FFmpeg
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
g723_1.c
Go to the documentation of this file.
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 #define BITSTREAM_READER_LE
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 "celp_math.h"
37 #include "g723_1_data.h"
38 #include "internal.h"
39 
40 #define CNG_RANDOM_SEED 12345
41 
42 typedef struct g723_1_context {
43  AVClass *class;
44 
45  G723_1_Subframe subframe[4];
48  enum Rate cur_rate;
49  uint8_t lsp_index[LSP_BANDS];
50  int pitch_lag[2];
52 
53  int16_t prev_lsp[LPC_ORDER];
54  int16_t sid_lsp[LPC_ORDER];
55  int16_t prev_excitation[PITCH_MAX];
56  int16_t excitation[PITCH_MAX + FRAME_LEN + 4];
57  int16_t synth_mem[LPC_ORDER];
58  int16_t fir_mem[LPC_ORDER];
59  int iir_mem[LPC_ORDER];
60 
65  int sid_gain;
66  int cur_gain;
68  int pf_gain; ///< formant postfilter
69  ///< gain scaling unit memory
71 
72  int16_t audio[FRAME_LEN + LPC_ORDER + PITCH_MAX + 4];
73  int16_t prev_data[HALF_FRAME_LEN];
74  int16_t prev_weight_sig[PITCH_MAX];
75 
76 
77  int16_t hpf_fir_mem; ///< highpass filter fir
78  int hpf_iir_mem; ///< and iir memories
79  int16_t perf_fir_mem[LPC_ORDER]; ///< perceptual filter fir
80  int16_t perf_iir_mem[LPC_ORDER]; ///< and iir memories
81 
82  int16_t harmonic_mem[PITCH_MAX];
84 
86 {
87  G723_1_Context *p = avctx->priv_data;
88 
91  avctx->channels = 1;
92  p->pf_gain = 1 << 12;
93 
94  memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
95  memcpy(p->sid_lsp, dc_lsp, LPC_ORDER * sizeof(*p->sid_lsp));
96 
99 
100  return 0;
101 }
102 
103 /**
104  * Unpack the frame into parameters.
105  *
106  * @param p the context
107  * @param buf pointer to the input buffer
108  * @param buf_size size of the input buffer
109  */
111  int buf_size)
112 {
113  GetBitContext gb;
114  int ad_cb_len;
115  int temp, info_bits, i;
116 
117  init_get_bits(&gb, buf, buf_size * 8);
118 
119  /* Extract frame type and rate info */
120  info_bits = get_bits(&gb, 2);
121 
122  if (info_bits == 3) {
124  return 0;
125  }
126 
127  /* Extract 24 bit lsp indices, 8 bit for each band */
128  p->lsp_index[2] = get_bits(&gb, 8);
129  p->lsp_index[1] = get_bits(&gb, 8);
130  p->lsp_index[0] = get_bits(&gb, 8);
131 
132  if (info_bits == 2) {
134  p->subframe[0].amp_index = get_bits(&gb, 6);
135  return 0;
136  }
137 
138  /* Extract the info common to both rates */
139  p->cur_rate = info_bits ? RATE_5300 : RATE_6300;
141 
142  p->pitch_lag[0] = get_bits(&gb, 7);
143  if (p->pitch_lag[0] > 123) /* test if forbidden code */
144  return -1;
145  p->pitch_lag[0] += PITCH_MIN;
146  p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
147 
148  p->pitch_lag[1] = get_bits(&gb, 7);
149  if (p->pitch_lag[1] > 123)
150  return -1;
151  p->pitch_lag[1] += PITCH_MIN;
152  p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
153  p->subframe[0].ad_cb_lag = 1;
154  p->subframe[2].ad_cb_lag = 1;
155 
156  for (i = 0; i < SUBFRAMES; i++) {
157  /* Extract combined gain */
158  temp = get_bits(&gb, 12);
159  ad_cb_len = 170;
160  p->subframe[i].dirac_train = 0;
161  if (p->cur_rate == RATE_6300 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
