318
319 for (i = 0; i < 8; i += 2) {
320 offset = ((cb_pos & 7) << 3) + cb_shift + i;
321 vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
322 cb_pos >>= 3;
323 cb_sign >>= 1;
324 }
325
326 /* Enhance harmonic components */
327 lag = pitch_contrib[subfrm->ad_cb_gain << 1] + pitch_lag +
328 subfrm->ad_cb_lag - 1;
329 beta = pitch_contrib[(subfrm->ad_cb_gain << 1) + 1];
330
331 if (lag < SUBFRAME_LEN - 2) {
332 for (i = lag; i < SUBFRAME_LEN; i++)
333 vector[i] += beta * vector[i - lag] >> 15;
334 }
335 }
336 }
337
338 /**
339 * Estimate maximum auto-correlation around pitch lag.
340 *
341 * @param buf buffer with offset applied
342 * @param offset offset of the excitation vector
343 * @param ccr_max pointer to the maximum auto-correlation
344 * @param pitch_lag decoded pitch lag
345 * @param length length of autocorrelation
346 * @param dir forward lag(1) / backward lag(-1)
347 */
autocorr_max(const int16_t * buf,int offset,int * ccr_max,int pitch_lag,int length,int dir)348 static int autocorr_max(const int16_t *buf, int offset, int *ccr_max,
349 int pitch_lag, int length, int dir)
350 {
351 int limit, ccr, lag = 0;
352 int i;
353
354 pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
355 if (dir > 0)
356 limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
357 else
358 limit = pitch_lag + 3;
359
360 for (i = pitch_lag - 3; i <= limit; i++) {
361 ccr = ff_g723_1_dot_product(buf, buf + dir * i, length);
362
363 if (ccr > *ccr_max) {
364 *ccr_max = ccr;
365 lag = i;
366 }
367 }
368 return lag;
369 }
370
371 /**
372 * Calculate pitch postfilter optimal and scaling gains.
373 *
374 * @param lag pitch postfilter forward/backward lag
375 * @param ppf pitch postfilter parameters
376 * @param cur_rate current bitrate
377 * @param tgt_eng target energy
378 * @param ccr cross-correlation
379 * @param res_eng residual energy
380 */
comp_ppf_gains(int lag,PPFParam * ppf,enum Rate cur_rate,int tgt_eng,int ccr,int res_eng)381 static void comp_ppf_gains(int lag, PPFParam *ppf, enum Rate cur_rate,
382 int tgt_eng, int ccr, int res_eng)
383 {
384 int pf_residual; /* square of postfiltered residual */
385 int temp1, temp2;
386
387 ppf->index = lag;
388
389 temp1 = tgt_eng * res_eng >> 1;
390 temp2 = ccr * ccr << 1;
391
392 if (temp2 > temp1) {
393 if (ccr >= res_eng) {
394 ppf->opt_gain = ppf_gain_weight[cur_rate];
395 } else {
396 ppf->opt_gain = (ccr << 15) / res_eng *
397 ppf_gain_weight[cur_rate] >> 15;
398 }
399 /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
400 temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
401 temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
402 pf_residual = av_sat_add32(temp1, temp2 + (1 << 15)) >> 16;
403
404 if (tgt_eng >= pf_residual << 1) {
405 temp1 = 0x7fff;
406 } else {
407 temp1 = (tgt_eng << 14) / pf_residual;
408 }
409
410 /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
411 ppf->sc_gain = square_root(temp1 << 16);
412 } else {
413 ppf->opt_gain = 0;
414 ppf->sc_gain = 0x7fff;
415 }
416
417 ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
418 }
419
420 /**
421 * Calculate pitch postfilter parameters.
