G.723.1 demuxer and decoder
[libav.git] / libavcodec / g723_1.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/audioconvert.h"
30 #include "libavutil/lzo.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 "celp_math.h"
37 #include "lsp.h"
38 #include "g723_1_data.h"
39
40 /**
41 * G723.1 frame types
42 */
43 enum FrameType {
44 ACTIVE_FRAME, ///< Active speech
45 SID_FRAME, ///< Silence Insertion Descriptor frame
46 UNTRANSMITTED_FRAME
47 };
48
49 enum Rate {
50 RATE_6300,
51 RATE_5300
52 };
53
54 /**
55 * G723.1 unpacked data subframe
56 */
57 typedef struct {
58 int ad_cb_lag; ///< adaptive codebook lag
59 int ad_cb_gain;
60 int dirac_train;
61 int pulse_sign;
62 int grid_index;
63 int amp_index;
64 int pulse_pos;
65 } G723_1_Subframe;
66
67 /**
68 * Pitch postfilter parameters
69 */
70 typedef struct {
71 int index; ///< postfilter backward/forward lag
72 int16_t opt_gain; ///< optimal gain
73 int16_t sc_gain; ///< scaling gain
74 } PPFParam;
75
76 typedef struct g723_1_context {
77 AVClass *class;
78 AVFrame frame;
79
80 G723_1_Subframe subframe[4];
81 enum FrameType cur_frame_type;
82 enum FrameType past_frame_type;
83 enum Rate cur_rate;
84 uint8_t lsp_index[LSP_BANDS];
85 int pitch_lag[2];
86 int erased_frames;
87
88 int16_t prev_lsp[LPC_ORDER];
89 int16_t prev_excitation[PITCH_MAX];
90 int16_t excitation[PITCH_MAX + FRAME_LEN];
91 int16_t synth_mem[LPC_ORDER];
92 int16_t fir_mem[LPC_ORDER];
93 int iir_mem[LPC_ORDER];
94
95 int random_seed;
96 int interp_index;
97 int interp_gain;
98 int sid_gain;
99 int cur_gain;
100 int reflection_coef;
101 int pf_gain;
102 int postfilter;
103
104 int16_t audio[FRAME_LEN + LPC_ORDER];
105 } G723_1_Context;
106
107 static av_cold int g723_1_decode_init(AVCodecContext *avctx)
108 {
109 G723_1_Context *p = avctx->priv_data;
110
111 avctx->channel_layout = AV_CH_LAYOUT_MONO;
112 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
113 avctx->channels = 1;
114 avctx->sample_rate = 8000;
115 p->pf_gain = 1 << 12;
116
117 avcodec_get_frame_defaults(&p->frame);
118 avctx->coded_frame = &p->frame;
119
120 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
121
122 return 0;
123 }
124
125 /**
126 * Unpack the frame into parameters.
127 *
128 * @param p the context
129 * @param buf pointer to the input buffer
130 * @param buf_size size of the input buffer
131 */
132 static int unpack_bitstream(G723_1_Context *p, const uint8_t *buf,
133 int buf_size)
134 {
135 GetBitContext gb;
136 int ad_cb_len;
137 int temp, info_bits, i;
138
139 init_get_bits(&gb, buf, buf_size * 8);
140
141 /* Extract frame type and rate info */
142 info_bits = get_bits(&gb, 2);
143
144 if (info_bits == 3) {
145 p->cur_frame_type = UNTRANSMITTED_FRAME;
146 return 0;
147 }
148
149 /* Extract 24 bit lsp indices, 8 bit for each band */
150 p->lsp_index[2] = get_bits(&gb, 8);
151 p->lsp_index[1] = get_bits(&gb, 8);
152 p->lsp_index[0] = get_bits(&gb, 8);
153
154 if (info_bits == 2) {
155 p->cur_frame_type = SID_FRAME;
156 p->subframe[0].amp_index = get_bits(&gb, 6);
157 return 0;
158 }
159
160 /* Extract the info common to both rates */
161 p->cur_rate = info_bits ? RATE_5300 : RATE_6300;
162 p->cur_frame_type = ACTIVE_FRAME;
163
164 p->pitch_lag[0] = get_bits(&gb, 7);
165 if (p->pitch_lag[0] > 123) /* test if forbidden code */
166 return -1;
167 p->pitch_lag[0] += PITCH_MIN;
168 p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
169
170 p->pitch_lag[1] = get_bits(&gb, 7);
171 if (p->pitch_lag[1] > 123)
172 return -1;
173 p->pitch_lag[1] += PITCH_MIN;
174 p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
175 p->subframe[0].ad_cb_lag = 1;
176 p->subframe[2].ad_cb_lag = 1;
177
178 for (i = 0; i < SUBFRAMES; i++) {
179 /* Extract combined gain */
180 temp = get_bits(&gb, 12);
181 ad_cb_len = 170;
182 p->subframe[i].dirac_train = 0;
183 if (p->cur_rate == RATE_6300 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
184 p->subframe[i].dirac_train = temp >> 11;
185 temp &= 0x7FF;
186 ad_cb_len = 85;
187 }
188 p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
189 if (p->subframe[i].ad_cb_gain < ad_cb_len) {
190 p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
191 GAIN_LEVELS;
192 } else {
193 return -1;
194 }
195 }
196
197 p->subframe[0].grid_index = get_bits(&gb, 1);
198 p->subframe[1].grid_index = get_bits(&gb, 1);
199 p->subframe[2].grid_index = get_bits(&gb, 1);
200 p->subframe[3].grid_index = get_bits(&gb, 1);
201
202 if (p->cur_rate == RATE_6300) {
203 skip_bits(&gb, 1); /* skip reserved bit */
204
205 /* Compute pulse_pos index using the 13-bit combined position index */
206 temp = get_bits(&gb, 13);