162  p->subframe[i].dirac_train = temp >> 11;
163  temp &= 0x7FF;
164  ad_cb_len = 85;
165  }
166  p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
167  if (p->subframe[i].ad_cb_gain < ad_cb_len) {
168  p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
169  GAIN_LEVELS;
170  } else {
171  return -1;
172  }
173  }
174 
175  p->subframe[0].grid_index = get_bits1(&gb);
176  p->subframe[1].grid_index = get_bits1(&gb);
177  p->subframe[2].grid_index = get_bits1(&gb);
178  p->subframe[3].grid_index = get_bits1(&gb);
179 
180  if (p->cur_rate == RATE_6300) {
181  skip_bits1(&gb); /* skip reserved bit */
182 
183  /* Compute pulse_pos index using the 13-bit combined position index */
184  temp = get_bits(&gb, 13);
185  p->subframe[0].pulse_pos = temp / 810;
186 
187  temp -= p->subframe[0].pulse_pos * 810;
188  p->subframe[1].pulse_pos = FASTDIV(temp, 90);
189 
190  temp -= p->subframe[1].pulse_pos * 90;
191  p->subframe[2].pulse_pos = FASTDIV(temp, 9);
192  p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
193 
194  p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
195  get_bits(&gb, 16);
196  p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
197  get_bits(&gb, 14);
198  p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
199  get_bits(&gb, 16);
200  p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
201  get_bits(&gb, 14);
202 
203  p->subframe[0].pulse_sign = get_bits(&gb, 6);
204  p->subframe[1].pulse_sign = get_bits(&gb, 5);
205  p->subframe[2].pulse_sign = get_bits(&gb, 6);
206  p->subframe[3].pulse_sign = get_bits(&gb, 5);
207  } else { /* 5300 bps */
208  p->subframe[0].pulse_pos = get_bits(&gb, 12);
209  p->subframe[1].pulse_pos = get_bits(&gb, 12);
210  p->subframe[2].pulse_pos = get_bits(&gb, 12);
211  p->subframe[3].pulse_pos = get_bits(&gb, 12);
212 
213  p->subframe[0].pulse_sign = get_bits(&gb, 4);
214  p->subframe[1].pulse_sign = get_bits(&gb, 4);
215  p->subframe[2].pulse_sign = get_bits(&gb, 4);
216  p->subframe[3].pulse_sign = get_bits(&gb, 4);
217  }
218 
219  return 0;
220 }
221 
222 /**
223  * Bitexact implementation of sqrt(val/2).
224  */
225 static int16_t square_root(unsigned val)
226 {
227  av_assert2(!(val & 0x80000000));
228 
229  return (ff_sqrt(val << 1) >> 1) & (~1);
230 }
231 
232 /**
233  * Calculate the number of left-shifts required for normalizing the input.
234  *
235  * @param num input number
236  * @param width width of the input, 15 or 31 bits
237  */
238 static int normalize_bits(int num, int width)
239 {
240  return width - av_log2(num) - 1;
241 }
242 
243 #define normalize_bits_int16(num) normalize_bits(num, 15)
244 #define normalize_bits_int32(num) normalize_bits(num, 31)
245 
246 /**
247  * Scale vector contents based on the largest of their absolutes.
248  */
249 static int scale_vector(int16_t *dst, const int16_t *vector, int length)
250 {
251  int bits, max = 0;
252  int i;
253 
254  for (i = 0; i < length; i++)
255  max |= FFABS(vector[i]);
256 
257  bits= 14 - av_log2_16bit(max);
258  bits= FFMAX(bits, 0);
259 
260  for (i = 0; i < length; i++)
261  dst[i] = vector[i] << bits >> 3;
262 
263  return bits - 3;
264 }
265 
266 /**
267  * Perform inverse quantization of LSP frequencies.