422 *
423 * @param p the context
424 * @param offset offset of the excitation vector
425 * @param pitch_lag decoded pitch lag
426 * @param ppf pitch postfilter parameters
427 * @param cur_rate current bitrate
428 */
comp_ppf_coeff(G723_1_ChannelContext * p,int offset,int pitch_lag,PPFParam * ppf,enum Rate cur_rate)429 static void comp_ppf_coeff(G723_1_ChannelContext *p, int offset, int pitch_lag,
430 PPFParam *ppf, enum Rate cur_rate)
431 {
432
433 int16_t scale;
434 int i;
435 int temp1, temp2;
436
437 /*
438 * 0 - target energy
439 * 1 - forward cross-correlation
440 * 2 - forward residual energy
441 * 3 - backward cross-correlation
442 * 4 - backward residual energy
443 */
444 int energy[5] = {0, 0, 0, 0, 0};
445 int16_t *buf = p->audio + LPC_ORDER + offset;
446 int fwd_lag = autocorr_max(buf, offset, &energy[1], pitch_lag,
447 SUBFRAME_LEN, 1);
448 int back_lag = autocorr_max(buf, offset, &energy[3], pitch_lag,
449 SUBFRAME_LEN, -1);
450
451 ppf->index = 0;
452 ppf->opt_gain = 0;
453 ppf->sc_gain = 0x7fff;
454
455 /* Case 0, Section 3.6 */
456 if (!back_lag && !fwd_lag)
457 return;
458
459 /* Compute target energy */
460 energy[0] = ff_g723_1_dot_product(buf, buf, SUBFRAME_LEN);
461
462 /* Compute forward residual energy */
463 if (fwd_lag)
464 energy[2] = ff_g723_1_dot_product(buf + fwd_lag, buf + fwd_lag,
465 SUBFRAME_LEN);
466
467 /* Compute backward residual energy */
468 if (back_lag)
469 energy[4] = ff_g723_1_dot_product(buf - back_lag, buf - back_lag,
470 SUBFRAME_LEN);
471
472 /* Normalize and shorten */
473 temp1 = 0;
474 for (i = 0; i < 5; i++)
475 temp1 = FFMAX(energy[i], temp1);
476
477 scale = ff_g723_1_normalize_bits(temp1, 31);
478 for (i = 0; i < 5; i++)
479 energy[i] = (energy[i] << scale) >> 16;
480
481 if (fwd_lag && !back_lag) { /* Case 1 */
482 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
483 energy[2]);
484 } else if (!fwd_lag) { /* Case 2 */
485 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
486 energy[4]);
487 } else { /* Case 3 */
488
489 /*
490 * Select the largest of energy[1]^2/energy[2]
491 * and energy[3]^2/energy[4]
492 */
493 temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
494 temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
495 if (temp1 >= temp2) {
496 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
497 energy[2]);
498 } else {
499 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
500 energy[4]);
501 }
502 }
503 }
504
505 /**
506 * Classify frames as voiced/unvoiced.
507 *
508 * @param p the context
509 * @param pitch_lag decoded pitch_lag
510 * @param exc_eng excitation energy estimation
511 * @param scale scaling factor of exc_eng
512 *
513 * @return residual interpolation index if voiced, 0 otherwise
514 */
comp_interp_index(G723_1_ChannelContext * p,int pitch_lag,int * exc_eng,int * scale)515 static int comp_interp_index(G723_1_ChannelContext *p, int pitch_lag,
516 int *exc_eng, int *scale)
517 {
518 int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
519 int16_t *buf = p->audio + LPC_ORDER;
520
521 int index, ccr, tgt_eng, best_eng, temp;
522
523 *scale = ff_g723_1_scale_vector(buf, p->excitation, FRAME_LEN + PITCH_MAX);
524 buf += offset;
525
526 /* Compute maximum backward cross-correlation */
527 ccr = 0;
528 index = autocorr_max(buf, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
529 ccr = av_sat_add32(ccr, 1 << 15) >> 16;
530
531 /* Compute target energy */
532 tgt_eng = ff_g723_1_dot_product(buf, buf, SUBFRAME_LEN * 2);
533 *exc_eng = av_sat_add32(tgt_eng, 1 << 15) >> 16;
534
535 if (ccr <= 0)
536 return 0;
537
538 /* Compute best energy */
539 best_eng = ff_g723_1_dot_product(buf - index, buf - index,
540 SUBFRAME_LEN * 2);
541 best_eng = av_sat_add32(best_eng, 1 << 15) >> 16;
542
543 temp = best_eng * *exc_eng >> 3;
544
545 if (temp < ccr * ccr) {
546 return index;
547 } else
548 return 0;
549 }
550
551 /**
552 * Perform residual interpolation based on frame classification.
