projects / libav.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
g723_1: Move sharable functions to a separate file
[libav.git] / libavcodec / g723_1dec.c
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 Libav.
7 *
8 * Libav 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 * Libav 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 Libav; 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 #define BITSTREAM_READER_LE
29 #include "libavutil/channel_layout.h"
30 #include "libavutil/mem.h"
31 #include "libavutil/opt.h"
32 #include "avcodec.h"
33 #include "get_bits.h"
34 #include "acelp_vectors.h"
35 #include "celp_filters.h"
36 #include "g723_1.h"
37 #include "internal.h"
38
39 #define CNG_RANDOM_SEED 12345
40
41 static av_cold int g723_1_decode_init(AVCodecContext *avctx)
42 {
43 G723_1_Context *p = avctx->priv_data;
44
45 avctx->channel_layout = AV_CH_LAYOUT_MONO;
46 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
47 avctx->channels = 1;
48 avctx->sample_rate = 8000;
49 p->pf_gain = 1 << 12;
50
51 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
52 memcpy(p->sid_lsp, dc_lsp, LPC_ORDER * sizeof(*p->sid_lsp));
53
54 p->cng_random_seed = CNG_RANDOM_SEED;
55 p->past_frame_type = SID_FRAME;
56
57 return 0;
58 }
59
60 /**
61 * Unpack the frame into parameters.
62 *
63 * @param p the context
64 * @param buf pointer to the input buffer
65 * @param buf_size size of the input buffer
66 */
67 static int unpack_bitstream(G723_1_Context *p, const uint8_t *buf,
68 int buf_size)
69 {
70 GetBitContext gb;
71 int ad_cb_len;
72 int temp, info_bits, i;
73
74 init_get_bits(&gb, buf, buf_size * 8);
75
76 /* Extract frame type and rate info */
77 info_bits = get_bits(&gb, 2);
78
79 if (info_bits == 3) {
80 p->cur_frame_type = UNTRANSMITTED_FRAME;
81 return 0;
82 }
83
84 /* Extract 24 bit lsp indices, 8 bit for each band */
85 p->lsp_index[2] = get_bits(&gb, 8);
86 p->lsp_index[1] = get_bits(&gb, 8);
87 p->lsp_index[0] = get_bits(&gb, 8);
88
89 if (info_bits == 2) {
90 p->cur_frame_type = SID_FRAME;
91 p->subframe[0].amp_index = get_bits(&gb, 6);
92 return 0;
93 }
94
95 /* Extract the info common to both rates */
96 p->cur_rate = info_bits ? RATE_5300 : RATE_6300;
97 p->cur_frame_type = ACTIVE_FRAME;
98
99 p->pitch_lag[0] = get_bits(&gb, 7);
100 if (p->pitch_lag[0] > 123) /* test if forbidden code */
101 return -1;
102 p->pitch_lag[0] += PITCH_MIN;
103 p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
104
105 p->pitch_lag[1] = get_bits(&gb, 7);
106 if (p->pitch_lag[1] > 123)
107 return -1;
108 p->pitch_lag[1] += PITCH_MIN;
109 p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
110 p->subframe[0].ad_cb_lag = 1;
111 p->subframe[2].ad_cb_lag = 1;
112
113 for (i = 0; i < SUBFRAMES; i++) {
114 /* Extract combined gain */
115 temp = get_bits(&gb, 12);
116 ad_cb_len = 170;
117 p->subframe[i].dirac_train = 0;
118 if (p->cur_rate == RATE_6300 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
119 p->subframe[i].dirac_train = temp >> 11;
120 temp &= 0x7FF;
121 ad_cb_len = 85;
122 }
123 p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
124 if (p->subframe[i].ad_cb_gain < ad_cb_len) {
125 p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
126 GAIN_LEVELS;
127 } else {
128 return -1;
129 }
130 }
131
132 p->subframe[0].grid_index = get_bits(&gb, 1);
133 p->subframe[1].grid_index = get_bits(&gb, 1);
134 p->subframe[2].grid_index = get_bits(&gb, 1);
135 p->subframe[3].grid_index = get_bits(&gb, 1);
136
137 if (p->cur_rate == RATE_6300) {
138 skip_bits(&gb, 1); /* skip reserved bit */
139
140 /* Compute pulse_pos index using the 13-bit combined position index */
141 temp = get_bits(&gb, 13);
142 p->subframe[0].pulse_pos = temp / 810;
143
144 temp -= p->subframe[0].pulse_pos * 810;
145 p->subframe[1].pulse_pos = FASTDIV(temp, 90);
146
147 temp -= p->subframe[1].pulse_pos * 90;
148 p->subframe[2].pulse_pos = FASTDIV(temp, 9);
149 p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
150
151 p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
152 get_bits(&gb, 16);
153 p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
154 get_bits(&gb, 14);
155 p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
156 get_bits(&gb, 16);
157 p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
158 get_bits(&gb, 14);
159
160 p->subframe[0].pulse_sign = get_bits(&gb, 6);
161 p->subframe[1].pulse_sign = get_bits(&gb, 5);
162 p->subframe[2].pulse_sign = get_bits(&gb, 6);
163 p->subframe[3].pulse_sign = get_bits(&gb, 5);
164 } else { /* 5300 bps */
165 p->subframe[0].pulse_pos = get_bits(&gb, 12);
166 p->subframe[1].pulse_pos = get_bits(&gb, 12);
167 p->subframe[2].pulse_pos = get_bits(&gb, 12);
168 p->subframe[3].pulse_pos = get_bits(&gb, 12);
169
170 p->subframe[0].pulse_sign = get_bits(&gb, 4);
171 p->subframe[1].pulse_sign = get_bits(&gb, 4);
172 p->subframe[2].pulse_sign = get_bits(&gb, 4);
173 p->subframe[3].pulse_sign = get_bits(&gb, 4);
174 }
175
176 return 0;
177 }
178
179 /**
180 * Bitexact implementation of sqrt(val/2).