268  *
269  * @param cur_lsp the current LSP vector
270  * @param prev_lsp the previous LSP vector
271  * @param lsp_index VQ indices
272  * @param bad_frame bad frame flag
273  */
274 static void inverse_quant(int16_t *cur_lsp, int16_t *prev_lsp,
275  uint8_t *lsp_index, int bad_frame)
276 {
277  int min_dist, pred;
278  int i, j, temp, stable;
279 
280  /* Check for frame erasure */
281  if (!bad_frame) {
282  min_dist = 0x100;
283  pred = 12288;
284  } else {
285  min_dist = 0x200;
286  pred = 23552;
287  lsp_index[0] = lsp_index[1] = lsp_index[2] = 0;
288  }
289 
290  /* Get the VQ table entry corresponding to the transmitted index */
291  cur_lsp[0] = lsp_band0[lsp_index[0]][0];
292  cur_lsp[1] = lsp_band0[lsp_index[0]][1];
293  cur_lsp[2] = lsp_band0[lsp_index[0]][2];
294  cur_lsp[3] = lsp_band1[lsp_index[1]][0];
295  cur_lsp[4] = lsp_band1[lsp_index[1]][1];
296  cur_lsp[5] = lsp_band1[lsp_index[1]][2];
297  cur_lsp[6] = lsp_band2[lsp_index[2]][0];
298  cur_lsp[7] = lsp_band2[lsp_index[2]][1];
299  cur_lsp[8] = lsp_band2[lsp_index[2]][2];
300  cur_lsp[9] = lsp_band2[lsp_index[2]][3];
301 
302  /* Add predicted vector & DC component to the previously quantized vector */
303  for (i = 0; i < LPC_ORDER; i++) {
304  temp = ((prev_lsp[i] - dc_lsp[i]) * pred + (1 << 14)) >> 15;
305  cur_lsp[i] += dc_lsp[i] + temp;
306  }
307 
308  for (i = 0; i < LPC_ORDER; i++) {
309  cur_lsp[0] = FFMAX(cur_lsp[0], 0x180);
310  cur_lsp[LPC_ORDER - 1] = FFMIN(cur_lsp[LPC_ORDER - 1], 0x7e00);
311 
312  /* Stability check */
313  for (j = 1; j < LPC_ORDER; j++) {
314  temp = min_dist + cur_lsp[j - 1] - cur_lsp[j];
315  if (temp > 0) {
316  temp >>= 1;
317  cur_lsp[j - 1] -= temp;
318  cur_lsp[j] += temp;
319  }
320  }
321  stable = 1;
322  for (j = 1; j < LPC_ORDER; j++) {
323  temp = cur_lsp[j - 1] + min_dist - cur_lsp[j] - 4;
324  if (temp > 0) {
325  stable = 0;
326  break;
327  }
328  }
329  if (stable)
330  break;
331  }
332  if (!stable)
333  memcpy(cur_lsp, prev_lsp, LPC_ORDER * sizeof(*cur_lsp));
334 }
335 
336 /**
337  * Bitexact implementation of 2ab scaled by 1/2^16.
338  *
339  * @param a 32 bit multiplicand
340  * @param b 16 bit multiplier
341  */
342 #define MULL2(a, b) \
343  MULL(a,b,15)
344 
345 /**
346  * Convert LSP frequencies to LPC coefficients.
347  *
348  * @param lpc buffer for LPC coefficients
349  */
350 static void lsp2lpc(int16_t *lpc)
351 {
352  int f1[LPC_ORDER / 2 + 1];
353  int f2[LPC_ORDER / 2 + 1];
354  int i, j;
355 
356  /* Calculate negative cosine */
357  for (j = 0; j < LPC_ORDER; j++) {
358  int index = (lpc[j] >> 7) & 0x1FF;
359  int offset = lpc[j] & 0x7f;
360  int temp1 = cos_tab[index] << 16;
361  int temp2 = (cos_tab[index + 1] - cos_tab[index]) *
362  ((offset << 8) + 0x80) << 1;
363 
364  lpc[j] = -(av_sat_dadd32(1 << 15, temp1 + temp2) >> 16);
365  }
366 
367  /*