207 p->subframe[0].pulse_pos = temp / 810;
208
209 temp -= p->subframe[0].pulse_pos * 810;
210 p->subframe[1].pulse_pos = FASTDIV(temp, 90);
211
212 temp -= p->subframe[1].pulse_pos * 90;
213 p->subframe[2].pulse_pos = FASTDIV(temp, 9);
214 p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
215
216 p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
217 get_bits(&gb, 16);
218 p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
219 get_bits(&gb, 14);
220 p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
221 get_bits(&gb, 16);
222 p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
223 get_bits(&gb, 14);
224
225 p->subframe[0].pulse_sign = get_bits(&gb, 6);
226 p->subframe[1].pulse_sign = get_bits(&gb, 5);
227 p->subframe[2].pulse_sign = get_bits(&gb, 6);
228 p->subframe[3].pulse_sign = get_bits(&gb, 5);
229 } else { /* 5300 bps */
230 p->subframe[0].pulse_pos = get_bits(&gb, 12);
231 p->subframe[1].pulse_pos = get_bits(&gb, 12);
232 p->subframe[2].pulse_pos = get_bits(&gb, 12);
233 p->subframe[3].pulse_pos = get_bits(&gb, 12);
234
235 p->subframe[0].pulse_sign = get_bits(&gb, 4);
236 p->subframe[1].pulse_sign = get_bits(&gb, 4);
237 p->subframe[2].pulse_sign = get_bits(&gb, 4);
238 p->subframe[3].pulse_sign = get_bits(&gb, 4);
239 }
240
241 return 0;
242 }
243
244 /**
245 * Bitexact implementation of sqrt(val/2).
246 */
247 static int16_t square_root(int val)
248 {
249 int16_t res = 0;
250 int16_t exp = 0x4000;
251 int i;
252
253 for (i = 0; i < 14; i ++) {
254 int res_exp = res + exp;
255 if (val >= res_exp * res_exp << 1)
256 res += exp;
257 exp >>= 1;
258 }
259 return res;
260 }
261
262 /**
263 * Calculate the number of left-shifts required for normalizing the input.
264 *
265 * @param num input number
266 * @param width width of the input, 16 bits(0) / 32 bits(1)
267 */
268 static int normalize_bits(int num, int width)
269 {
270 if (!num)
271 return 0;
272 if (num == -1)
273 return width;
274 if (num < 0)
275 num = ~num;
276
277 return width - av_log2(num);
278 }
279
280 /**
281 * Scale vector contents based on the largest of their absolutes.
282 */
283 static int scale_vector(int16_t *vector, int length)
284 {
285 int bits, scale, max = 0;
286 int i;
287
288
289 for (i = 0; i < length; i++)
290 max = FFMAX(max, FFABS(vector[i]));
291
292 bits = normalize_bits(max, 15);
293 scale = (bits == 15) ? 0x7FFF : (1 << bits);
294
295 for (i = 0; i < length; i++)
296 vector[i] = (vector[i] * scale) >> 4;
297
298 return bits - 3;
299 }
300
301 /**
302 * Perform inverse quantization of LSP frequencies.
303 *
304 * @param cur_lsp the current LSP vector
305 * @param prev_lsp the previous LSP vector
306 * @param lsp_index VQ indices
307 * @param bad_frame bad frame flag
308 */
309 static void inverse_quant(int16_t *cur_lsp, int16_t *prev_lsp,
310 uint8_t *lsp_index, int bad_frame)
311 {
312 int min_dist, pred;
313 int i, j, temp, stable;
314
315 /* Check for frame erasure */
316 if (!bad_frame) {
317 min_dist = 0x100;
318 pred = 12288;
319 } else {
320 min_dist = 0x200;
321 pred = 23552;
322 lsp_index[0] = lsp_index[1] = lsp_index[2] = 0;
323 }
324
325 /* Get the VQ table entry corresponding to the transmitted index */
326 cur_lsp[0] = lsp_band0[lsp_index[0]][0];
327 cur_lsp[1] = lsp_band0[lsp_index[0]][1];
328 cur_lsp[2] = lsp_band0[lsp_index[0]][2];
329 cur_lsp[3] = lsp_band1[lsp_index[1]][0];
330 cur_lsp[4] = lsp_band1[lsp_index[1]][1];
331 cur_lsp[5] = lsp_band1[lsp_index[1]][2];
332 cur_lsp[6] = lsp_band2[lsp_index[2]][0];
333 cur_lsp[7] = lsp_band2[lsp_index[2]][1];
334 cur_lsp[8] = lsp_band2[lsp_index[2]][2];
335 cur_lsp[9] = lsp_band2[lsp_index[2]][3];
336
337 /* Add predicted vector & DC component to the previously quantized vector */
338 for (i = 0; i < LPC_ORDER; i++) {
339 temp = ((prev_lsp[i] - dc_lsp[i]) * pred + (1 << 14)) >> 15;
340 cur_lsp[i] += dc_lsp[i] + temp;
341 }
342
343 for (i = 0; i < LPC_ORDER; i++) {
344 cur_lsp[0] = FFMAX(cur_lsp[0], 0x180);
345 cur_lsp[LPC_ORDER - 1] = FFMIN(cur_lsp[LPC_ORDER - 1], 0x7e00);
346
347 /* Stability check */
348 for (j = 1; j < LPC_ORDER; j++) {
349 temp = min_dist + cur_lsp[j - 1] - cur_lsp[j];
350 if (temp > 0) {
351 temp >>= 1;
352 cur_lsp[j - 1] -= temp;
353 cur_lsp[j] += temp;
354 }
355 }
356 stable = 1;
357 for (j = 1; j < LPC_ORDER; j++) {
358 temp = cur_lsp[j - 1] + min_dist - cur_lsp[j] - 4;
359 if (temp > 0) {
360 stable = 0;
361 break;
362 }
363 }
364 if (stable)
365 break;
366 }
367 if (!stable)
368 memcpy(cur_lsp, prev_lsp, LPC_ORDER * sizeof(*cur_lsp));
369 }
370
371 /**
372 * Bitexact implementation of 2ab scaled by 1/2^16.