553 *
554 * @param buf decoded excitation vector
555 * @param out output vector
556 * @param lag decoded pitch lag
557 * @param gain interpolated gain
558 * @param rseed seed for random number generator
559 */
residual_interp(int16_t * buf,int16_t * out,int lag,int gain,int * rseed)560 static void residual_interp(int16_t *buf, int16_t *out, int lag,
561 int gain, int *rseed)
562 {
563 int i;
564 if (lag) { /* Voiced */
565 int16_t *vector_ptr = buf + PITCH_MAX;
566 /* Attenuate */
567 for (i = 0; i < lag; i++)
568 out[i] = vector_ptr[i - lag] * 3 >> 2;
569 av_memcpy_backptr((uint8_t*)(out + lag), lag * sizeof(*out),
570 (FRAME_LEN - lag) * sizeof(*out));
571 } else { /* Unvoiced */
572 for (i = 0; i < FRAME_LEN; i++) {
573 *rseed = (int16_t)(*rseed * 521 + 259);
574 out[i] = gain * *rseed >> 15;
575 }
576 memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(*buf));
577 }
578 }
579
580 /**
581 * Perform IIR filtering.
582 *
583 * @param fir_coef FIR coefficients
584 * @param iir_coef IIR coefficients
585 * @param src source vector
586 * @param dest destination vector
587 * @param width width of the output, 16 bits(0) / 32 bits(1)
588 */
589 #define iir_filter(fir_coef, iir_coef, src, dest, width)\
590 {\
591 int m, n;\
592 int res_shift = 16 & ~-(width);\
593 int in_shift = 16 - res_shift;\
594 \
595 for (m = 0; m < SUBFRAME_LEN; m++) {\
596 int64_t filter = 0;\
597 for (n = 1; n <= LPC_ORDER; n++) {\
598 filter -= (fir_coef)[n - 1] * (src)[m - n] -\
599 (iir_coef)[n - 1] * ((dest)[m - n] >> in_shift);\
600 }\
601 \
602 (dest)[m] = av_clipl_int32(((src)[m] * 65536) + (filter * 8) +\
603 (1 << 15)) >> res_shift;\
604 }\
605 }
606
607 /**
608 * Adjust gain of postfiltered signal.
609 *
610 * @param p the context
611 * @param buf postfiltered output vector
612 * @param energy input energy coefficient
613 */
gain_scale(G723_1_ChannelContext * p,int16_t * buf,int energy)614 static void gain_scale(G723_1_ChannelContext *p, int16_t * buf, int energy)
615 {
616 int num, denom, gain, bits1, bits2;
617 int i;
618
619 num = energy;
620 denom = 0;
621 for (i = 0; i < SUBFRAME_LEN; i++) {
622 int temp = buf[i] >> 2;
623 temp *= temp;
624 denom = av_sat_dadd32(denom, temp);
625 }
626
627 if (num && denom) {
628 bits1 = ff_g723_1_normalize_bits(num, 31);
629 bits2 = ff_g723_1_normalize_bits(denom, 31);
630 num = num << bits1 >> 1;
631 denom <<= bits2;
632
633 bits2 = 5 + bits1 - bits2;
634 bits2 = av_clip_uintp2(bits2, 5);
635
636 gain = (num >> 1) / (denom >> 16);
637 gain = square_root(gain << 16 >> bits2);
638 } else {
639 gain = 1 << 12;
640 }
641
642 for (i = 0; i < SUBFRAME_LEN; i++) {
643 p->pf_gain = (15 * p->pf_gain + gain + (1 << 3)) >> 4;
644 buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
645 (1 << 10)) >> 11);
646 }
647 }
648
649 /**
650 * Perform formant filtering.