181 */
182 static int16_t square_root(int val)
183 {
184 int16_t res = 0;
185 int16_t exp = 0x4000;
186 int i;
187
188 for (i = 0; i < 14; i ++) {
189 int res_exp = res + exp;
190 if (val >= res_exp * res_exp << 1)
191 res += exp;
192 exp >>= 1;
193 }
194 return res;
195 }
196
197 /**
198 * Bitexact implementation of 2ab scaled by 1/2^16.
199 *
200 * @param a 32 bit multiplicand
201 * @param b 16 bit multiplier
202 */
203 #define MULL2(a, b) \
204 ((((a) >> 16) * (b) << 1) + (((a) & 0xffff) * (b) >> 15))
205
206 /**
207 * Generate fixed codebook excitation vector.
208 *
209 * @param vector decoded excitation vector
210 * @param subfrm current subframe
211 * @param cur_rate current bitrate
212 * @param pitch_lag closed loop pitch lag
213 * @param index current subframe index
214 */
215 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe *subfrm,
216 enum Rate cur_rate, int pitch_lag, int index)
217 {
218 int temp, i, j;
219
220 memset(vector, 0, SUBFRAME_LEN * sizeof(*vector));
221
222 if (cur_rate == RATE_6300) {
223 if (subfrm->pulse_pos >= max_pos[index])
224 return;
225
226 /* Decode amplitudes and positions */
227 j = PULSE_MAX - pulses[index];
228 temp = subfrm->pulse_pos;
229 for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
230 temp -= combinatorial_table[j][i];
231 if (temp >= 0)
232 continue;
233 temp += combinatorial_table[j++][i];
234 if (subfrm->pulse_sign & (1 << (PULSE_MAX - j))) {
235 vector[subfrm->grid_index + GRID_SIZE * i] =
236 -fixed_cb_gain[subfrm->amp_index];
237 } else {
238 vector[subfrm->grid_index + GRID_SIZE * i] =
239 fixed_cb_gain[subfrm->amp_index];
240 }
241 if (j == PULSE_MAX)
242 break;
243 }
244 if (subfrm->dirac_train == 1)
245 ff_g723_1_gen_dirac_train(vector, pitch_lag);
246 } else { /* 5300 bps */
247 int cb_gain = fixed_cb_gain[subfrm->amp_index];
248 int cb_shift = subfrm->grid_index;
249 int cb_sign = subfrm->pulse_sign;
250 int cb_pos = subfrm->pulse_pos;
251 int offset, beta, lag;
252
253 for (i = 0; i < 8; i += 2) {
254 offset = ((cb_pos & 7) << 3) + cb_shift + i;
255 vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
256 cb_pos >>= 3;
257 cb_sign >>= 1;
258 }
259
260 /* Enhance harmonic components */
261 lag = pitch_contrib[subfrm->ad_cb_gain << 1] + pitch_lag +
262 subfrm->ad_cb_lag - 1;
263 beta = pitch_contrib[(subfrm->ad_cb_gain << 1) + 1];
264
265 if (lag < SUBFRAME_LEN - 2) {
266 for (i = lag; i < SUBFRAME_LEN; i++)
267 vector[i] += beta * vector[i - lag] >> 15;
268 }
269 }
270 }
271
272 /**
273 * Estimate maximum auto-correlation around pitch lag.
