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 "codec_internal.h"
38 #include "get_bits.h"
39 #include "internal.h"
40 #include "g723_1.h"
41
42 #define CNG_RANDOM_SEED 12345
43
44 /**
45 * Postfilter gain weighting factors scaled by 2^15
46 */
47 static const int16_t ppf_gain_weight[2] = {0x1800, 0x2000};
48
49 static const int16_t pitch_contrib[340] = {
50 60, 0, 0, 2489, 60, 0, 0, 5217,
51 1, 6171, 0, 3953, 0, 10364, 1, 9357,
52 -1, 8843, 1, 9396, 0, 5794, -1, 10816,
53 2, 11606, -2, 12072, 0, 8616, 1, 12170,
54 0, 14440, 0, 7787, -1, 13721, 0, 18205,
55 0, 14471, 0, 15807, 1, 15275, 0, 13480,
56 -1, 18375, -1, 0, 1, 11194, -1, 13010,
57 1, 18836, -2, 20354, 1, 16233, -1, 0,
58 60, 0, 0, 12130, 0, 13385, 1, 17834,
59 1, 20875, 0, 21996, 1, 0, 1, 18277,
60 -1, 21321, 1, 13738, -1, 19094, -1, 20387,
61 -1, 0, 0, 21008, 60, 0, -2, 22807,
62 0, 15900, 1, 0, 0, 17989, -1, 22259,
63 1, 24395, 1, 23138, 0, 23948, 1, 22997,
64 2, 22604, -1, 25942, 0, 26246, 1, 25321,
65 0, 26423, 0, 24061, 0, 27247, 60, 0,
66 -1, 25572, 1, 23918, 1, 25930, 2, 26408,
67 -1, 19049, 1, 27357, -1, 24538, 60, 0,
68 -1, 25093, 0, 28549, 1, 0, 0, 22793,
69 -1, 25659, 0, 29377, 0, 30276, 0, 26198,
70 1, 22521, -1, 28919, 0, 27384, 1, 30162,
71 -1, 0, 0, 24237, -1, 30062, 0, 21763,
72 1, 30917, 60, 0, 0, 31284, 0, 29433,
73 1, 26821, 1, 28655, 0, 31327, 2, 30799,
74 1, 31389, 0, 32322, 1, 31760, -2, 31830,
75 0, 26936, -1, 31180, 1, 30875, 0, 27873,
76 -1, 30429, 1, 31050, 0, 0, 0, 31912,
77 1, 31611, 0, 31565, 0, 25557, 0, 31357,
78 60, 0, 1, 29536, 1, 28985, -1, 26984,
79 -1, 31587, 2, 30836, -2, 31133, 0, 30243,
80 -1, 30742, -1, 32090, 60, 0, 2, 30902,
81 60, 0, 0, 30027, 0, 29042, 60, 0,
82 0, 31756, 0, 24553, 0, 25636, -2, 30501,
83 60, 0, -1, 29617, 0, 30649, 60, 0,
84 0, 29274, 2, 30415, 0, 27480, 0, 31213,
85 -1, 28147, 0, 30600, 1, 31652, 2, 29068,
86 60, 0, 1, 28571, 1, 28730, 1, 31422,
87 0, 28257, 0, 24797, 60, 0, 0, 0,
88 60, 0, 0, 22105, 0, 27852, 60, 0,
89 60, 0, -1, 24214, 0, 24642, 0, 23305,
90 60, 0, 60, 0, 1, 22883, 0, 21601,
91 60, 0, 2, 25650, 60, 0, -2, 31253,
92 -2, 25144, 0, 17998
93 };
94
95 /**
96 * Size of the MP-MLQ fixed excitation codebooks
97 */
98 static const int32_t max_pos[4] = {593775, 142506, 593775, 142506};
99
100 /**
101 * 0.65^i (Zero part) and 0.75^i (Pole part) scaled by 2^15
102 */
103 static const int16_t postfilter_tbl[2][LPC_ORDER] = {
104 /* Zero */
105 {21299, 13844, 8999, 5849, 3802, 2471, 1606, 1044, 679, 441},
106 /* Pole */
107 {24576, 18432, 13824, 10368, 7776, 5832, 4374, 3281, 2460, 1845}
108 };
109
110 static const int cng_adaptive_cb_lag[4] = { 1, 0, 1, 3 };
111
112 static const int cng_filt[4] = { 273, 998, 499, 333 };
113
114 static const int cng_bseg[3] = { 2048, 18432, 231233 };
115
g723_1_decode_init(AVCodecContext * avctx)116 static av_cold int g723_1_decode_init(AVCodecContext *avctx)
117 {
118 G723_1_Context *s = avctx->priv_data;
119
120 avctx->sample_fmt = AV_SAMPLE_FMT_S16P;
121 if (avctx->ch_layout.nb_channels < 1 || avctx->ch_layout.nb_channels > 2) {
122 av_log(avctx, AV_LOG_ERROR, "Only mono and stereo are supported (requested channels: %d).\n",
123 avctx->ch_layout.nb_channels);
124 return AVERROR(EINVAL);
125 }
126 for (int ch = 0; ch < avctx->ch_layout.nb_channels; ch++) {
127 G723_1_ChannelContext *p = &s->ch[ch];
128
129 p->pf_gain = 1 << 12;
130
131 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
132 memcpy(p->sid_lsp, dc_lsp, LPC_ORDER * sizeof(*p->sid_lsp));
133
134 p->cng_random_seed = CNG_RANDOM_SEED;
135 p->past_frame_type = SID_FRAME;
136 }
137
138 return 0;
139 }
140
141 /**
142 * Unpack the frame into parameters.