373 *
374 * @param a 32 bit multiplicand
375 * @param b 16 bit multiplier
376 */
377 #define MULL2(a, b) \
378 ((((a) >> 16) * (b) << 1) + (((a) & 0xffff) * (b) >> 15))
379
380 /**
381 * Convert LSP frequencies to LPC coefficients.
382 *
383 * @param lpc buffer for LPC coefficients
384 */
385 static void lsp2lpc(int16_t *lpc)
386 {
387 int f1[LPC_ORDER / 2 + 1];
388 int f2[LPC_ORDER / 2 + 1];
389 int i, j;
390
391 /* Calculate negative cosine */
392 for (j = 0; j < LPC_ORDER; j++) {
393 int index = lpc[j] >> 7;
394 int offset = lpc[j] & 0x7f;
395 int64_t temp1 = cos_tab[index] << 16;
396 int temp2 = (cos_tab[index + 1] - cos_tab[index]) *
397 ((offset << 8) + 0x80) << 1;
398
399 lpc[j] = -(av_clipl_int32(((temp1 + temp2) << 1) + (1 << 15)) >> 16);
400 }
401
402 /*
403 * Compute sum and difference polynomial coefficients
404 * (bitexact alternative to lsp2poly() in lsp.c)
405 */
406 /* Initialize with values in Q28 */
407 f1[0] = 1 << 28;
408 f1[1] = (lpc[0] << 14) + (lpc[2] << 14);
409 f1[2] = lpc[0] * lpc[2] + (2 << 28);
410
411 f2[0] = 1 << 28;
412 f2[1] = (lpc[1] << 14) + (lpc[3] << 14);
413 f2[2] = lpc[1] * lpc[3] + (2 << 28);
414
415 /*
416 * Calculate and scale the coefficients by 1/2 in
417 * each iteration for a final scaling factor of Q25
418 */
419 for (i = 2; i < LPC_ORDER / 2; i++) {
420 f1[i + 1] = f1[i - 1] + MULL2(f1[i], lpc[2 * i]);
421 f2[i + 1] = f2[i - 1] + MULL2(f2[i], lpc[2 * i + 1]);
422
423 for (j = i; j >= 2; j--) {
424 f1[j] = MULL2(f1[j - 1], lpc[2 * i]) +
425 (f1[j] >> 1) + (f1[j - 2] >> 1);
426 f2[j] = MULL2(f2[j - 1], lpc[2 * i + 1]) +
427 (f2[j] >> 1) + (f2[j - 2] >> 1);
428 }
429
430 f1[0] >>= 1;
431 f2[0] >>= 1;
432 f1[1] = ((lpc[2 * i] << 16 >> i) + f1[1]) >> 1;
433 f2[1] = ((lpc[2 * i + 1] << 16 >> i) + f2[1]) >> 1;
434 }
435
436 /* Convert polynomial coefficients to LPC coefficients */
437 for (i = 0; i < LPC_ORDER / 2; i++) {
438 int64_t ff1 = f1[i + 1] + f1[i];
439 int64_t ff2 = f2[i + 1] - f2[i];
440
441 lpc[i] = av_clipl_int32(((ff1 + ff2) << 3) + (1 << 15)) >> 16;
442 lpc[LPC_ORDER - i - 1] = av_clipl_int32(((ff1 - ff2) << 3) +
443 (1 << 15)) >> 16;
444 }
445 }
446
447 /**
448 * Quantize LSP frequencies by interpolation and convert them to
449 * the corresponding LPC coefficients.