274 *
275 * @param buf buffer with offset applied
276 * @param offset offset of the excitation vector
277 * @param ccr_max pointer to the maximum auto-correlation
278 * @param pitch_lag decoded pitch lag
279 * @param length length of autocorrelation
280 * @param dir forward lag(1) / backward lag(-1)
281 */
282 static int autocorr_max(const int16_t *buf, int offset, int *ccr_max,
283 int pitch_lag, int length, int dir)
284 {
285 int limit, ccr, lag = 0;
286 int i;
287
288 pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
289 if (dir > 0)
290 limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
291 else
292 limit = pitch_lag + 3;
293
294 for (i = pitch_lag - 3; i <= limit; i++) {
295 ccr = ff_g723_1_dot_product(buf, buf + dir * i, length);
296
297 if (ccr > *ccr_max) {
298 *ccr_max = ccr;
299 lag = i;
300 }
301 }
302 return lag;
303 }
304
305 /**
306 * Calculate pitch postfilter optimal and scaling gains.
307 *
308 * @param lag pitch postfilter forward/backward lag
309 * @param ppf pitch postfilter parameters
310 * @param cur_rate current bitrate
311 * @param tgt_eng target energy
312 * @param ccr cross-correlation
313 * @param res_eng residual energy
314 */
315 static void comp_ppf_gains(int lag, PPFParam *ppf, enum Rate cur_rate,
316 int tgt_eng, int ccr, int res_eng)
317 {
318 int pf_residual; /* square of postfiltered residual */
319 int temp1, temp2;
320
321 ppf->index = lag;
322
323 temp1 = tgt_eng * res_eng >> 1;
324 temp2 = ccr * ccr << 1;
325
326 if (temp2 > temp1) {
327 if (ccr >= res_eng) {
328 ppf->opt_gain = ppf_gain_weight[cur_rate];
329 } else {
330 ppf->opt_gain = (ccr << 15) / res_eng *
331 ppf_gain_weight[cur_rate] >> 15;
332 }
333 /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
334 temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
335 temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
336 pf_residual = av_sat_add32(temp1, temp2 + (1 << 15)) >> 16;
337
338 if (tgt_eng >= pf_residual << 1) {
339 temp1 = 0x7fff;
340 } else {
341 temp1 = (tgt_eng << 14) / pf_residual;
342 }
343
344 /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
345 ppf->sc_gain = square_root(temp1 << 16);
346 } else {
347 ppf->opt_gain = 0;
348 ppf->sc_gain = 0x7fff;
349 }
350
351 ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
352 }
353
354 /**
355 * Calculate pitch postfilter parameters.
356 *
357 * @param p the context
358 * @param offset offset of the excitation vector
359 * @param pitch_lag decoded pitch lag
360 * @param ppf pitch postfilter parameters
361 * @param cur_rate current bitrate
362 */
363 static void comp_ppf_coeff(G723_1_Context *p, int offset, int pitch_lag,
364 PPFParam *ppf, enum Rate cur_rate)
365 {
366
367 int16_t scale;
368 int i;
369 int temp1, temp2;
370
371 /*
372 * 0 - target energy
373 * 1 - forward cross-correlation
374 * 2 - forward residual energy
375 * 3 - backward cross-correlation
376 * 4 - backward residual energy
377 */
378 int energy[5] = {0, 0, 0, 0, 0};
379 int16_t *buf = p->audio + LPC_ORDER + offset;
380 int fwd_lag = autocorr_max(buf, offset, &energy[1], pitch_lag,
381 SUBFRAME_LEN, 1);
382 int back_lag = autocorr_max(buf, offset, &energy[3], pitch_lag,
383 SUBFRAME_LEN, -1);
384
385 ppf->index = 0;
386 ppf->opt_gain = 0;
387 ppf->sc_gain = 0x7fff;
388
389 /* Case 0, Section 3.6 */
390 if (!back_lag && !fwd_lag)
391 return;
392
393 /* Compute target energy */
394 energy[0] = ff_g723_1_dot_product(buf, buf, SUBFRAME_LEN);
395
396 /* Compute forward residual energy */
397 if (fwd_lag)
398 energy[2] = ff_g723_1_dot_product(buf + fwd_lag, buf + fwd_lag,
399 SUBFRAME_LEN);
400
401 /* Compute backward residual energy */
402 if (back_lag)
403 energy[4] = ff_g723_1_dot_product(buf - back_lag, buf - back_lag,
404 SUBFRAME_LEN);
405
406 /* Normalize and shorten */
407 temp1 = 0;
408 for (i = 0; i < 5; i++)
409 temp1 = FFMAX(energy[i], temp1);
410
411 scale = ff_g723_1_normalize_bits(temp1, 31);
412 for (i = 0; i < 5; i++)
413 energy[i] = (energy[i] << scale) >> 16;
414
415 if (fwd_lag && !back_lag) { /* Case 1 */
416 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
417 energy[2]);
418 } else if (!fwd_lag) { /* Case 2 */
419 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
420 energy[4]);
421 } else { /* Case 3 */
422
423 /*
424 * Select the largest of energy[1]^2/energy[2]
425 * and energy[3]^2/energy[4]
426 */
427 temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
428 temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
429 if (temp1 >= temp2) {
430 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
431 energy[2]);
432 } else {
433 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
434 energy[4]);
435 }
436 }
437 }
438
439 /**
440 * Classify frames as voiced/unvoiced.