143 *
144 * @param p the context
145 * @param buf pointer to the input buffer
146 * @param buf_size size of the input buffer
147 */
unpack_bitstream(G723_1_ChannelContext * p,const uint8_t * buf,int buf_size)148 static int unpack_bitstream(G723_1_ChannelContext *p, const uint8_t *buf,
149 int buf_size)
150 {
151 GetBitContext gb;
152 int ad_cb_len;
153 int temp, info_bits, i;
154 int ret;
155
156 ret = init_get_bits8(&gb, buf, buf_size);
157 if (ret < 0)
158 return ret;
159
160 /* Extract frame type and rate info */
161 info_bits = get_bits(&gb, 2);
162
163 if (info_bits == 3) {
164 p->cur_frame_type = UNTRANSMITTED_FRAME;
165 return 0;
166 }
167
168 /* Extract 24 bit lsp indices, 8 bit for each band */
169 p->lsp_index[2] = get_bits(&gb, 8);
170 p->lsp_index[1] = get_bits(&gb, 8);
171 p->lsp_index[0] = get_bits(&gb, 8);
172
173 if (info_bits == 2) {
174 p->cur_frame_type = SID_FRAME;
175 p->subframe[0].amp_index = get_bits(&gb, 6);
176 return 0;
177 }
178
179 /* Extract the info common to both rates */
180 p->cur_rate = info_bits ? RATE_5300 : RATE_6300;
181 p->cur_frame_type = ACTIVE_FRAME;
182
183 p->pitch_lag[0] = get_bits(&gb, 7);
184 if (p->pitch_lag[0] > 123) /* test if forbidden code */
185 return -1;
186 p->pitch_lag[0] += PITCH_MIN;
187 p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
188
189 p->pitch_lag[1] = get_bits(&gb, 7);
190 if (p->pitch_lag[1] > 123)
191 return -1;
192 p->pitch_lag[1] += PITCH_MIN;
193 p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
194 p->subframe[0].ad_cb_lag = 1;
195 p->subframe[2].ad_cb_lag = 1;
196
197 for (i = 0; i < SUBFRAMES; i++) {
198 /* Extract combined gain */
199 temp = get_bits(&gb, 12);
200 ad_cb_len = 170;
201 p->subframe[i].dirac_train = 0;
202 if (p->cur_rate == RATE_6300 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
203 p->subframe[i].dirac_train = temp >> 11;
204 temp &= 0x7FF;
205 ad_cb_len = 85;
206 }
207 p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
208 if (p->subframe[i].ad_cb_gain < ad_cb_len) {
209 p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
210 GAIN_LEVELS;
211 } else {
212 return -1;
213 }
214 }
215
216 p->subframe[0].grid_index = get_bits1(&gb);
217 p->subframe[1].grid_index = get_bits1(&gb);
218 p->subframe[2].grid_index = get_bits1(&gb);
219 p->subframe[3].grid_index = get_bits1(&gb);
220
221 if (p->cur_rate == RATE_6300) {
222 skip_bits1(&gb); /* skip reserved bit */
223
224 /* Compute pulse_pos index using the 13-bit combined position index */
225 temp = get_bits(&gb, 13);
226 p->subframe[0].pulse_pos = temp / 810;
227
228 temp -= p->subframe[0].pulse_pos * 810;
229 p->subframe[1].pulse_pos = FASTDIV(temp, 90);
230
231 temp -= p->subframe[1].pulse_pos * 90;
232 p->subframe[2].pulse_pos = FASTDIV(temp, 9);
233 p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
234
235 p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
236 get_bits(&gb, 16);
237 p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
238 get_bits(&gb, 14);
239 p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
240 get_bits(&gb, 16);
241 p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
242 get_bits(&gb, 14);
243
244 p->subframe[0].pulse_sign = get_bits(&gb, 6);
245 p->subframe[1].pulse_sign = get_bits(&gb, 5);
246 p->subframe[2].pulse_sign = get_bits(&gb, 6);
247 p->subframe[3].pulse_sign = get_bits(&gb, 5);
248 } else { /* 5300 bps */
249 p->subframe[0].pulse_pos = get_bits(&gb, 12);
250 p->subframe[1].pulse_pos = get_bits(&gb, 12);
251 p->subframe[2].pulse_pos = get_bits(&gb, 12);
252 p->subframe[3].pulse_pos = get_bits(&gb, 12);
253
254 p->subframe[0].pulse_sign = get_bits(&gb, 4);
255 p->subframe[1].pulse_sign = get_bits(&gb, 4);
256 p->subframe[2].pulse_sign = get_bits(&gb, 4);
257 p->subframe[3].pulse_sign = get_bits(&gb, 4);
258 }
259
260 return 0;
261 }
262
263 /**
264 * Bitexact implementation of sqrt(val/2).
265 */
square_root(unsigned val)266 static int16_t square_root(unsigned val)
267 {
268 av_assert2(!(val & 0x80000000));
269
270 return (ff_sqrt(val << 1) >> 1) & (~1);
271 }
272
273 /**
274 * Generate fixed codebook excitation vector.