450 *
451 * @param lpc buffer for LPC coefficients
452 * @param cur_lsp the current LSP vector
453 * @param prev_lsp the previous LSP vector
454 */
455 static void lsp_interpolate(int16_t *lpc, int16_t *cur_lsp, int16_t *prev_lsp)
456 {
457 int i;
458 int16_t *lpc_ptr = lpc;
459
460 /* cur_lsp * 0.25 + prev_lsp * 0.75 */
461 ff_acelp_weighted_vector_sum(lpc, cur_lsp, prev_lsp,
462 4096, 12288, 1 << 13, 14, LPC_ORDER);
463 ff_acelp_weighted_vector_sum(lpc + LPC_ORDER, cur_lsp, prev_lsp,
464 8192, 8192, 1 << 13, 14, LPC_ORDER);
465 ff_acelp_weighted_vector_sum(lpc + 2 * LPC_ORDER, cur_lsp, prev_lsp,
466 12288, 4096, 1 << 13, 14, LPC_ORDER);
467 memcpy(lpc + 3 * LPC_ORDER, cur_lsp, LPC_ORDER * sizeof(*lpc));
468
469 for (i = 0; i < SUBFRAMES; i++) {
470 lsp2lpc(lpc_ptr);
471 lpc_ptr += LPC_ORDER;
472 }
473 }
474
475 /**
476 * Generate a train of dirac functions with period as pitch lag.
477 */
478 static void gen_dirac_train(int16_t *buf, int pitch_lag)
479 {
480 int16_t vector[SUBFRAME_LEN];
481 int i, j;
482
483 memcpy(vector, buf, SUBFRAME_LEN * sizeof(*vector));
484 for (i = pitch_lag; i < SUBFRAME_LEN; i += pitch_lag) {
485 for (j = 0; j < SUBFRAME_LEN - i; j++)
486 buf[i + j] += vector[j];
487 }
488 }
489
490 /**
491 * Generate fixed codebook excitation vector.
492 *
493 * @param vector decoded excitation vector
494 * @param subfrm current subframe
495 * @param cur_rate current bitrate
496 * @param pitch_lag closed loop pitch lag
497 * @param index current subframe index
498 */
499 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe subfrm,
500 enum Rate cur_rate, int pitch_lag, int index)
501 {
502 int temp, i, j;
503
504 memset(vector, 0, SUBFRAME_LEN * sizeof(*vector));
505
506 if (cur_rate == RATE_6300) {
507 if (subfrm.pulse_pos >= max_pos[index])
508 return;
509
510 /* Decode amplitudes and positions */
511 j = PULSE_MAX - pulses[index];
512 temp = subfrm.pulse_pos;
513 for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
514 temp -= combinatorial_table[j][i];
515 if (temp >= 0)
516 continue;
517 temp += combinatorial_table[j++][i];
518 if (subfrm.pulse_sign & (1 << (PULSE_MAX - j))) {
519 vector[subfrm.grid_index + GRID_SIZE * i] =
520 -fixed_cb_gain[subfrm.amp_index];
521 } else {
522 vector[subfrm.grid_index + GRID_SIZE * i] =
523 fixed_cb_gain[subfrm.amp_index];
524 }
525 if (j == PULSE_MAX)
526 break;
527 }
528 if (subfrm.dirac_train == 1)
529 gen_dirac_train(vector, pitch_lag);
530 } else { /* 5300 bps */
531 int cb_gain = fixed_cb_gain[subfrm.amp_index];
532 int cb_shift = subfrm.grid_index;
533 int cb_sign = subfrm.pulse_sign;
534 int cb_pos = subfrm.pulse_pos;
535 int offset, beta, lag;
536
537 for (i = 0; i < 8; i += 2) {
538 offset = ((cb_pos & 7) << 3) + cb_shift + i;
539 vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
540 cb_pos >>= 3;
541 cb_sign >>= 1;
542 }
543
544 /* Enhance harmonic components */
545 lag = pitch_contrib[subfrm.ad_cb_gain << 1] + pitch_lag +
546 subfrm.ad_cb_lag - 1;
547 beta = pitch_contrib[(subfrm.ad_cb_gain << 1) + 1];
548
549 if (lag < SUBFRAME_LEN - 2) {
550 for (i = lag; i < SUBFRAME_LEN; i++)
551 vector[i] += beta * vector[i - lag] >> 15;
552 }
553 }
554 }
555
556 /**
557 * Get delayed contribution from the previous excitation vector.
558 */
559 static void get_residual(int16_t *residual, int16_t *prev_excitation, int lag)
560 {
561 int offset = PITCH_MAX - PITCH_ORDER / 2 - lag;
562 int i;
563
564 residual[0] = prev_excitation[offset];
565 residual[1] = prev_excitation[offset + 1];
566
567 offset += 2;
568 for (i = 2; i < SUBFRAME_LEN + PITCH_ORDER - 1; i++)
569 residual[i] = prev_excitation[offset + (i - 2) % lag];
570 }
571
572 static int dot_product(const int16_t *a, const int16_t *b, int length,
573 int shift)
574 {
575 int i, sum = 0;
576
577 for (i = 0; i < length; i++) {
578 int64_t prod = av_clipl_int32(MUL64(a[i], b[i]) << shift);
579 sum = av_clipl_int32(sum + prod);
580 }
581 return sum;
582 }
583
584 /**
585 * Generate adaptive codebook excitation.