651 *
652 * @param p the context
653 * @param lpc quantized lpc coefficients
654 * @param buf input buffer
655 * @param dst output buffer
656 */
formant_postfilter(G723_1_ChannelContext * p,int16_t * lpc,int16_t * buf,int16_t * dst)657 static void formant_postfilter(G723_1_ChannelContext *p, int16_t *lpc,
658 int16_t *buf, int16_t *dst)
659 {
660 int16_t filter_coef[2][LPC_ORDER];
661 int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
662 int i, j, k;
663
664 memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(*buf));
665 memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(*filter_signal));
666
667 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
668 for (k = 0; k < LPC_ORDER; k++) {
669 filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
670 (1 << 14)) >> 15;
671 filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
672 (1 << 14)) >> 15;
673 }
674 iir_filter(filter_coef[0], filter_coef[1], buf + i, filter_signal + i, 1);
675 lpc += LPC_ORDER;
676 }
677
678 memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(int16_t));
679 memcpy(p->iir_mem, filter_signal + FRAME_LEN, LPC_ORDER * sizeof(int));
680
681 buf += LPC_ORDER;
682 signal_ptr = filter_signal + LPC_ORDER;
683 for (i = 0; i < SUBFRAMES; i++) {
684 int temp;
685 int auto_corr[2];
686 int scale, energy;
687
688 /* Normalize */
689 scale = ff_g723_1_scale_vector(dst, buf, SUBFRAME_LEN);
690
691 /* Compute auto correlation coefficients */
692 auto_corr[0] = ff_g723_1_dot_product(dst, dst + 1, SUBFRAME_LEN - 1);
693 auto_corr[1] = ff_g723_1_dot_product(dst, dst, SUBFRAME_LEN);
694
695 /* Compute reflection coefficient */
696 temp = auto_corr[1] >> 16;
697 if (temp) {
698 temp = (auto_corr[0] >> 2) / temp;
699 }
700 p->reflection_coef = (3 * p->reflection_coef + temp + 2) >> 2;
701 temp = -p->reflection_coef >> 1 & ~3;
702
703 /* Compensation filter */
704 for (j = 0; j < SUBFRAME_LEN; j++) {
705 dst[j] = av_sat_dadd32(signal_ptr[j],
706 (signal_ptr[j - 1] >> 16) * temp) >> 16;
707 }
708
709 /* Compute normalized signal energy */
710 temp = 2 * scale + 4;
711 if (temp < 0) {
712 energy = av_clipl_int32((int64_t)auto_corr[1] << -temp);
713 } else
714 energy = auto_corr[1] >> temp;
715
716 gain_scale(p, dst, energy);
717
718 buf += SUBFRAME_LEN;
719 signal_ptr += SUBFRAME_LEN;
720 dst += SUBFRAME_LEN;
721 }
722 }
723
sid_gain_to_lsp_index(int gain)724 static int sid_gain_to_lsp_index(int gain)
725 {
726 if (gain < 0x10)
727 return gain << 6;
728 else if (gain < 0x20)
729 return gain - 8 << 7;
730 else
731 return gain - 20 << 8;
732 }
733
cng_rand(int * state,int base)734 static inline int cng_rand(int *state, int base)
735 {
736 *state = (*state * 521 + 259) & 0xFFFF;
737 return (*state & 0x7FFF) * base >> 15;
738 }
739
estimate_sid_gain(G723_1_ChannelContext * p)740 static int estimate_sid_gain(G723_1_ChannelContext *p)