275 *
276 * @param vector decoded excitation vector
277 * @param subfrm current subframe
278 * @param cur_rate current bitrate
279 * @param pitch_lag closed loop pitch lag
280 * @param index current subframe index
281 */
gen_fcb_excitation(int16_t * vector,G723_1_Subframe * subfrm,enum Rate cur_rate,int pitch_lag,int index)282 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe *subfrm,
283 enum Rate cur_rate, int pitch_lag, int index)
284 {
285 int temp, i, j;
286
287 memset(vector, 0, SUBFRAME_LEN * sizeof(*vector));
288
289 if (cur_rate == RATE_6300) {
290 if (subfrm->pulse_pos >= max_pos[index])
291 return;
292
293 /* Decode amplitudes and positions */
294 j = PULSE_MAX - pulses[index];
295 temp = subfrm->pulse_pos;
296 for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
297 temp -= ff_g723_1_combinatorial_table[j][i];
298 if (temp >= 0)
299 continue;
300 temp += ff_g723_1_combinatorial_table[j++][i];
301 if (subfrm->pulse_sign & (1 << (PULSE_MAX - j))) {
302 vector[subfrm->grid_index + GRID_SIZE * i] =
303 -ff_g723_1_fixed_cb_gain[subfrm->amp_index];
304 } else {
305 vector[subfrm->grid_index + GRID_SIZE * i] =
306 ff_g723_1_fixed_cb_gain[subfrm->amp_index];
307 }
308 if (j == PULSE_MAX)
309 break;
310 }
311 if (subfrm->dirac_train == 1)
312 ff_g723_1_gen_dirac_train(vector, pitch_lag);
313 } else { /* 5300 bps */
314 int cb_gain = ff_g723_1_fixed_cb_gain[subfrm->amp_index];
315 int cb_shift = subfrm->grid_index;
316 int cb_sign = subfrm->pulse_sign;
317 int cb_pos = subfrm->pulse_pos;
318 int offset, beta, lag;
319
320 for (i = 0; i < 8; i += 2) {
321 offset = ((cb_pos & 7) << 3) + cb_shift + i;
322 vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
323 cb_pos >>= 3;
324 cb_sign >>= 1;
325 }
326
327 /* Enhance harmonic components */
328 lag = pitch_contrib[subfrm->ad_cb_gain << 1] + pitch_lag +
329 subfrm->ad_cb_lag - 1;
330 beta = pitch_contrib[(subfrm->ad_cb_gain << 1) + 1];
331
332 if (lag < SUBFRAME_LEN - 2) {
333 for (i = lag; i < SUBFRAME_LEN; i++)
334 vector[i] += beta * vector[i - lag] >> 15;
335 }
336 }
337 }
338
339 /**
340 * Estimate maximum auto-correlation around pitch lag.
341 *
342 * @param buf buffer with offset applied
343 * @param offset offset of the excitation vector
344 * @param ccr_max pointer to the maximum auto-correlation
345 * @param pitch_lag decoded pitch lag
346 * @param length length of autocorrelation
347 * @param dir forward lag(1) / backward lag(-1)
348 */
autocorr_max(const int16_t * buf,int offset,int * ccr_max,int pitch_lag,int length,int dir)349 static int autocorr_max(const int16_t *buf, int offset, int *ccr_max,
350 int pitch_lag, int length, int dir)
351 {
352 int limit, ccr, lag = 0;
353 int i;
354
355 pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
356 if (dir > 0)
357 limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
358 else
359 limit = pitch_lag + 3;
360
361 for (i = pitch_lag - 3; i <= limit; i++) {
362 ccr = ff_g723_1_dot_product(buf, buf + dir * i, length);
363
364 if (ccr > *ccr_max) {
365 *ccr_max = ccr;
366 lag = i;
367 }
368 }
369 return lag;
370 }
371
372 /**
373 * Calculate pitch postfilter optimal and scaling gains.
374 *
375 * @param lag pitch postfilter forward/backward lag
376 * @param ppf pitch postfilter parameters
377 * @param cur_rate current bitrate
378 * @param tgt_eng target energy
379 * @param ccr cross-correlation
380 * @param res_eng residual energy
381 */
comp_ppf_gains(int lag,PPFParam * ppf,enum Rate cur_rate,int tgt_eng,int ccr,int res_eng)382 static void comp_ppf_gains(int lag, PPFParam *ppf, enum Rate cur_rate,
383 int tgt_eng, int ccr, int res_eng)
384 {
385 int pf_residual; /* square of postfiltered residual */
386 int temp1, temp2;
387
388 ppf->index = lag;
389
390 temp1 = tgt_eng * res_eng >> 1;
391 temp2 = ccr * ccr << 1;
392
393 if (temp2 > temp1) {
394 if (ccr >= res_eng) {
395 ppf->opt_gain = ppf_gain_weight[cur_rate];
396 } else {
397 ppf->opt_gain = (ccr << 15) / res_eng *
398 ppf_gain_weight[cur_rate] >> 15;
399 }
400 /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
401 temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
402 temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
403 pf_residual = av_sat_add32(temp1, temp2 + (1 << 15)) >> 16;
404
405 if (tgt_eng >= pf_residual << 1) {
406 temp1 = 0x7fff;
407 } else {
408 temp1 = (tgt_eng << 14) / pf_residual;
409 }
410
411 /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
412 ppf->sc_gain = square_root(temp1 << 16);
413 } else {
414 ppf->opt_gain = 0;
415 ppf->sc_gain = 0x7fff;
416 }
417
418 ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
419 }
420
421 /**
422 * Calculate pitch postfilter parameters.