441 *
442 * @param p the context
443 * @param pitch_lag decoded pitch_lag
444 * @param exc_eng excitation energy estimation
445 * @param scale scaling factor of exc_eng
446 *
447 * @return residual interpolation index if voiced, 0 otherwise
448 */
449 static int comp_interp_index(G723_1_Context *p, int pitch_lag,
450 int *exc_eng, int *scale)
451 {
452 int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
453 int16_t *buf = p->audio + LPC_ORDER;
454
455 int index, ccr, tgt_eng, best_eng, temp;
456
457 *scale = ff_g723_1_scale_vector(buf, p->excitation, FRAME_LEN + PITCH_MAX);
458 buf += offset;
459
460 /* Compute maximum backward cross-correlation */
461 ccr = 0;
462 index = autocorr_max(buf, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
463 ccr = av_sat_add32(ccr, 1 << 15) >> 16;
464
465 /* Compute target energy */
466 tgt_eng = ff_g723_1_dot_product(buf, buf, SUBFRAME_LEN * 2);
467 *exc_eng = av_sat_add32(tgt_eng, 1 << 15) >> 16;
468
469 if (ccr <= 0)
470 return 0;
471
472 /* Compute best energy */
473 best_eng = ff_g723_1_dot_product(buf - index, buf - index,
474 SUBFRAME_LEN * 2);
475 best_eng = av_sat_add32(best_eng, 1 << 15) >> 16;
476
477 temp = best_eng * *exc_eng >> 3;
478
479 if (temp < ccr * ccr)
480 return index;
481 else
482 return 0;
483 }
484
485 /**
486 * Peform residual interpolation based on frame classification.
487 *
488 * @param buf decoded excitation vector
489 * @param out output vector
490 * @param lag decoded pitch lag
491 * @param gain interpolated gain
492 * @param rseed seed for random number generator
493 */
494 static void residual_interp(int16_t *buf, int16_t *out, int lag,
495 int gain, int *rseed)
496 {
497 int i;
498 if (lag) { /* Voiced */
499 int16_t *vector_ptr = buf + PITCH_MAX;
500 /* Attenuate */
501 for (i = 0; i < lag; i++)
502 out[i] = vector_ptr[i - lag] * 3 >> 2;
503 av_memcpy_backptr((uint8_t*)(out + lag), lag * sizeof(*out),
504 (FRAME_LEN - lag) * sizeof(*out));
505 } else { /* Unvoiced */
506 for (i = 0; i < FRAME_LEN; i++) {
507 *rseed = *rseed * 521 + 259;
508 out[i] = gain * *rseed >> 15;
509 }
510 memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(*buf));
511 }
512 }
513
514 /**
515 * Perform IIR filtering.
516 *
517 * @param fir_coef FIR coefficients
518 * @param iir_coef IIR coefficients
519 * @param src source vector
520 * @param dest destination vector
521 */
522 static void iir_filter(int16_t *fir_coef, int16_t *iir_coef,
523 int16_t *src, int *dest)
524 {
525 int m, n;
526
527 for (m = 0; m < SUBFRAME_LEN; m++) {
528 int64_t filter = 0;
529 for (n = 1; n <= LPC_ORDER; n++) {
530 filter -= fir_coef[n - 1] * src[m - n] -
531 iir_coef[n - 1] * (dest[m - n] >> 16);
532 }
533
534 dest[m] = av_clipl_int32((src[m] << 16) + (filter << 3) + (1 << 15));
535 }
536 }
537
538 /**
539 * Adjust gain of postfiltered signal.