741 {
742 int i, shift, seg, seg2, t, val, val_add, x, y;
743
744 shift = 16 - p->cur_gain * 2;
745 if (shift > 0) {
746 if (p->sid_gain == 0) {
747 t = 0;
748 } else if (shift >= 31 || (int32_t)((uint32_t)p->sid_gain << shift) >> shift != p->sid_gain) {
749 if (p->sid_gain < 0) t = INT32_MIN;
750 else t = INT32_MAX;
751 } else
752 t = p->sid_gain * (1 << shift);
753 } else if(shift < -31) {
754 t = (p->sid_gain < 0) ? -1 : 0;
755 }else
756 t = p->sid_gain >> -shift;
757 x = av_clipl_int32(t * (int64_t)cng_filt[0] >> 16);
758
759 if (x >= cng_bseg[2])
760 return 0x3F;
761
762 if (x >= cng_bseg[1]) {
763 shift = 4;
764 seg = 3;
765 } else {
766 shift = 3;
767 seg = (x >= cng_bseg[0]);
768 }
769 seg2 = FFMIN(seg, 3);
770
771 val = 1 << shift;
772 val_add = val >> 1;
773 for (i = 0; i < shift; i++) {
774 t = seg * 32 + (val << seg2);
775 t *= t;
776 if (x >= t)
777 val += val_add;
778 else
779 val -= val_add;
780 val_add >>= 1;
781 }
782
783 t = seg * 32 + (val << seg2);
784 y = t * t - x;
785 if (y <= 0) {
786 t = seg * 32 + (val + 1 << seg2);
787 t = t * t - x;
788 val = (seg2 - 1) * 16 + val;
789 if (t >= y)
790 val++;
791 } else {
792 t = seg * 32 + (val - 1 << seg2);
793 t = t * t - x;
794 val = (seg2 - 1) * 16 + val;
795 if (t >= y)
796 val--;
797 }
798
799 return val;
800 }
801
generate_noise(G723_1_ChannelContext * p)802 static void generate_noise(G723_1_ChannelContext *p)
803 {
804 int i, j, idx, t;
805 int off[SUBFRAMES];
806 int signs[SUBFRAMES / 2 * 11], pos[SUBFRAMES / 2 * 11];
807 int tmp[SUBFRAME_LEN * 2];
808 int16_t *vector_ptr;
809 int64_t sum;
810 int b0, c, delta, x, shift;
811
812 p->pitch_lag[0] = cng_rand(&p->cng_random_seed, 21) + 123;
813 p->pitch_lag[1] = cng_rand(&p->cng_random_seed, 19) + 123;
814
815 for (i = 0; i < SUBFRAMES; i++) {
816 p->subframe[i].ad_cb_gain = cng_rand(&p->cng_random_seed, 50) + 1;
817 p->subframe[i].ad_cb_lag = cng_adaptive_cb_lag[i];
818 }
819
820 for (i = 0; i < SUBFRAMES / 2; i++) {
821 t = cng_rand(&p->cng_random_seed, 1 << 13);
822 off[i * 2] = t & 1;
823 off[i * 2 + 1] = ((t >> 1) & 1) + SUBFRAME_LEN;
824 t >>= 2;
825 for (j = 0; j < 11; j++) {
826 signs[i * 11 + j] = ((t & 1) * 2 - 1) * (1 << 14);
827 t >>= 1;
828 }
829 }
830
831 idx = 0;
832 for (i = 0; i < SUBFRAMES; i++) {
833 for (j = 0; j < SUBFRAME_LEN / 2; j++)
834 tmp[j] = j;
835 t = SUBFRAME_LEN / 2;
836 for (j = 0; j < pulses[i]; j++, idx++) {
837 int idx2 = cng_rand(&p->cng_random_seed, t);
838
839 pos[idx] = tmp[idx2] * 2 + off[i];
840 tmp[idx2] = tmp[--t];
841 }
842 }
843
844 vector_ptr = p->audio + LPC_ORDER;
845 memcpy(vector_ptr, p->prev_excitation,
846 PITCH_MAX * sizeof(*p->excitation));
847 for (i = 0; i < SUBFRAMES; i += 2) {
848 ff_g723_1_gen_acb_excitation(vector_ptr, vector_ptr,
849 p->pitch_lag[i >> 1], &p->subframe[i],
850 p->cur_rate);
851 ff_g723_1_gen_acb_excitation(vector_ptr + SUBFRAME_LEN,