423 *
424 * @param p the context
425 * @param offset offset of the excitation vector
426 * @param pitch_lag decoded pitch lag
427 * @param ppf pitch postfilter parameters
428 * @param cur_rate current bitrate
429 */
comp_ppf_coeff(G723_1_ChannelContext * p,int offset,int pitch_lag,PPFParam * ppf,enum Rate cur_rate)430 static void comp_ppf_coeff(G723_1_ChannelContext *p, int offset, int pitch_lag,
431 PPFParam *ppf, enum Rate cur_rate)
432 {
433
434 int16_t scale;
435 int i;
436 int temp1, temp2;
437
438 /*
439 * 0 - target energy
440 * 1 - forward cross-correlation
441 * 2 - forward residual energy
442 * 3 - backward cross-correlation
443 * 4 - backward residual energy
444 */
445 int energy[5] = {0, 0, 0, 0, 0};
446 int16_t *buf = p->audio + LPC_ORDER + offset;
447 int fwd_lag = autocorr_max(buf, offset, &energy[1], pitch_lag,
448 SUBFRAME_LEN, 1);
449 int back_lag = autocorr_max(buf, offset, &energy[3], pitch_lag,
450 SUBFRAME_LEN, -1);
451
452 ppf->index = 0;
453 ppf->opt_gain = 0;
454 ppf->sc_gain = 0x7fff;
455
456 /* Case 0, Section 3.6 */
457 if (!back_lag && !fwd_lag)
458 return;
459
460 /* Compute target energy */
461 energy[0] = ff_g723_1_dot_product(buf, buf, SUBFRAME_LEN);
462
463 /* Compute forward residual energy */
464 if (fwd_lag)
465 energy[2] = ff_g723_1_dot_product(buf + fwd_lag, buf + fwd_lag,
466 SUBFRAME_LEN);
467
468 /* Compute backward residual energy */
469 if (back_lag)
470 energy[4] = ff_g723_1_dot_product(buf - back_lag, buf - back_lag,
471 SUBFRAME_LEN);
472
473 /* Normalize and shorten */
474 temp1 = 0;
475 for (i = 0; i < 5; i++)
476 temp1 = FFMAX(energy[i], temp1);
477
478 scale = ff_g723_1_normalize_bits(temp1, 31);
479 for (i = 0; i < 5; i++)
480 energy[i] = (energy[i] << scale) >> 16;
481
482 if (fwd_lag && !back_lag) { /* Case 1 */
483 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
484 energy[2]);
485 } else if (!fwd_lag) { /* Case 2 */
486 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
487 energy[4]);
488 } else { /* Case 3 */
489
490 /*
491 * Select the largest of energy[1]^2/energy[2]
492 * and energy[3]^2/energy[4]
493 */
494 temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
495 temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
496 if (temp1 >= temp2) {
497 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
498 energy[2]);
499 } else {
500 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
501 energy[4]);
502 }
503 }
504 }
505
506 /**
507 * Classify frames as voiced/unvoiced.
508 *
509 * @param p the context
510 * @param pitch_lag decoded pitch_lag
511 * @param exc_eng excitation energy estimation
512 * @param scale scaling factor of exc_eng
513 *
514 * @return residual interpolation index if voiced, 0 otherwise
515 */
comp_interp_index(G723_1_ChannelContext * p,int pitch_lag,int * exc_eng,int * scale)516 static int comp_interp_index(G723_1_ChannelContext *p, int pitch_lag,
517 int *exc_eng, int *scale)
518 {
519 int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
520 int16_t *buf = p->audio + LPC_ORDER;
521
522 int index, ccr, tgt_eng, best_eng, temp;
523
524 *scale = ff_g723_1_scale_vector(buf, p->excitation, FRAME_LEN + PITCH_MAX);
525 buf += offset;
526
527 /* Compute maximum backward cross-correlation */
528 ccr = 0;
529 index = autocorr_max(buf, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
530 ccr = av_sat_add32(ccr, 1 << 15) >> 16;
531
532 /* Compute target energy */
533 tgt_eng = ff_g723_1_dot_product(buf, buf, SUBFRAME_LEN * 2);
534 *exc_eng = av_sat_add32(tgt_eng, 1 << 15) >> 16;
535
536 if (ccr <= 0)
537 return 0;
538
539 /* Compute best energy */
540 best_eng = ff_g723_1_dot_product(buf - index, buf - index,
541 SUBFRAME_LEN * 2);
542 best_eng = av_sat_add32(best_eng, 1 << 15) >> 16;
543
544 temp = best_eng * *exc_eng >> 3;
545
546 if (temp < ccr * ccr) {
547 return index;
548 } else
549 return 0;
550 }
551
552 /**
553 * Perform residual interpolation based on frame classification.
554 *
555 * @param buf decoded excitation vector
556 * @param out output vector
557 * @param lag decoded pitch lag
558 * @param gain interpolated gain
559 * @param rseed seed for random number generator
560 */
residual_interp(int16_t * buf,int16_t * out,int lag,int gain,int * rseed)561 static void residual_interp(int16_t *buf, int16_t *out, int lag,
562 int gain, int *rseed)
563 {
564 int i;
565 if (lag) { /* Voiced */
566 int16_t *vector_ptr = buf + PITCH_MAX;
567 /* Attenuate */
568 for (i = 0; i < lag; i++)
569 out[i] = vector_ptr[i - lag] * 3 >> 2;
570 av_memcpy_backptr((uint8_t*)(out + lag), lag * sizeof(*out),
571 (FRAME_LEN - lag) * sizeof(*out));
572 } else { /* Unvoiced */
573 for (i = 0; i < FRAME_LEN; i++) {
574 *rseed = (int16_t)(*rseed * 521 + 259);
575 out[i] = gain * *rseed >> 15;
576 }
577 memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(*buf));
578 }
579 }
580
581 /**
582 * Perform IIR filtering.