368  * Compute sum and difference polynomial coefficients
369  * (bitexact alternative to lsp2poly() in lsp.c)
370  */
371  /* Initialize with values in Q28 */
372  f1[0] = 1 << 28;
373  f1[1] = (lpc[0] << 14) + (lpc[2] << 14);
374  f1[2] = lpc[0] * lpc[2] + (2 << 28);
375 
376  f2[0] = 1 << 28;
377  f2[1] = (lpc[1] << 14) + (lpc[3] << 14);
378  f2[2] = lpc[1] * lpc[3] + (2 << 28);
379 
380  /*
381  * Calculate and scale the coefficients by 1/2 in
382  * each iteration for a final scaling factor of Q25
383  */
384  for (i = 2; i < LPC_ORDER / 2; i++) {
385  f1[i + 1] = f1[i - 1] + MULL2(f1[i], lpc[2 * i]);
386  f2[i + 1] = f2[i - 1] + MULL2(f2[i], lpc[2 * i + 1]);
387 
388  for (j = i; j >= 2; j--) {
389  f1[j] = MULL2(f1[j - 1], lpc[2 * i]) +
390  (f1[j] >> 1) + (f1[j - 2] >> 1);
391  f2[j] = MULL2(f2[j - 1], lpc[2 * i + 1]) +
392  (f2[j] >> 1) + (f2[j - 2] >> 1);
393  }
394 
395  f1[0] >>= 1;
396  f2[0] >>= 1;
397  f1[1] = ((lpc[2 * i] << 16 >> i) + f1[1]) >> 1;
398  f2[1] = ((lpc[2 * i + 1] << 16 >> i) + f2[1]) >> 1;
399  }
400 
401  /* Convert polynomial coefficients to LPC coefficients */
402  for (i = 0; i < LPC_ORDER / 2; i++) {
403  int64_t ff1 = f1[i + 1] + f1[i];
404  int64_t ff2 = f2[i + 1] - f2[i];
405 
406  lpc[i] = av_clipl_int32(((ff1 + ff2) << 3) + (1 << 15)) >> 16;
407  lpc[LPC_ORDER - i - 1] = av_clipl_int32(((ff1 - ff2) << 3) +
408  (1 << 15)) >> 16;
409  }
410 }
411 
412 /**
413  * Quantize LSP frequencies by interpolation and convert them to
414  * the corresponding LPC coefficients.
415  *
416  * @param lpc buffer for LPC coefficients
417  * @param cur_lsp the current LSP vector
418  * @param prev_lsp the previous LSP vector
419  */
420 static void lsp_interpolate(int16_t *lpc, int16_t *cur_lsp, int16_t *prev_lsp)
421 {
422  int i;
423  int16_t *lpc_ptr = lpc;
424 
425  /* cur_lsp * 0.25 + prev_lsp * 0.75 */
426  ff_acelp_weighted_vector_sum(lpc, cur_lsp, prev_lsp,
427  4096, 12288, 1 << 13, 14, LPC_ORDER);
428  ff_acelp_weighted_vector_sum(lpc + LPC_ORDER, cur_lsp, prev_lsp,
429  8192, 8192, 1 << 13, 14, LPC_ORDER);
430  ff_acelp_weighted_vector_sum(lpc + 2 * LPC_ORDER, cur_lsp, prev_lsp,
431  12288, 4096, 1 << 13, 14, LPC_ORDER);
432  memcpy(lpc + 3 * LPC_ORDER, cur_lsp, LPC_ORDER * sizeof(*lpc));
433 
434  for (i = 0; i < SUBFRAMES; i++) {
435  lsp2lpc(lpc_ptr);
436  lpc_ptr += LPC_ORDER;
437  }
438 }
439 
440 /**
441  * Generate a train of dirac functions with period as pitch lag.
442  */
443 static void gen_dirac_train(int16_t *buf, int pitch_lag)
444 {
445  int16_t vector[SUBFRAME_LEN];
446  int i, j;
447 
448  memcpy(vector, buf, SUBFRAME_LEN * sizeof(*vector));
449  for (i = pitch_lag; i < SUBFRAME_LEN; i += pitch_lag) {
450  for (j = 0; j < SUBFRAME_LEN - i; j++)
451  buf[i + j] += vector[j];
452  }
453 }
454 
455 /**
456  * Generate fixed codebook excitation vector.