540 *
541 * @param p the context
542 * @param buf postfiltered output vector
543 * @param energy input energy coefficient
544 */
545 static void gain_scale(G723_1_Context *p, int16_t * buf, int energy)
546 {
547 int num, denom, gain, bits1, bits2;
548 int i;
549
550 num = energy;
551 denom = 0;
552 for (i = 0; i < SUBFRAME_LEN; i++) {
553 int temp = buf[i] >> 2;
554 temp *= temp;
555 denom = av_sat_dadd32(denom, temp);
556 }
557
558 if (num && denom) {
559 bits1 = ff_g723_1_normalize_bits(num, 31);
560 bits2 = ff_g723_1_normalize_bits(denom, 31);
561 num = num << bits1 >> 1;
562 denom <<= bits2;
563
564 bits2 = 5 + bits1 - bits2;
565 bits2 = FFMAX(0, bits2);
566
567 gain = (num >> 1) / (denom >> 16);
568 gain = square_root(gain << 16 >> bits2);
569 } else {
570 gain = 1 << 12;
571 }
572
573 for (i = 0; i < SUBFRAME_LEN; i++) {
574 p->pf_gain = (15 * p->pf_gain + gain + (1 << 3)) >> 4;
575 buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
576 (1 << 10)) >> 11);
577 }
578 }
579
580 /**
581 * Perform formant filtering.
582 *
583 * @param p the context
584 * @param lpc quantized lpc coefficients
585 * @param buf input buffer
586 * @param dst output buffer
587 */
588 static void formant_postfilter(G723_1_Context *p, int16_t *lpc,
589 int16_t *buf, int16_t *dst)
590 {
591 int16_t filter_coef[2][LPC_ORDER];
592 int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
593 int i, j, k;
594
595 memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(*buf));
596 memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(*filter_signal));
597
598 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
599 for (k = 0; k < LPC_ORDER; k++) {
600 filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
601 (1 << 14)) >> 15;
602 filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
603 (1 << 14)) >> 15;
604 }
605 iir_filter(filter_coef[0], filter_coef[1], buf + i, filter_signal + i);
606 lpc += LPC_ORDER;
607 }
608
609 memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(*p->fir_mem));
610 memcpy(p->iir_mem, filter_signal + FRAME_LEN,
611 LPC_ORDER * sizeof(*p->iir_mem));
612
613 buf += LPC_ORDER;
614 signal_ptr = filter_signal + LPC_ORDER;
615 for (i = 0; i < SUBFRAMES; i++) {
616 int temp;
617 int auto_corr[2];
618 int scale, energy;
619
620 /* Normalize */
621 scale = ff_g723_1_scale_vector(dst, buf, SUBFRAME_LEN);
622
623 /* Compute auto correlation coefficients */
624 auto_corr[0] = ff_g723_1_dot_product(dst, dst + 1, SUBFRAME_LEN - 1);
625 auto_corr[1] = ff_g723_1_dot_product(dst, dst, SUBFRAME_LEN);
626
627 /* Compute reflection coefficient */
628 temp = auto_corr[1] >> 16;
629 if (temp) {
630 temp = (auto_corr[0] >> 2) / temp;
631 }
632 p->reflection_coef = (3 * p->reflection_coef + temp + 2) >> 2;
633 temp = -p->reflection_coef >> 1 & ~3;
634
635 /* Compensation filter */
636 for (j = 0; j < SUBFRAME_LEN; j++) {
637 dst[j] = av_sat_dadd32(signal_ptr[j],
638 (signal_ptr[j - 1] >> 16) * temp) >> 16;
639 }
640
641 /* Compute normalized signal energy */
642 temp = 2 * scale + 4;
643 if (temp < 0) {
644 energy = av_clipl_int32((int64_t)auto_corr[1] << -temp);
645 } else
646 energy = auto_corr[1] >> temp;
647
648 gain_scale(p, dst, energy);
649
650 buf += SUBFRAME_LEN;
651 signal_ptr += SUBFRAME_LEN;
652 dst += SUBFRAME_LEN;
653 }
654 }
655
656 static int sid_gain_to_lsp_index(int gain)
657 {
658 if (gain < 0x10)
659 return gain << 6;
660 else if (gain < 0x20)
661 return gain - 8 << 7;
662 else
663 return gain - 20 << 8;
664 }
665
666 static inline int cng_rand(int *state, int base)
667 {
668 *state = (*state * 521 + 259) & 0xFFFF;