583 *
584 * @param fir_coef FIR coefficients
585 * @param iir_coef IIR coefficients
586 * @param src source vector
587 * @param dest destination vector
588 * @param width width of the output, 16 bits(0) / 32 bits(1)
589 */
590 #define iir_filter(fir_coef, iir_coef, src, dest, width)\
591 {\
592 int m, n;\
593 int res_shift = 16 & ~-(width);\
594 int in_shift = 16 - res_shift;\
595 \
596 for (m = 0; m < SUBFRAME_LEN; m++) {\
597 int64_t filter = 0;\
598 for (n = 1; n <= LPC_ORDER; n++) {\
599 filter -= (fir_coef)[n - 1] * (src)[m - n] -\
600 (iir_coef)[n - 1] * ((dest)[m - n] >> in_shift);\
601 }\
602 \
603 (dest)[m] = av_clipl_int32(((src)[m] * 65536) + (filter * 8) +\
604 (1 << 15)) >> res_shift;\
605 }\
606 }
607
608 /**
609 * Adjust gain of postfiltered signal.
610 *
611 * @param p the context
612 * @param buf postfiltered output vector
613 * @param energy input energy coefficient
614 */
gain_scale(G723_1_ChannelContext * p,int16_t * buf,int energy)615 static void gain_scale(G723_1_ChannelContext *p, int16_t * buf, int energy)
616 {
617 int num, denom, gain, bits1, bits2;
618 int i;
619
620 num = energy;
621 denom = 0;
622 for (i = 0; i < SUBFRAME_LEN; i++) {
623 int temp = buf[i] >> 2;
624 temp *= temp;
625 denom = av_sat_dadd32(denom, temp);
626 }
627
628 if (num && denom) {
629 bits1 = ff_g723_1_normalize_bits(num, 31);
630 bits2 = ff_g723_1_normalize_bits(denom, 31);
631 num = num << bits1 >> 1;
632 denom <<= bits2;
633
634 bits2 = 5 + bits1 - bits2;
635 bits2 = av_clip_uintp2(bits2, 5);
636
637 gain = (num >> 1) / (denom >> 16);
638 gain = square_root(gain << 16 >> bits2);
639 } else {
640 gain = 1 << 12;
641 }
642
643 for (i = 0; i < SUBFRAME_LEN; i++) {
644 p->pf_gain = (15 * p->pf_gain + gain + (1 << 3)) >> 4;
645 buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
646 (1 << 10)) >> 11);
647 }
648 }
649
650 /**
651 * Perform formant filtering.
652 *
653 * @param p the context
654 * @param lpc quantized lpc coefficients
655 * @param buf input buffer
656 * @param dst output buffer
657 */
formant_postfilter(G723_1_ChannelContext * p,int16_t * lpc,int16_t * buf,int16_t * dst)658 static void formant_postfilter(G723_1_ChannelContext *p, int16_t *lpc,
659 int16_t *buf, int16_t *dst)
660 {
661 int16_t filter_coef[2][LPC_ORDER];
662 int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
663 int i, j, k;
664
665 memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(*buf));
666 memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(*filter_signal));
667
668 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
669 for (k = 0; k < LPC_ORDER; k++) {
670 filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
671 (1 << 14)) >> 15;
672 filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
673 (1 << 14)) >> 15;
674 }
675 iir_filter(filter_coef[0], filter_coef[1], buf + i, filter_signal + i, 1);
676 lpc += LPC_ORDER;
677 }
678
679 memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(int16_t));
680 memcpy(p->iir_mem, filter_signal + FRAME_LEN, LPC_ORDER * sizeof(int));
681
682 buf += LPC_ORDER;
683 signal_ptr = filter_signal + LPC_ORDER;
684 for (i = 0; i < SUBFRAMES; i++) {
685 int temp;
686 int auto_corr[2];
687 int scale, energy;
688
689 /* Normalize */
690 scale = ff_g723_1_scale_vector(dst, buf, SUBFRAME_LEN);
691
692 /* Compute auto correlation coefficients */
693 auto_corr[0] = ff_g723_1_dot_product(dst, dst + 1, SUBFRAME_LEN - 1);
694 auto_corr[1] = ff_g723_1_dot_product(dst, dst, SUBFRAME_LEN);
695
696 /* Compute reflection coefficient */
697 temp = auto_corr[1] >> 16;
698 if (temp) {
699 temp = (auto_corr[0] >> 2) / temp;
700 }
701 p->reflection_coef = (3 * p->reflection_coef + temp + 2) >> 2;
702 temp = -p->reflection_coef >> 1 & ~3;
703
704 /* Compensation filter */
705 for (j = 0; j < SUBFRAME_LEN; j++) {
706 dst[j] = av_sat_dadd32(signal_ptr[j],
707 (signal_ptr[j - 1] >> 16) * temp) >> 16;
708 }
709
710 /* Compute normalized signal energy */
711 temp = 2 * scale + 4;
712 if (temp < 0) {
713 energy = av_clipl_int32((int64_t)auto_corr[1] << -temp);
714 } else