586 */
587 static void gen_acb_excitation(int16_t *vector, int16_t *prev_excitation,
588 int pitch_lag, G723_1_Subframe subfrm,
589 enum Rate cur_rate)
590 {
591 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
592 const int16_t *cb_ptr;
593 int lag = pitch_lag + subfrm.ad_cb_lag - 1;
594
595 int i;
596 int64_t sum;
597
598 get_residual(residual, prev_excitation, lag);
599
600 /* Select quantization table */
601 if (cur_rate == RATE_6300 && pitch_lag < SUBFRAME_LEN - 2)
602 cb_ptr = adaptive_cb_gain85;
603 else
604 cb_ptr = adaptive_cb_gain170;
605
606 /* Calculate adaptive vector */
607 cb_ptr += subfrm.ad_cb_gain * 20;
608 for (i = 0; i < SUBFRAME_LEN; i++) {
609 sum = dot_product(residual + i, cb_ptr, PITCH_ORDER, 1);
610 vector[i] = av_clipl_int32((sum << 1) + (1 << 15)) >> 16;
611 }
612 }
613
614 /**
615 * Estimate maximum auto-correlation around pitch lag.
616 *
617 * @param p the context
618 * @param offset offset of the excitation vector
619 * @param ccr_max pointer to the maximum auto-correlation
620 * @param pitch_lag decoded pitch lag
621 * @param length length of autocorrelation
622 * @param dir forward lag(1) / backward lag(-1)
623 */
624 static int autocorr_max(G723_1_Context *p, int offset, int *ccr_max,
625 int pitch_lag, int length, int dir)
626 {
627 int limit, ccr, lag = 0;
628 int16_t *buf = p->excitation + offset;
629 int i;
630
631 pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
632 limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
633
634 for (i = pitch_lag - 3; i <= limit; i++) {
635 ccr = dot_product(buf, buf + dir * i, length, 1);
636
637 if (ccr > *ccr_max) {
638 *ccr_max = ccr;
639 lag = i;
640 }
641 }
642 return lag;
643 }
644
645 /**
646 * Calculate pitch postfilter optimal and scaling gains.
647 *
648 * @param lag pitch postfilter forward/backward lag
649 * @param ppf pitch postfilter parameters
650 * @param cur_rate current bitrate
651 * @param tgt_eng target energy
652 * @param ccr cross-correlation
653 * @param res_eng residual energy
654 */
655 static void comp_ppf_gains(int lag, PPFParam *ppf, enum Rate cur_rate,
656 int tgt_eng, int ccr, int res_eng)
657 {
658 int pf_residual; /* square of postfiltered residual */
659 int64_t temp1, temp2;
660
661 ppf->index = lag;
662
663 temp1 = tgt_eng * res_eng >> 1;
664 temp2 = ccr * ccr << 1;
665
666 if (temp2 > temp1) {
667 if (ccr >= res_eng) {
668 ppf->opt_gain = ppf_gain_weight[cur_rate];
669 } else {
670 ppf->opt_gain = (ccr << 15) / res_eng *
671 ppf_gain_weight[cur_rate] >> 15;
672 }
673 /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
674 temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
675 temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
676 pf_residual = av_clipl_int32(temp1 + temp2 + (1 << 15)) >> 16;
677
678 if (tgt_eng >= pf_residual << 1) {
679 temp1 = 0x7fff;
680 } else {
681 temp1 = (tgt_eng << 14) / pf_residual;
682 }
683
684 /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
685 ppf->sc_gain = square_root(temp1 << 16);
686 } else {
687 ppf->opt_gain = 0;
688 ppf->sc_gain = 0x7fff;
689 }
690
691 ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
692 }
693
694 /**
695 * Calculate pitch postfilter parameters.
696 *
697 * @param p the context
698 * @param offset offset of the excitation vector
699 * @param pitch_lag decoded pitch lag
700 * @param ppf pitch postfilter parameters
701 * @param cur_rate current bitrate
702 */
703 static void comp_ppf_coeff(G723_1_Context *p, int offset, int pitch_lag,
704 PPFParam *ppf, enum Rate cur_rate)
705 {
706
707 int16_t scale;
708 int i;
709 int64_t temp1, temp2;
710
711 /*
712 * 0 - target energy
713 * 1 - forward cross-correlation
714 * 2 - forward residual energy
715 * 3 - backward cross-correlation
716 * 4 - backward residual energy
717 */
718 int energy[5] = {0, 0, 0, 0, 0};
719 int16_t *buf = p->excitation + offset;
720 int fwd_lag = autocorr_max(p, offset, &energy[1], pitch_lag,
721 SUBFRAME_LEN, 1);
722 int back_lag = autocorr_max(p, offset, &energy[3], pitch_lag,
723 SUBFRAME_LEN, -1);
724
725 ppf->index = 0;
726 ppf->opt_gain = 0;
727 ppf->sc_gain = 0x7fff;
728
729 /* Case 0, Section 3.6 */
730 if (!back_lag && !fwd_lag)
731 return;
732
733 /* Compute target energy */
734 energy[0] = dot_product(buf, buf, SUBFRAME_LEN, 1);
735
736 /* Compute forward residual energy */
737 if (fwd_lag)
738 energy[2] = dot_product(buf + fwd_lag, buf + fwd_lag,
739 SUBFRAME_LEN, 1);
740
741 /* Compute backward residual energy */
742 if (back_lag)
743 energy[4] = dot_product(buf - back_lag, buf - back_lag,
744 SUBFRAME_LEN, 1);
745
746 /* Normalize and shorten */
747 temp1 = 0;
748 for (i = 0; i < 5; i++)
749 temp1 = FFMAX(energy[i], temp1);
750
751 scale = normalize_bits(temp1, 31);
752 for (i = 0; i < 5; i++)
753 energy[i] = (energy[i] << scale) >> 16;
754
755 if (fwd_lag && !back_lag) { /* Case 1 */
756 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
757 energy[2]);
758 } else if (!fwd_lag) { /* Case 2 */
759 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
760 energy[4]);
761 } else { /* Case 3 */
762
763 /*
764 * Select the largest of energy[1]^2/energy[2]
765 * and energy[3]^2/energy[4]
766 */
767 temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
768 temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
769 if (temp1 >= temp2) {
770 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
771 energy[2]);
772 } else {
773 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
774 energy[4]);
775 }
776 }
777 }
778
779 /**
780 * Classify frames as voiced/unvoiced.