852 vector_ptr + SUBFRAME_LEN,
853 p->pitch_lag[i >> 1], &p->subframe[i + 1],
854 p->cur_rate);
855
856 t = 0;
857 for (j = 0; j < SUBFRAME_LEN * 2; j++)
858 t |= FFABS(vector_ptr[j]);
859 t = FFMIN(t, 0x7FFF);
860 if (!t) {
861 shift = 0;
862 } else {
863 shift = -10 + av_log2(t);
864 if (shift < -2)
865 shift = -2;
866 }
867 sum = 0;
868 if (shift < 0) {
869 for (j = 0; j < SUBFRAME_LEN * 2; j++) {
870 t = vector_ptr[j] * (1 << -shift);
871 sum += t * t;
872 tmp[j] = t;
873 }
874 } else {
875 for (j = 0; j < SUBFRAME_LEN * 2; j++) {
876 t = vector_ptr[j] >> shift;
877 sum += t * t;
878 tmp[j] = t;
879 }
880 }
881
882 b0 = 0;
883 for (j = 0; j < 11; j++)
884 b0 += tmp[pos[(i / 2) * 11 + j]] * signs[(i / 2) * 11 + j];
885 b0 = b0 * 2 * 2979LL + (1 << 29) >> 30; // approximated division by 11
886
887 c = p->cur_gain * (p->cur_gain * SUBFRAME_LEN >> 5);
888 if (shift * 2 + 3 >= 0)
889 c >>= shift * 2 + 3;
890 else
891 c <<= -(shift * 2 + 3);
892 c = (av_clipl_int32(sum << 1) - c) * 2979LL >> 15;
893
894 delta = b0 * b0 * 2 - c;
895 if (delta <= 0) {
896 x = -b0;
897 } else {
898 delta = square_root(delta);
899 x = delta - b0;
900 t = delta + b0;
901 if (FFABS(t) < FFABS(x))
902 x = -t;
903 }
904 shift++;
905 if (shift < 0)
906 x >>= -shift;
907 else
908 x *= 1 << shift;
909 x = av_clip(x, -10000, 10000);
910
911 for (j = 0; j < 11; j++) {
912 idx = (i / 2) * 11 + j;
913 vector_ptr[pos[idx]] = av_clip_int16(vector_ptr[pos[idx]] +
914 (x * signs[idx] >> 15));
915 }
916
917 /* copy decoded data to serve as a history for the next decoded subframes */
918 memcpy(vector_ptr + PITCH_MAX, vector_ptr,
919 sizeof(*vector_ptr) * SUBFRAME_LEN * 2);
920 vector_ptr += SUBFRAME_LEN * 2;
921 }
922 /* Save the excitation for the next frame */
923 memcpy(p->prev_excitation, p->audio + LPC_ORDER + FRAME_LEN,
924 PITCH_MAX * sizeof(*p->excitation));
925 }
926
g723_1_decode_frame(AVCodecContext * avctx,void * data,int * got_frame_ptr,AVPacket * avpkt)927 static int g723_1_decode_frame(AVCodecContext *avctx, void *data,
928 int *got_frame_ptr, AVPacket *avpkt)
929 {
930 G723_1_Context *s = avctx->priv_data;
931 AVFrame *frame = data;
932 const uint8_t *buf = avpkt->data;
933 int buf_size = avpkt->size;
934 int dec_mode = buf[0] & 3;
935
936 PPFParam ppf[SUBFRAMES];
937 int16_t cur_lsp[LPC_ORDER];
938 int16_t lpc[SUBFRAMES * LPC_ORDER];
939 int16_t acb_vector[SUBFRAME_LEN];
940 int16_t *out;
941 int bad_frame = 0, i, j, ret;
942
943 if (buf_size < frame_size[dec_mode] * avctx->channels) {
944 if (buf_size)
945 av_log(avctx, AV_LOG_WARNING,
946 "Expected %d bytes, got %d - skipping packet\n",
947 frame_size[dec_mode], buf_size);
948 *got_frame_ptr = 0;
949 return buf_size;
950 }
951
952 frame->nb_samples = FRAME_LEN;
953 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
954 return ret;
955