715 energy = auto_corr[1] >> temp;
716
717 gain_scale(p, dst, energy);
718
719 buf += SUBFRAME_LEN;
720 signal_ptr += SUBFRAME_LEN;
721 dst += SUBFRAME_LEN;
722 }
723 }
724
sid_gain_to_lsp_index(int gain)725 static int sid_gain_to_lsp_index(int gain)
726 {
727 if (gain < 0x10)
728 return gain << 6;
729 else if (gain < 0x20)
730 return gain - 8 << 7;
731 else
732 return gain - 20 << 8;
733 }
734
cng_rand(int * state,int base)735 static inline int cng_rand(int *state, int base)
736 {
737 *state = (*state * 521 + 259) & 0xFFFF;
738 return (*state & 0x7FFF) * base >> 15;
739 }
740
estimate_sid_gain(G723_1_ChannelContext * p)741 static int estimate_sid_gain(G723_1_ChannelContext *p)
742 {
743 int i, shift, seg, seg2, t, val, val_add, x, y;
744
745 shift = 16 - p->cur_gain * 2;
746 if (shift > 0) {
747 if (p->sid_gain == 0) {
748 t = 0;
749 } else if (shift >= 31 || (int32_t)((uint32_t)p->sid_gain << shift) >> shift != p->sid_gain) {
750 if (p->sid_gain < 0) t = INT32_MIN;
751 else t = INT32_MAX;
752 } else
753 t = p->sid_gain * (1 << shift);
754 } else if(shift < -31) {
755 t = (p->sid_gain < 0) ? -1 : 0;
756 }else
757 t = p->sid_gain >> -shift;
758 x = av_clipl_int32(t * (int64_t)cng_filt[0] >> 16);
759
760 if (x >= cng_bseg[2])
761 return 0x3F;
762
763 if (x >= cng_bseg[1]) {
764 shift = 4;
765 seg = 3;
766 } else {
767 shift = 3;
768 seg = (x >= cng_bseg[0]);
769 }
770 seg2 = FFMIN(seg, 3);
771
772 val = 1 << shift;
773 val_add = val >> 1;
774 for (i = 0; i < shift; i++) {
775 t = seg * 32 + (val << seg2);
776 t *= t;
777 if (x >= t)
778 val += val_add;
779 else
780 val -= val_add;
781 val_add >>= 1;
782 }
783
784 t = seg * 32 + (val << seg2);
785 y = t * t - x;
786 if (y <= 0) {
787 t = seg * 32 + (val + 1 << seg2);
788 t = t * t - x;
789 val = (seg2 - 1) * 16 + val;
790 if (t >= y)
791 val++;
792 } else {
793 t = seg * 32 + (val - 1 << seg2);
794 t = t * t - x;
795 val = (seg2 - 1) * 16 + val;
796 if (t >= y)
797 val--;
798 }
799
800 return val;
801 }
802
generate_noise(G723_1_ChannelContext * p)803 static void generate_noise(G723_1_ChannelContext *p)
804 {
805 int i, j, idx, t;
806 int off[SUBFRAMES];
807 int signs[SUBFRAMES / 2 * 11], pos[SUBFRAMES / 2 * 11];
808 int tmp[SUBFRAME_LEN * 2];
809 int16_t *vector_ptr;
810 int64_t sum;
811 int b0, c, delta, x, shift;
812
813 p->pitch_lag[0] = cng_rand(&p->cng_random_seed, 21) + 123;
814 p->pitch_lag[1] = cng_rand(&p->cng_random_seed, 19) + 123;
815
816 for (i = 0; i < SUBFRAMES; i++) {
817 p->subframe[i].ad_cb_gain = cng_rand(&p->cng_random_seed, 50) + 1;
818 p->subframe[i].ad_cb_lag = cng_adaptive_cb_lag[i];
819 }
820
821 for (i = 0; i < SUBFRAMES / 2; i++) {
822 t = cng_rand(&p->cng_random_seed, 1 << 13);
823 off[i * 2] = t & 1;
824 off[i * 2 + 1] = ((t >> 1) & 1) + SUBFRAME_LEN;
825 t >>= 2;
826 for (j = 0; j < 11; j++) {
827 signs[i * 11 + j] = ((t & 1) * 2 - 1) * (1 << 14);
828 t >>= 1;
829 }
830 }
831
832 idx = 0;
833 for (i = 0; i < SUBFRAMES; i++) {
834 for (j = 0; j < SUBFRAME_LEN / 2; j++)
835 tmp[j] = j;
836 t = SUBFRAME_LEN / 2;
837 for (j = 0; j < pulses[i]; j++, idx++) {
838 int idx2 = cng_rand(&p->cng_random_seed, t);
839
840 pos[idx] = tmp[idx2] * 2 + off[i];
841 tmp[idx2] = tmp[--t];
842 }
843 }
844
845 vector_ptr = p->audio + LPC_ORDER;
846 memcpy(vector_ptr, p->prev_excitation,
847 PITCH_MAX * sizeof(*p->excitation));
848 for (i = 0; i < SUBFRAMES; i += 2) {
849 ff_g723_1_gen_acb_excitation(vector_ptr, vector_ptr,
850 p->pitch_lag[i >> 1], &p->subframe[i],
851 p->cur_rate);
852 ff_g723_1_gen_acb_excitation(vector_ptr + SUBFRAME_LEN,
853 vector_ptr + SUBFRAME_LEN,
854 p->pitch_lag[i >> 1], &p->subframe[i + 1],
855 p->cur_rate);
856
857 t = 0;
858 for (j = 0; j < SUBFRAME_LEN * 2; j++)
859 t |= FFABS(vector_ptr[j]);
860 t = FFMIN(t, 0x7FFF);
861 if (!t) {
862 shift = 0;
863 } else {
864 shift = -10 + av_log2(t);
865 if (shift < -2)
866 shift = -2;
867 }
868 sum = 0;
869 if (shift < 0) {
870 for (j = 0; j < SUBFRAME_LEN * 2; j++) {
871 t = vector_ptr[j] * (1 << -shift);
872 sum += t * t;
873 tmp[j] = t;
874 }
875 } else {