781 *
782 * @param p the context
783 * @param pitch_lag decoded pitch_lag
784 * @param exc_eng excitation energy estimation
785 * @param scale scaling factor of exc_eng
786 *
787 * @return residual interpolation index if voiced, 0 otherwise
788 */
789 static int comp_interp_index(G723_1_Context *p, int pitch_lag,
790 int *exc_eng, int *scale)
791 {
792 int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
793 int16_t *buf = p->excitation + offset;
794
795 int index, ccr, tgt_eng, best_eng, temp;
796
797 *scale = scale_vector(p->excitation, FRAME_LEN + PITCH_MAX);
798
799 /* Compute maximum backward cross-correlation */
800 ccr = 0;
801 index = autocorr_max(p, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
802 ccr = av_clipl_int32((int64_t)ccr + (1 << 15)) >> 16;
803
804 /* Compute target energy */
805 tgt_eng = dot_product(buf, buf, SUBFRAME_LEN * 2, 1);
806 *exc_eng = av_clipl_int32((int64_t)tgt_eng + (1 << 15)) >> 16;
807
808 if (ccr <= 0)
809 return 0;
810
811 /* Compute best energy */
812 best_eng = dot_product(buf - index, buf - index,
813 SUBFRAME_LEN * 2, 1);
814 best_eng = av_clipl_int32((int64_t)best_eng + (1 << 15)) >> 16;
815
816 temp = best_eng * *exc_eng >> 3;
817
818 if (temp < ccr * ccr)
819 return index;
820 else
821 return 0;
822 }
823
824 /**
825 * Peform residual interpolation based on frame classification.
826 *
827 * @param buf decoded excitation vector
828 * @param out output vector
829 * @param lag decoded pitch lag
830 * @param gain interpolated gain
831 * @param rseed seed for random number generator
832 */
833 static void residual_interp(int16_t *buf, int16_t *out, int lag,
834 int gain, int *rseed)
835 {
836 int i;
837 if (lag) { /* Voiced */
838 int16_t *vector_ptr = buf + PITCH_MAX;
839 /* Attenuate */
840 for (i = 0; i < lag; i++)
841 vector_ptr[i - lag] = vector_ptr[i - lag] * 3 >> 2;
842 av_memcpy_backptr((uint8_t*)vector_ptr, lag * sizeof(*vector_ptr),
843 FRAME_LEN * sizeof(*vector_ptr));
844 memcpy(out, vector_ptr, FRAME_LEN * sizeof(*vector_ptr));
845 } else { /* Unvoiced */
846 for (i = 0; i < FRAME_LEN; i++) {
847 *rseed = *rseed * 521 + 259;
848 out[i] = gain * *rseed >> 15;
849 }
850 memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(*buf));
851 }
852 }
853
854 /**
855 * Perform IIR filtering.
856 *
857 * @param fir_coef FIR coefficients
858 * @param iir_coef IIR coefficients
859 * @param src source vector
860 * @param dest destination vector
861 */
862 static inline void iir_filter(int16_t *fir_coef, int16_t *iir_coef,
863 int16_t *src, int *dest)
864 {
865 int m, n;
866
867 for (m = 0; m < SUBFRAME_LEN; m++) {
868 int64_t filter = 0;
869 for (n = 1; n <= LPC_ORDER; n++) {
870 filter -= fir_coef[n - 1] * src[m - n] -
871 iir_coef[n - 1] * (dest[m - n] >> 16);
872 }
873
874 dest[m] = av_clipl_int32((src[m] << 16) + (filter << 3) + (1 << 15));
875 }
876 }
877
878 /**
879 * Adjust gain of postfiltered signal.
880 *
881 * @param p the context
882 * @param buf postfiltered output vector
883 * @param energy input energy coefficient
884 */
885 static void gain_scale(G723_1_Context *p, int16_t * buf, int energy)
886 {
887 int num, denom, gain, bits1, bits2;
888 int i;
889
890 num = energy;
891 denom = 0;
892 for (i = 0; i < SUBFRAME_LEN; i++) {
893 int64_t temp = buf[i] >> 2;
894 temp = av_clipl_int32(MUL64(temp, temp) << 1);
895 denom = av_clipl_int32(denom + temp);
896 }
897
898 if (num && denom) {
899 bits1 = normalize_bits(num, 31);
900 bits2 = normalize_bits(denom, 31);
901 num = num << bits1 >> 1;
902 denom <<= bits2;
903
904 bits2 = 5 + bits1 - bits2;
905 bits2 = FFMAX(0, bits2);
906
907 gain = (num >> 1) / (denom >> 16);
908 gain = square_root(gain << 16 >> bits2);
909 } else {
910 gain = 1 << 12;
911 }
912
913 for (i = 0; i < SUBFRAME_LEN; i++) {
914 p->pf_gain = ((p->pf_gain << 4) - p->pf_gain + gain + (1 << 3)) >> 4;
915 buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
916 (1 << 10)) >> 11);
917 }
918 }
919
920 /**
921 * Perform formant filtering.