669 return (*state & 0x7FFF) * base >> 15;
670 }
671
672 static int estimate_sid_gain(G723_1_Context *p)
673 {
674 int i, shift, seg, seg2, t, val, val_add, x, y;
675
676 shift = 16 - p->cur_gain * 2;
677 if (shift > 0)
678 t = p->sid_gain << shift;
679 else
680 t = p->sid_gain >> -shift;
681 x = t * cng_filt[0] >> 16;
682
683 if (x >= cng_bseg[2])
684 return 0x3F;
685
686 if (x >= cng_bseg[1]) {
687 shift = 4;
688 seg = 3;
689 } else {
690 shift = 3;
691 seg = (x >= cng_bseg[0]);
692 }
693 seg2 = FFMIN(seg, 3);
694
695 val = 1 << shift;
696 val_add = val >> 1;
697 for (i = 0; i < shift; i++) {
698 t = seg * 32 + (val << seg2);
699 t *= t;
700 if (x >= t)
701 val += val_add;
702 else
703 val -= val_add;
704 val_add >>= 1;
705 }
706
707 t = seg * 32 + (val << seg2);
708 y = t * t - x;
709 if (y <= 0) {
710 t = seg * 32 + (val + 1 << seg2);
711 t = t * t - x;
712 val = (seg2 - 1 << 4) + val;
713 if (t >= y)
714 val++;
715 } else {
716 t = seg * 32 + (val - 1 << seg2);
717 t = t * t - x;
718 val = (seg2 - 1 << 4) + val;
719 if (t >= y)
720 val--;
721 }
722
723 return val;
724 }
725
726 static void generate_noise(G723_1_Context *p)
727 {
728 int i, j, idx, t;
729 int off[SUBFRAMES];
730 int signs[SUBFRAMES / 2 * 11], pos[SUBFRAMES / 2 * 11];
731 int tmp[SUBFRAME_LEN * 2];
732 int16_t *vector_ptr;
733 int64_t sum;
734 int b0, c, delta, x, shift;
735
736 p->pitch_lag[0] = cng_rand(&p->cng_random_seed, 21) + 123;
737 p->pitch_lag[1] = cng_rand(&p->cng_random_seed, 19) + 123;
738
739 for (i = 0; i < SUBFRAMES; i++) {
740 p->subframe[i].ad_cb_gain = cng_rand(&p->cng_random_seed, 50) + 1;
741 p->subframe[i].ad_cb_lag = cng_adaptive_cb_lag[i];
742 }
743
744 for (i = 0; i < SUBFRAMES / 2; i++) {
745 t = cng_rand(&p->cng_random_seed, 1 << 13);
746 off[i * 2] = t & 1;
747 off[i * 2 + 1] = ((t >> 1) & 1) + SUBFRAME_LEN;
748 t >>= 2;
749 for (j = 0; j < 11; j++) {
750 signs[i * 11 + j] = (t & 1) * 2 - 1 << 14;
751 t >>= 1;
752 }
753 }
754
755 idx = 0;
756 for (i = 0; i < SUBFRAMES; i++) {
757 for (j = 0; j < SUBFRAME_LEN / 2; j++)
758 tmp[j] = j;
759 t = SUBFRAME_LEN / 2;
760 for (j = 0; j < pulses[i]; j++, idx++) {
761 int idx2 = cng_rand(&p->cng_random_seed, t);
762
763 pos[idx] = tmp[idx2] * 2 + off[i];
764 tmp[idx2] = tmp[--t];
765 }
766 }
767
768 vector_ptr = p->audio + LPC_ORDER;
769 memcpy(vector_ptr, p->prev_excitation,
770 PITCH_MAX * sizeof(*p->excitation));
771 for (i = 0; i < SUBFRAMES; i += 2) {
772 ff_g723_1_gen_acb_excitation(vector_ptr, vector_ptr,
773 p->pitch_lag[i >> 1], &p->subframe[i],
774 p->cur_rate);
775 ff_g723_1_gen_acb_excitation(vector_ptr + SUBFRAME_LEN,
776 vector_ptr + SUBFRAME_LEN,
777 p->pitch_lag[i >> 1], &p->subframe[i + 1],
778 p->cur_rate);
779
780 t = 0;
781 for (j = 0; j < SUBFRAME_LEN * 2; j++)
782 t |= FFABS(vector_ptr[j]);
783 t = FFMIN(t, 0x7FFF);
784 if (!t) {
785 shift = 0;
786 } else {
787 shift = -10 + av_log2(t);
788 if (shift < -2)
789 shift = -2;
790 }
791 sum = 0;
792 if (shift < 0) {
793 for (j = 0; j < SUBFRAME_LEN * 2; j++) {
794 t = vector_ptr[j] << -shift;
795 sum += t * t;
796 tmp[j] = t;
797 }
798 } else {
799 for (j = 0; j < SUBFRAME_LEN * 2; j++) {
800 t = vector_ptr[j] >> shift;
801 sum += t * t;
802 tmp[j] = t;
803 }
804 }
805
806 b0 = 0;
807 for (j = 0; j < 11; j++)
808 b0 += tmp[pos[(i / 2) * 11 + j]] * signs[(i / 2) * 11 + j];
809 b0 = b0 * 2 * 2979LL + (1 << 29) >> 30; // approximated division by 11
810
811 c = p->cur_gain * (p->cur_gain * SUBFRAME_LEN >> 5);
812 if (shift * 2 + 3 >= 0)
813 c >>= shift * 2 + 3;
814 else
815 c <<= -(shift * 2 + 3);
816 c = (av_clipl_int32(sum << 1) - c) * 2979LL >> 15;
817