876 for (j = 0; j < SUBFRAME_LEN * 2; j++) {
877 t = vector_ptr[j] >> shift;
878 sum += t * t;
879 tmp[j] = t;
880 }
881 }
882
883 b0 = 0;
884 for (j = 0; j < 11; j++)
885 b0 += tmp[pos[(i / 2) * 11 + j]] * signs[(i / 2) * 11 + j];
886 b0 = b0 * 2 * 2979LL + (1 << 29) >> 30; // approximated division by 11
887
888 c = p->cur_gain * (p->cur_gain * SUBFRAME_LEN >> 5);
889 if (shift * 2 + 3 >= 0)
890 c >>= shift * 2 + 3;
891 else
892 c <<= -(shift * 2 + 3);
893 c = (av_clipl_int32(sum << 1) - c) * 2979LL >> 15;
894
895 delta = b0 * b0 * 2 - c;
896 if (delta <= 0) {
897 x = -b0;
898 } else {
899 delta = square_root(delta);
900 x = delta - b0;
901 t = delta + b0;
902 if (FFABS(t) < FFABS(x))
903 x = -t;
904 }
905 shift++;
906 if (shift < 0)
907 x >>= -shift;
908 else
909 x *= 1 << shift;
910 x = av_clip(x, -10000, 10000);
911
912 for (j = 0; j < 11; j++) {
913 idx = (i / 2) * 11 + j;
914 vector_ptr[pos[idx]] = av_clip_int16(vector_ptr[pos[idx]] +
915 (x * signs[idx] >> 15));
916 }
917
918 /* copy decoded data to serve as a history for the next decoded subframes */
919 memcpy(vector_ptr + PITCH_MAX, vector_ptr,
920 sizeof(*vector_ptr) * SUBFRAME_LEN * 2);
921 vector_ptr += SUBFRAME_LEN * 2;
922 }
923 /* Save the excitation for the next frame */
924 memcpy(p->prev_excitation, p->audio + LPC_ORDER + FRAME_LEN,
925 PITCH_MAX * sizeof(*p->excitation));
926 }
927
g723_1_decode_frame(AVCodecContext * avctx,AVFrame * frame,int * got_frame_ptr,AVPacket * avpkt)928 static int g723_1_decode_frame(AVCodecContext *avctx, AVFrame *frame,
929 int *got_frame_ptr, AVPacket *avpkt)
930 {
931 G723_1_Context *s = avctx->priv_data;
932 const uint8_t *buf = avpkt->data;
933 int buf_size = avpkt->size;
934 int dec_mode = buf[0] & 3;
935 int channels = avctx->ch_layout.nb_channels;
936
937 PPFParam ppf[SUBFRAMES];
938 int16_t cur_lsp[LPC_ORDER];
939 int16_t lpc[SUBFRAMES * LPC_ORDER];
940 int16_t acb_vector[SUBFRAME_LEN];
941 int16_t *out;
942 int bad_frame = 0, i, j, ret;
943
944 if (buf_size < frame_size[dec_mode] * channels) {
945 if (buf_size)
946 av_log(avctx, AV_LOG_WARNING,
947 "Expected %d bytes, got %d - skipping packet\n",
948 frame_size[dec_mode], buf_size);
949 *got_frame_ptr = 0;
950 return buf_size;
951 }
952
953 frame->nb_samples = FRAME_LEN;
954 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
955 return ret;
956
957 for (int ch = 0; ch < channels; ch++) {
958 G723_1_ChannelContext *p = &s->ch[ch];
959 int16_t *audio = p->audio;
960
961 if (unpack_bitstream(p, buf + ch * (buf_size / channels),
962 buf_size / channels) < 0) {
963 bad_frame = 1;
964 if (p->past_frame_type == ACTIVE_FRAME)
965 p->cur_frame_type = ACTIVE_FRAME;
966 else
967 p->cur_frame_type = UNTRANSMITTED_FRAME;
968 }
969
970 out = (int16_t *)frame->extended_data[ch];
971
972 if (p->cur_frame_type == ACTIVE_FRAME) {
973 if (!bad_frame)
974 p->erased_frames = 0;
975 else if (p->erased_frames != 3)
976 p->erased_frames++;
977
978 ff_g723_1_inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
979 ff_g723_1_lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
980
981 /* Save the lsp_vector for the next frame */
982 memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
983
984 /* Generate the excitation for the frame */
985 memcpy(p->excitation, p->prev_excitation,
986 PITCH_MAX * sizeof(*p->excitation));
987 if (!p->erased_frames) {
988 int16_t *vector_ptr = p->excitation + PITCH_MAX;
989
990 /* Update interpolation gain memory */
991 p->interp_gain = ff_g723_1_fixed_cb_gain[(p->subframe[2].amp_index +
992 p->subframe[3].amp_index) >> 1];
993 for (i = 0; i < SUBFRAMES; i++) {
994 gen_fcb_excitation(vector_ptr, &p->subframe[i], p->cur_rate,
995 p->pitch_lag[i >> 1], i);
996 ff_g723_1_gen_acb_excitation(acb_vector,
997 &p->excitation[SUBFRAME_LEN * i],
998 p->pitch_lag[i >> 1],
999 &p->subframe[i], p->cur_rate);
1000 /* Get the total excitation */
1001 for (j = 0; j < SUBFRAME_LEN; j++) {
1002 int v = av_clip_int16(vector_ptr[j] * 2);
1003 vector_ptr[j] = av_clip_int16(v + acb_vector[j]);
1004 }
1005 vector_ptr += SUBFRAME_LEN;
1006 }
1007