922 *
923 * @param p the context
924 * @param lpc quantized lpc coefficients
925 * @param buf output buffer
926 */
927 static void formant_postfilter(G723_1_Context *p, int16_t *lpc, int16_t *buf)
928 {
929 int16_t filter_coef[2][LPC_ORDER], *buf_ptr;
930 int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
931 int i, j, k;
932
933 memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(*buf));
934 memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(*filter_signal));
935
936 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
937 for (k = 0; k < LPC_ORDER; k++) {
938 filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
939 (1 << 14)) >> 15;
940 filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
941 (1 << 14)) >> 15;
942 }
943 iir_filter(filter_coef[0], filter_coef[1], buf + i,
944 filter_signal + i);
945 }
946
947 memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(*p->fir_mem));
948 memcpy(p->iir_mem, filter_signal + FRAME_LEN,
949 LPC_ORDER * sizeof(*p->iir_mem));
950
951 buf_ptr = buf + LPC_ORDER;
952 signal_ptr = filter_signal + LPC_ORDER;
953 for (i = 0; i < SUBFRAMES; i++) {
954 int16_t temp_vector[SUBFRAME_LEN];
955 int16_t temp;
956 int auto_corr[2];
957 int scale, energy;
958
959 /* Normalize */
960 memcpy(temp_vector, buf_ptr, SUBFRAME_LEN * sizeof(*temp_vector));
961 scale = scale_vector(temp_vector, SUBFRAME_LEN);
962
963 /* Compute auto correlation coefficients */
964 auto_corr[0] = dot_product(temp_vector, temp_vector + 1,
965 SUBFRAME_LEN - 1, 1);
966 auto_corr[1] = dot_product(temp_vector, temp_vector, SUBFRAME_LEN, 1);
967
968 /* Compute reflection coefficient */
969 temp = auto_corr[1] >> 16;
970 if (temp) {
971 temp = (auto_corr[0] >> 2) / temp;
972 }
973 p->reflection_coef = ((p->reflection_coef << 2) - p->reflection_coef +
974 temp + 2) >> 2;
975 temp = (p->reflection_coef * 0xffffc >> 3) & 0xfffc;
976
977 /* Compensation filter */
978 for (j = 0; j < SUBFRAME_LEN; j++) {
979 buf_ptr[j] = av_clipl_int32(signal_ptr[j] +
980 ((signal_ptr[j - 1] >> 16) *
981 temp << 1)) >> 16;
982 }
983
984 /* Compute normalized signal energy */
985 temp = 2 * scale + 4;
986 if (temp < 0) {
987 energy = av_clipl_int32((int64_t)auto_corr[1] << -temp);
988 } else
989 energy = auto_corr[1] >> temp;
990
991 gain_scale(p, buf_ptr, energy);
992
993 buf_ptr += SUBFRAME_LEN;
994 signal_ptr += SUBFRAME_LEN;
995 }
996 }
997
998 static int g723_1_decode_frame(AVCodecContext *avctx, void *data,
999 int *got_frame_ptr, AVPacket *avpkt)
1000 {
1001 G723_1_Context *p = avctx->priv_data;
1002 const uint8_t *buf = avpkt->data;
1003 int buf_size = avpkt->size;
1004 int dec_mode = buf[0] & 3;
1005
1006 PPFParam ppf[SUBFRAMES];
1007 int16_t cur_lsp[LPC_ORDER];
1008 int16_t lpc[SUBFRAMES * LPC_ORDER];
1009 int16_t acb_vector[SUBFRAME_LEN];
1010 int16_t *vector_ptr;
1011 int bad_frame = 0, i, j, ret;
1012
1013 if (buf_size < frame_size[dec_mode]) {
1014 if (buf_size)
1015 av_log(avctx, AV_LOG_WARNING,
1016 "Expected %d bytes, got %d - skipping packet\n",
1017 frame_size[dec_mode], buf_size);
1018 *got_frame_ptr = 0;
1019 return buf_size;
1020 }
1021
1022 if (unpack_bitstream(p, buf, buf_size) < 0) {
1023 bad_frame = 1;
1024 if (p->past_frame_type == ACTIVE_FRAME)
1025 p->cur_frame_type = ACTIVE_FRAME;
1026 else
1027 p->cur_frame_type = UNTRANSMITTED_FRAME;
1028 }
1029
1030 p->frame.nb_samples = FRAME_LEN;
1031 if ((ret = avctx->get_buffer(avctx, &p->frame)) < 0) {
1032 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1033 return ret;
1034 }
1035
1036 if (p->cur_frame_type == ACTIVE_FRAME) {
1037 if (!bad_frame)
1038 p->erased_frames = 0;
1039 else if (p->erased_frames != 3)
1040 p->erased_frames++;
1041
1042 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
1043 lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
1044
1045 /* Save the lsp_vector for the next frame */
1046 memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1047
1048 /* Generate the excitation for the frame */
1049 memcpy(p->excitation, p->prev_excitation,
1050 PITCH_MAX * sizeof(*p->excitation));
1051 vector_ptr = p->excitation + PITCH_MAX;