956 for (int ch = 0; ch < avctx->channels; ch++) {
957 G723_1_ChannelContext *p = &s->ch[ch];
958 int16_t *audio = p->audio;
959
960 if (unpack_bitstream(p, buf + ch * (buf_size / avctx->channels),
961 buf_size / avctx->channels) < 0) {
962 bad_frame = 1;
963 if (p->past_frame_type == ACTIVE_FRAME)
964 p->cur_frame_type = ACTIVE_FRAME;
965 else
966 p->cur_frame_type = UNTRANSMITTED_FRAME;
967 }
968
969 out = (int16_t *)frame->extended_data[ch];
970
971 if (p->cur_frame_type == ACTIVE_FRAME) {
972 if (!bad_frame)
973 p->erased_frames = 0;
974 else if (p->erased_frames != 3)
975 p->erased_frames++;
976
977 ff_g723_1_inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
978 ff_g723_1_lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
979
980 /* Save the lsp_vector for the next frame */
981 memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
982
983 /* Generate the excitation for the frame */
984 memcpy(p->excitation, p->prev_excitation,
985 PITCH_MAX * sizeof(*p->excitation));
986 if (!p->erased_frames) {
987 int16_t *vector_ptr = p->excitation + PITCH_MAX;
988
989 /* Update interpolation gain memory */
990 p->interp_gain = ff_g723_1_fixed_cb_gain[(p->subframe[2].amp_index +
991 p->subframe[3].amp_index) >> 1];
992 for (i = 0; i < SUBFRAMES; i++) {
993 gen_fcb_excitation(vector_ptr, &p->subframe[i], p->cur_rate,
994 p->pitch_lag[i >> 1], i);
995 ff_g723_1_gen_acb_excitation(acb_vector,
996 &p->excitation[SUBFRAME_LEN * i],
997 p->pitch_lag[i >> 1],
998 &p->subframe[i], p->cur_rate);
999 /* Get the total excitation */
1000 for (j = 0; j < SUBFRAME_LEN; j++) {
1001 int v = av_clip_int16(vector_ptr[j] * 2);
1002 vector_ptr[j] = av_clip_int16(v + acb_vector[j]);
1003 }
1004 vector_ptr += SUBFRAME_LEN;
1005 }
1006
1007 vector_ptr = p->excitation + PITCH_MAX;
1008
1009 p->interp_index = comp_interp_index(p, p->pitch_lag[1],
1010 &p->sid_gain, &p->cur_gain);
1011
1012 /* Perform pitch postfiltering */
1013 if (s->postfilter) {
1014 i = PITCH_MAX;
1015 for (j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1016 comp_ppf_coeff(p, i, p->pitch_lag[j >> 1],
1017 ppf + j, p->cur_rate);
1018
1019 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1020 ff_acelp_weighted_vector_sum(p->audio + LPC_ORDER + i,
1021 vector_ptr + i,
1022 vector_ptr + i + ppf[j].index,
1023 ppf[j].sc_gain,
1024 ppf[j].opt_gain,
1025 1 << 14, 15, SUBFRAME_LEN);
1026 } else {
1027 audio = vector_ptr - LPC_ORDER;
1028 }
1029
1030 /* Save the excitation for the next frame */
1031 memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
1032 PITCH_MAX * sizeof(*p->excitation));
1033 } else {
1034 p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
1035 if (p->erased_frames == 3) {
1036 /* Mute output */
1037 memset(p->excitation, 0,
1038 (FRAME_LEN + PITCH_MAX) * sizeof(*p->excitation));
1039 memset(p->prev_excitation, 0,
1040 PITCH_MAX * sizeof(*p->excitation));