1008 vector_ptr = p->excitation + PITCH_MAX;
1009
1010 p->interp_index = comp_interp_index(p, p->pitch_lag[1],
1011 &p->sid_gain, &p->cur_gain);
1012
1013 /* Perform pitch postfiltering */
1014 if (s->postfilter) {
1015 i = PITCH_MAX;
1016 for (j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1017 comp_ppf_coeff(p, i, p->pitch_lag[j >> 1],
1018 ppf + j, p->cur_rate);
1019
1020 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1021 ff_acelp_weighted_vector_sum(p->audio + LPC_ORDER + i,
1022 vector_ptr + i,
1023 vector_ptr + i + ppf[j].index,
1024 ppf[j].sc_gain,
1025 ppf[j].opt_gain,
1026 1 << 14, 15, SUBFRAME_LEN);
1027 } else {
1028 audio = vector_ptr - LPC_ORDER;
1029 }
1030
1031 /* Save the excitation for the next frame */
1032 memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
1033 PITCH_MAX * sizeof(*p->excitation));
1034 } else {
1035 p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
1036 if (p->erased_frames == 3) {
1037 /* Mute output */
1038 memset(p->excitation, 0,
1039 (FRAME_LEN + PITCH_MAX) * sizeof(*p->excitation));
1040 memset(p->prev_excitation, 0,
1041 PITCH_MAX * sizeof(*p->excitation));
1042 memset(frame->data[0], 0,
1043 (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
1044 } else {
1045 int16_t *buf = p->audio + LPC_ORDER;
1046
1047 /* Regenerate frame */
1048 residual_interp(p->excitation, buf, p->interp_index,
1049 p->interp_gain, &p->random_seed);
1050
1051 /* Save the excitation for the next frame */
1052 memcpy(p->prev_excitation, buf + (FRAME_LEN - PITCH_MAX),
1053 PITCH_MAX * sizeof(*p->excitation));
1054 }
1055 }
1056 p->cng_random_seed = CNG_RANDOM_SEED;
1057 } else {
1058 if (p->cur_frame_type == SID_FRAME) {
1059 p->sid_gain = sid_gain_to_lsp_index(p->subframe[0].amp_index);
1060 ff_g723_1_inverse_quant(p->sid_lsp, p->prev_lsp, p->lsp_index, 0);
1061 } else if (p->past_frame_type == ACTIVE_FRAME) {
1062 p->sid_gain = estimate_sid_gain(p);
1063 }
1064
1065 if (p->past_frame_type == ACTIVE_FRAME)
1066 p->cur_gain = p->sid_gain;
1067 else
1068 p->cur_gain = (p->cur_gain * 7 + p->sid_gain) >> 3;
1069 generate_noise(p);
1070 ff_g723_1_lsp_interpolate(lpc, p->sid_lsp, p->prev_lsp);
1071 /* Save the lsp_vector for the next frame */
1072 memcpy(p->prev_lsp, p->sid_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1073 }
1074
1075 p->past_frame_type = p->cur_frame_type;
1076
1077 memcpy(p->audio, p->synth_mem, LPC_ORDER * sizeof(*p->audio));
1078 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1079 ff_celp_lp_synthesis_filter(p->audio + i, &lpc[j * LPC_ORDER],
1080 audio + i, SUBFRAME_LEN, LPC_ORDER,
1081 0, 1, 1 << 12);
1082 memcpy(p->synth_mem, p->audio + FRAME_LEN, LPC_ORDER * sizeof(*p->audio));
1083
1084 if (s->postfilter) {
1085 formant_postfilter(p, lpc, p->audio, out);
1086 } else { // if output is not postfiltered it should be scaled by 2
1087 for (i = 0; i < FRAME_LEN; i++)
1088 out[i] = av_clip_int16(2 * p->audio[LPC_ORDER + i]);
1089 }
1090 }
1091
1092 *got_frame_ptr = 1;
1093
1094 return frame_size[dec_mode] * channels;
1095 }
1096
1097 #define OFFSET(x) offsetof(G723_1_Context, x)
1098 #define AD AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1099
1100 static const AVOption options[] = {
1101 { "postfilter", "enable postfilter", OFFSET(postfilter), AV_OPT_TYPE_BOOL,
1102 { .i64 = 1 }, 0, 1, AD },
1103 { NULL }
1104 };
1105
1106
1107 static const AVClass g723_1dec_class = {
1108 .class_name = "G.723.1 decoder",
1109 .item_name = av_default_item_name,
1110 .option = options,
1111 .version = LIBAVUTIL_VERSION_INT,
1112 };
1113
1114 const FFCodec ff_g723_1_decoder = {
1115 .p.name = "g723_1",
1116 .p.long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1117 .p.type = AVMEDIA_TYPE_AUDIO,
1118 .p.id = AV_CODEC_ID_G723_1,
1119 .priv_data_size = sizeof(G723_1_Context),
1120 .init = g723_1_decode_init,
1121 FF_CODEC_DECODE_CB(g723_1_decode_frame),
1122 .p.capabilities = AV_CODEC_CAP_SUBFRAMES | AV_CODEC_CAP_DR1,
1123 .p.priv_class = &g723_1dec_class,
1124 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE,
1125 };
1126