457  *
458  * @param vector decoded excitation vector
459  * @param subfrm current subframe
460  * @param cur_rate current bitrate
461  * @param pitch_lag closed loop pitch lag
462  * @param index current subframe index
463  */
464 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe *subfrm,
465  enum Rate cur_rate, int pitch_lag, int index)
466 {
467  int temp, i, j;
468 
469  memset(vector, 0, SUBFRAME_LEN * sizeof(*vector));
470 
471  if (cur_rate == RATE_6300) {
472  if (subfrm->pulse_pos >= max_pos[index])
473  return;
474 
475  /* Decode amplitudes and positions */
476  j = PULSE_MAX - pulses[index];
477  temp = subfrm->pulse_pos;
478  for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
479  temp -= combinatorial_table[j][i];
480  if (temp >= 0)
481  continue;
482  temp += combinatorial_table[j++][i];
483  if (subfrm->pulse_sign & (1 << (PULSE_MAX - j))) {
484  vector[subfrm->grid_index + GRID_SIZE * i] =
485  -fixed_cb_gain[subfrm->amp_index];
486  } else {
487  vector[subfrm->grid_index + GRID_SIZE * i] =
488  fixed_cb_gain[subfrm->amp_index];
489  }
490  if (j == PULSE_MAX)
491  break;
492  }
493  if (subfrm->dirac_train == 1)
494  gen_dirac_train(vector, pitch_lag);
495  } else { /* 5300 bps */
496  int cb_gain = fixed_cb_gain[subfrm->amp_index];
497  int cb_shift = subfrm->grid_index;
498  int cb_sign = subfrm->pulse_sign;
499  int cb_pos = subfrm->pulse_pos;
500  int offset, beta, lag;
501 
502  for (i = 0; i < 8; i += 2) {
503  offset = ((cb_pos & 7) << 3) + cb_shift + i;
504  vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
505  cb_pos >>= 3;
506  cb_sign >>= 1;
507  }
508 
509  /* Enhance harmonic components */
510  lag = pitch_contrib[subfrm->ad_cb_gain << 1] + pitch_lag +
511  subfrm->ad_cb_lag - 1;
512  beta = pitch_contrib[(subfrm->ad_cb_gain << 1) + 1];
513 
514  if (lag < SUBFRAME_LEN - 2) {
515  for (i = lag; i < SUBFRAME_LEN; i++)
516  vector[i] += beta * vector[i - lag] >> 15;
517  }
518  }
519 }
520 
521 /**
522  * Get delayed contribution from the previous excitation vector.
523  */
524 static void get_residual(int16_t *residual, int16_t *prev_excitation, int lag)
525 {
526  int offset = PITCH_MAX - PITCH_ORDER / 2 - lag;
527  int i;
528 
529  residual[0] = prev_excitation[offset];
530  residual[1] = prev_excitation[offset + 1];
531 
532  offset += 2;
533  for (i = 2; i < SUBFRAME_LEN + PITCH_ORDER - 1; i++)
534  residual[i] = prev_excitation[offset + (i - 2) % lag];
535 }
536 
537 static int dot_product(const int16_t *a, const int16_t *b, int length)
538 {
539  int sum = ff_dot_product(a,b,length);
540  return av_sat_add32(sum, sum);
541 }
542 
543 /**
544  * Generate adaptive codebook excitation.
545  */
546 static void gen_acb_excitation(int16_t *vector, int16_t *prev_excitation,
547  int pitch_lag, G723_1_Subframe *subfrm,
548  enum Rate cur_rate)
549 {
550  int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
551  const int16_t *cb_ptr;
552  int lag = pitch_lag + subfrm->ad_cb_lag - 1;
553 
554  int i;
555  int sum;
556 
557  get_residual(residual, prev_excitation, lag);
558 
559  /* Select quantization table */
560  if (cur_rate == RATE_6300 && pitch_lag < SUBFRAME_LEN - 2) {
561  cb_ptr = adaptive_cb_gain85;
562  } else
563  cb_ptr = adaptive_cb_gain170;
564 
565  /* Calculate adaptive vector */
566  cb_ptr += subfrm->ad_cb_gain * 20;
567  for (i = 0; i < SUBFRAME_LEN; i++) {
568  sum = ff_dot_product(residual + i, cb_ptr, PITCH_ORDER);
569  vector[i] = av_sat_dadd32(1 << 15, av_sat_add32(sum, sum)) >> 16;
570  }
571 }
572 
573 /**
574  * Estimate maximum auto-correlation around pitch lag.