1041 memset(frame->data[0], 0,
1042 (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
1043 } else {
1044 int16_t *buf = p->audio + LPC_ORDER;
1045
1046 /* Regenerate frame */
1047 residual_interp(p->excitation, buf, p->interp_index,
1048 p->interp_gain, &p->random_seed);
1049
1050 /* Save the excitation for the next frame */
1051 memcpy(p->prev_excitation, buf + (FRAME_LEN - PITCH_MAX),
1052 PITCH_MAX * sizeof(*p->excitation));
1053 }
1054 }
1055 p->cng_random_seed = CNG_RANDOM_SEED;
1056 } else {
1057 if (p->cur_frame_type == SID_FRAME) {
1058 p->sid_gain = sid_gain_to_lsp_index(p->subframe[0].amp_index);
1059 ff_g723_1_inverse_quant(p->sid_lsp, p->prev_lsp, p->lsp_index, 0);
1060 } else if (p->past_frame_type == ACTIVE_FRAME) {
1061 p->sid_gain = estimate_sid_gain(p);
1062 }
1063
1064 if (p->past_frame_type == ACTIVE_FRAME)
1065 p->cur_gain = p->sid_gain;
1066 else
1067 p->cur_gain = (p->cur_gain * 7 + p->sid_gain) >> 3;
1068 generate_noise(p);
1069 ff_g723_1_lsp_interpolate(lpc, p->sid_lsp, p->prev_lsp);
1070 /* Save the lsp_vector for the next frame */
1071 memcpy(p->prev_lsp, p->sid_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1072 }
1073
1074 p->past_frame_type = p->cur_frame_type;
1075
1076 memcpy(p->audio, p->synth_mem, LPC_ORDER * sizeof(*p->audio));
1077 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1078 ff_celp_lp_synthesis_filter(p->audio + i, &lpc[j * LPC_ORDER],
1079 audio + i, SUBFRAME_LEN, LPC_ORDER,
1080 0, 1, 1 << 12);
1081 memcpy(p->synth_mem, p->audio + FRAME_LEN, LPC_ORDER * sizeof(*p->audio));
1082
1083 if (s->postfilter) {
1084 formant_postfilter(p, lpc, p->audio, out);
1085 } else { // if output is not postfiltered it should be scaled by 2
1086 for (i = 0; i < FRAME_LEN; i++)
1087 out[i] = av_clip_int16(2 * p->audio[LPC_ORDER + i]);
1088 }
1089 }
1090
1091 *got_frame_ptr = 1;
1092
1093 return frame_size[dec_mode] * avctx->channels;
1094 }
1095
1096 #define OFFSET(x) offsetof(G723_1_Context, x)
1097 #define AD AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1098
1099 static const AVOption options[] = {
1100 { "postfilter", "enable postfilter", OFFSET(postfilter), AV_OPT_TYPE_BOOL,
1101 { .i64 = 1 }, 0, 1, AD },
1102 { NULL }
1103 };
1104
1105
1106 static const AVClass g723_1dec_class = {
1107 .class_name = "G.723.1 decoder",
1108 .item_name = av_default_item_name,
1109 .option = options,
1110 .version = LIBAVUTIL_VERSION_INT,
1111 };
1112
1113 AVCodec ff_g723_1_decoder = {
1114 .name = "g723_1",
1115 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1116 .type = AVMEDIA_TYPE_AUDIO,
1117 .id = AV_CODEC_ID_G723_1,
1118 .priv_data_size = sizeof(G723_1_Context),
1119 .init = g723_1_decode_init,
1120 .decode = g723_1_decode_frame,
1121 .capabilities = AV_CODEC_CAP_SUBFRAMES | AV_CODEC_CAP_DR1,
1122 .priv_class = &g723_1dec_class,
1123 };
1124