818 delta = b0 * b0 * 2 - c;
819 if (delta <= 0) {
820 x = -b0;
821 } else {
822 delta = square_root(delta);
823 x = delta - b0;
824 t = delta + b0;
825 if (FFABS(t) < FFABS(x))
826 x = -t;
827 }
828 shift++;
829 if (shift < 0)
830 x >>= -shift;
831 else
832 x <<= shift;
833 x = av_clip(x, -10000, 10000);
834
835 for (j = 0; j < 11; j++) {
836 idx = (i / 2) * 11 + j;
837 vector_ptr[pos[idx]] = av_clip_int16(vector_ptr[pos[idx]] +
838 (x * signs[idx] >> 15));
839 }
840
841 /* copy decoded data to serve as a history for the next decoded subframes */
842 memcpy(vector_ptr + PITCH_MAX, vector_ptr,
843 sizeof(*vector_ptr) * SUBFRAME_LEN * 2);
844 vector_ptr += SUBFRAME_LEN * 2;
845 }
846 /* Save the excitation for the next frame */
847 memcpy(p->prev_excitation, p->audio + LPC_ORDER + FRAME_LEN,
848 PITCH_MAX * sizeof(*p->excitation));
849 }
850
851 static int g723_1_decode_frame(AVCodecContext *avctx, void *data,
852 int *got_frame_ptr, AVPacket *avpkt)
853 {
854 G723_1_Context *p = avctx->priv_data;
855 AVFrame *frame = data;
856 const uint8_t *buf = avpkt->data;
857 int buf_size = avpkt->size;
858 int dec_mode = buf[0] & 3;
859
860 PPFParam ppf[SUBFRAMES];
861 int16_t cur_lsp[LPC_ORDER];
862 int16_t lpc[SUBFRAMES * LPC_ORDER];
863 int16_t acb_vector[SUBFRAME_LEN];
864 int16_t *out;
865 int bad_frame = 0, i, j, ret;
866 int16_t *audio = p->audio;
867
868 if (buf_size < frame_size[dec_mode]) {
869 if (buf_size)
870 av_log(avctx, AV_LOG_WARNING,
871 "Expected %d bytes, got %d - skipping packet\n",
872 frame_size[dec_mode], buf_size);
873 *got_frame_ptr = 0;
874 return buf_size;
875 }
876
877 if (unpack_bitstream(p, buf, buf_size) < 0) {
878 bad_frame = 1;
879 if (p->past_frame_type == ACTIVE_FRAME)
880 p->cur_frame_type = ACTIVE_FRAME;
881 else
882 p->cur_frame_type = UNTRANSMITTED_FRAME;
883 }
884
885 frame->nb_samples = FRAME_LEN;
886 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0) {
887 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
888 return ret;
889 }
890
891 out = (int16_t *)frame->data[0];
892
893 if (p->cur_frame_type == ACTIVE_FRAME) {
894 if (!bad_frame)
895 p->erased_frames = 0;
896 else if (p->erased_frames != 3)
897 p->erased_frames++;
898
899 ff_g723_1_inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
900 ff_g723_1_lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
901
902 /* Save the lsp_vector for the next frame */
903 memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
904
905 /* Generate the excitation for the frame */
906 memcpy(p->excitation, p->prev_excitation,
907 PITCH_MAX * sizeof(*p->excitation));
908 if (!p->erased_frames) {
909 int16_t *vector_ptr = p->excitation + PITCH_MAX;
910
911 /* Update interpolation gain memory */
912 p->interp_gain = fixed_cb_gain[(p->subframe[2].amp_index +
913 p->subframe[3].amp_index) >> 1];
914 for (i = 0; i < SUBFRAMES; i++) {
915 gen_fcb_excitation(vector_ptr, &p->subframe[i], p->cur_rate,
916 p->pitch_lag[i >> 1], i);
917 ff_g723_1_gen_acb_excitation(acb_vector,
918 &p->excitation[SUBFRAME_LEN * i],
919 p->pitch_lag[i >> 1],
920 &p->subframe[i], p->cur_rate);
921 /* Get the total excitation */
922 for (j = 0; j < SUBFRAME_LEN; j++) {
923 int v = av_clip_int16(vector_ptr[j] << 1);
924 vector_ptr[j] = av_clip_int16(v + acb_vector[j]);
925 }
926 vector_ptr += SUBFRAME_LEN;
927 }
928
929 vector_ptr = p->excitation + PITCH_MAX;
930
931 p->interp_index = comp_interp_index(p, p->pitch_lag[1],
932 &p->sid_gain, &p->cur_gain);
933
934 /* Peform pitch postfiltering */
935 if (p->postfilter) {
936 i = PITCH_MAX;
937 for (j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