1052 if (!p->erased_frames) {
1053 /* Update interpolation gain memory */
1054 p->interp_gain = fixed_cb_gain[(p->subframe[2].amp_index +
1055 p->subframe[3].amp_index) >> 1];
1056 for (i = 0; i < SUBFRAMES; i++) {
1057 gen_fcb_excitation(vector_ptr, p->subframe[i], p->cur_rate,
1058 p->pitch_lag[i >> 1], i);
1059 gen_acb_excitation(acb_vector, &p->excitation[SUBFRAME_LEN * i],
1060 p->pitch_lag[i >> 1], p->subframe[i],
1061 p->cur_rate);
1062 /* Get the total excitation */
1063 for (j = 0; j < SUBFRAME_LEN; j++) {
1064 vector_ptr[j] = av_clip_int16(vector_ptr[j] << 1);
1065 vector_ptr[j] = av_clip_int16(vector_ptr[j] +
1066 acb_vector[j]);
1067 }
1068 vector_ptr += SUBFRAME_LEN;
1069 }
1070
1071 vector_ptr = p->excitation + PITCH_MAX;
1072
1073 /* Save the excitation */
1074 memcpy(p->audio, vector_ptr, FRAME_LEN * sizeof(*p->audio));
1075
1076 p->interp_index = comp_interp_index(p, p->pitch_lag[1],
1077 &p->sid_gain, &p->cur_gain);
1078
1079 if (p->postfilter) {
1080 i = PITCH_MAX;
1081 for (j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1082 comp_ppf_coeff(p, i, p->pitch_lag[j >> 1],
1083 ppf + j, p->cur_rate);
1084 }
1085
1086 /* Restore the original excitation */
1087 memcpy(p->excitation, p->prev_excitation,
1088 PITCH_MAX * sizeof(*p->excitation));
1089 memcpy(vector_ptr, p->audio, FRAME_LEN * sizeof(*vector_ptr));
1090
1091 /* Peform pitch postfiltering */
1092 if (p->postfilter)
1093 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1094 ff_acelp_weighted_vector_sum(p->audio + LPC_ORDER + i,
1095 vector_ptr + i,
1096 vector_ptr + i + ppf[j].index,
1097 ppf[j].sc_gain,
1098 ppf[j].opt_gain,
1099 1 << 14, 15, SUBFRAME_LEN);
1100
1101 } else {
1102 p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
1103 if (p->erased_frames == 3) {
1104 /* Mute output */
1105 memset(p->excitation, 0,
1106 (FRAME_LEN + PITCH_MAX) * sizeof(*p->excitation));
1107 memset(p->frame.data[0], 0,
1108 (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
1109 } else {
1110 /* Regenerate frame */
1111 residual_interp(p->excitation, p->audio + LPC_ORDER, p->interp_index,
1112 p->interp_gain, &p->random_seed);
1113 }
1114 }
1115 /* Save the excitation for the next frame */
1116 memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
1117 PITCH_MAX * sizeof(*p->excitation));
1118 } else {
1119 memset(p->frame.data[0], 0, FRAME_LEN * 2);
1120 av_log(avctx, AV_LOG_WARNING,
1121 "G.723.1: Comfort noise generation not supported yet\n");
1122
1123 *got_frame_ptr = 1;
1124 *(AVFrame *)data = p->frame;
1125 return frame_size[dec_mode];
1126 }
1127
1128 p->past_frame_type = p->cur_frame_type;
1129
1130 memcpy(p->audio, p->synth_mem, LPC_ORDER * sizeof(*p->audio));
1131 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1132 ff_celp_lp_synthesis_filter(p->audio + i, &lpc[j * LPC_ORDER],
1133 p->audio + i, SUBFRAME_LEN, LPC_ORDER,
1134 0, 1, 1 << 12);
1135 memcpy(p->synth_mem, p->audio + FRAME_LEN, LPC_ORDER * sizeof(*p->audio));
1136
1137 if (p->postfilter)
1138 formant_postfilter(p, lpc, p->audio);
1139
1140 memcpy(p->frame.data[0], p->audio + LPC_ORDER, FRAME_LEN * 2);
1141
1142 *got_frame_ptr = 1;
1143 *(AVFrame *)data = p->frame;
1144
1145 return frame_size[dec_mode];
1146 }
1147
1148 #define OFFSET(x) offsetof(G723_1_Context, x)
1149 #define AD AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1150
1151 static const AVOption options[] = {
1152 { "postfilter", "postfilter on/off", OFFSET(postfilter), AV_OPT_TYPE_INT,
1153 { 1 }, 0, 1, AD },
1154 { NULL }
1155 };
1156
1157
1158 static const AVClass g723_1dec_class = {
1159 .class_name = "G.723.1 decoder",
1160 .item_name = av_default_item_name,
1161 .option = options,
1162 .version = LIBAVUTIL_VERSION_INT,
1163 };
1164
1165 AVCodec ff_g723_1_decoder = {
1166 .name = "g723_1",
1167 .type = AVMEDIA_TYPE_AUDIO,
1168 .id = CODEC_ID_G723_1,
1169 .priv_data_size = sizeof(G723_1_Context),
1170 .init = g723_1_decode_init,
1171 .decode = g723_1_decode_frame,
1172 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1173 .capabilities = CODEC_CAP_SUBFRAMES,
1174 .priv_class = &g723_1dec_class,
1175 };