938 comp_ppf_coeff(p, i, p->pitch_lag[j >> 1],
939 ppf + j, p->cur_rate);
940
941 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
942 ff_acelp_weighted_vector_sum(p->audio + LPC_ORDER + i,
943 vector_ptr + i,
944 vector_ptr + i + ppf[j].index,
945 ppf[j].sc_gain,
946 ppf[j].opt_gain,
947 1 << 14, 15, SUBFRAME_LEN);
948 } else {
949 audio = vector_ptr - LPC_ORDER;
950 }
951
952 /* Save the excitation for the next frame */
953 memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
954 PITCH_MAX * sizeof(*p->excitation));
955 } else {
956 p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
957 if (p->erased_frames == 3) {
958 /* Mute output */
959 memset(p->excitation, 0,
960 (FRAME_LEN + PITCH_MAX) * sizeof(*p->excitation));
961 memset(p->prev_excitation, 0,
962 PITCH_MAX * sizeof(*p->excitation));
963 memset(frame->data[0], 0,
964 (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
965 } else {
966 int16_t *buf = p->audio + LPC_ORDER;
967
968 /* Regenerate frame */
969 residual_interp(p->excitation, buf, p->interp_index,
970 p->interp_gain, &p->random_seed);
971
972 /* Save the excitation for the next frame */
973 memcpy(p->prev_excitation, buf + (FRAME_LEN - PITCH_MAX),
974 PITCH_MAX * sizeof(*p->excitation));
975 }
976 }
977 p->cng_random_seed = CNG_RANDOM_SEED;
978 } else {
979 if (p->cur_frame_type == SID_FRAME) {
980 p->sid_gain = sid_gain_to_lsp_index(p->subframe[0].amp_index);
981 ff_g723_1_inverse_quant(p->sid_lsp, p->prev_lsp, p->lsp_index, 0);
982 } else if (p->past_frame_type == ACTIVE_FRAME) {
983 p->sid_gain = estimate_sid_gain(p);
984 }
985
986 if (p->past_frame_type == ACTIVE_FRAME)
987 p->cur_gain = p->sid_gain;
988 else
989 p->cur_gain = (p->cur_gain * 7 + p->sid_gain) >> 3;
990 generate_noise(p);
991 ff_g723_1_lsp_interpolate(lpc, p->sid_lsp, p->prev_lsp);
992 /* Save the lsp_vector for the next frame */
993 memcpy(p->prev_lsp, p->sid_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
994 }
995
996 p->past_frame_type = p->cur_frame_type;
997
998 memcpy(p->audio, p->synth_mem, LPC_ORDER * sizeof(*p->audio));
999 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1000 ff_celp_lp_synthesis_filter(p->audio + i, &lpc[j * LPC_ORDER],
1001 audio + i, SUBFRAME_LEN, LPC_ORDER,
1002 0, 1, 1 << 12);
1003 memcpy(p->synth_mem, p->audio + FRAME_LEN, LPC_ORDER * sizeof(*p->audio));
1004
1005 if (p->postfilter) {
1006 formant_postfilter(p, lpc, p->audio, out);
1007 } else { // if output is not postfiltered it should be scaled by 2
1008 for (i = 0; i < FRAME_LEN; i++)
1009 out[i] = av_clip_int16(p->audio[LPC_ORDER + i] << 1);
1010 }
1011
1012 *got_frame_ptr = 1;
1013
1014 return frame_size[dec_mode];
1015 }
1016
1017 #define OFFSET(x) offsetof(G723_1_Context, x)
1018 #define AD AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1019
1020 static const AVOption options[] = {
1021 { "postfilter", "postfilter on/off", OFFSET(postfilter), AV_OPT_TYPE_INT,
1022 { .i64 = 1 }, 0, 1, AD },
1023 { NULL }
1024 };
1025
1026
1027 static const AVClass g723_1dec_class = {
1028 .class_name = "G.723.1 decoder",
1029 .item_name = av_default_item_name,
1030 .option = options,
1031 .version = LIBAVUTIL_VERSION_INT,
1032 };
1033
1034 AVCodec ff_g723_1_decoder = {
1035 .name = "g723_1",
1036 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1037 .type = AVMEDIA_TYPE_AUDIO,
1038 .id = AV_CODEC_ID_G723_1,
1039 .priv_data_size = sizeof(G723_1_Context),
1040 .init = g723_1_decode_init,
1041 .decode = g723_1_decode_frame,
1042 .capabilities = AV_CODEC_CAP_SUBFRAMES | AV_CODEC_CAP_DR1,
1043 .priv_class = &g723_1dec_class,
1044 };
Libav master