575  *
576  * @param buf buffer with offset applied
577  * @param offset offset of the excitation vector
578  * @param ccr_max pointer to the maximum auto-correlation
579  * @param pitch_lag decoded pitch lag
580  * @param length length of autocorrelation
581  * @param dir forward lag(1) / backward lag(-1)
582  */
583 static int autocorr_max(const int16_t *buf, int offset, int *ccr_max,
584  int pitch_lag, int length, int dir)
585 {
586  int limit, ccr, lag = 0;
587  int i;
588 
589  pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
590  if (dir > 0)
591  limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
592  else
593  limit = pitch_lag + 3;
594 
595  for (i = pitch_lag - 3; i <= limit; i++) {
596  ccr = dot_product(buf, buf + dir * i, length);
597 
598  if (ccr > *ccr_max) {
599  *ccr_max = ccr;
600  lag = i;
601  }
602  }
603  return lag;
604 }
605 
606 /**
607  * Calculate pitch postfilter optimal and scaling gains.
608  *
609  * @param lag pitch postfilter forward/backward lag
610  * @param ppf pitch postfilter parameters
611  * @param cur_rate current bitrate
612  * @param tgt_eng target energy
613  * @param ccr cross-correlation
614  * @param res_eng residual energy
615  */
616 static void comp_ppf_gains(int lag, PPFParam *ppf, enum Rate cur_rate,
617  int tgt_eng, int ccr, int res_eng)
618 {
619  int pf_residual; /* square of postfiltered residual */
620  int temp1, temp2;
621 
622  ppf->index = lag;
623 
624  temp1 = tgt_eng * res_eng >> 1;
625  temp2 = ccr * ccr << 1;
626 
627  if (temp2 > temp1) {
628  if (ccr >= res_eng) {
629  ppf->opt_gain = ppf_gain_weight[cur_rate];
630  } else {
631  ppf->opt_gain = (ccr << 15) / res_eng *
632  ppf_gain_weight[cur_rate] >> 15;
633  }
634  /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
635  temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
636  temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
637  pf_residual = av_sat_add32(temp1, temp2 + (1 << 15)) >> 16;
638 
639  if (tgt_eng >= pf_residual << 1) {
640  temp1 = 0x7fff;
641  } else {
642  temp1 = (tgt_eng << 14) / pf_residual;
643  }
644 
645  /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
646  ppf->sc_gain = square_root(temp1 << 16);
647  } else {
648  ppf->opt_gain = 0;
649  ppf->sc_gain = 0x7fff;
650  }
651 
652  ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
653 }
654 
655 /**
656  * Calculate pitch postfilter parameters.
657  *
658  * @param p the context
659  * @param offset offset of the excitation vector
660  * @param pitch_lag decoded pitch lag
661  * @param ppf pitch postfilter parameters
662  * @param cur_rate current bitrate
663  */
664 static void comp_ppf_coeff(G723_1_Context *p, int offset, int pitch_lag,
665  PPFParam *ppf, enum Rate cur_rate)
666 {
667 
668  int16_t scale;
669  int i;
670  int temp1, temp2;
671 
672  /*
673  * 0 - target energy
674  * 1 - forward cross-correlation
675  * 2 - forward residual energy
676  * 3 - backward cross-correlation
677  * 4 - backward residual energy
678  */
679  int energy[5] = {0, 0, 0, 0, 0};
680  int16_t *buf = p->audio + LPC_ORDER + offset;
681  int fwd_lag = autocorr_max(buf, offset, &energy[1], pitch_lag,
682  SUBFRAME_LEN, 1);
683  int back_lag = autocorr_max(buf, offset, &energy[3], pitch_lag,
684  SUBFRAME_LEN, -1);
685 
686  ppf->index = 0;
687  ppf->opt_gain = 0;
688  ppf->sc_gain = 0x7fff;
689 
690  /* Case 0, Section 3.6 */
691  if (!back_lag && !fwd_lag)
692  return;
693 
694  /* Compute target energy */
695  energy[0] = dot_product(buf, buf, SUBFRAME_LEN);
696 
697  /* Compute forward residual energy */
698  if (fwd_lag)
699  energy[2] = dot_product(buf + fwd_lag, buf + fwd_lag, SUBFRAME_LEN);
700 
701  /* Compute backward residual energy */
702  if (back_lag)
703  energy[4] = dot_product(buf - back_lag, buf - back_lag, SUBFRAME_LEN);
704 
705  /* Normalize and shorten */
706  temp1 = 0;
707  for (i = 0; i < 5; i++)
708  temp1 = FFMAX(energy[i], temp1);
709 
710  scale = normalize_bits(temp1, 31);
711  for (i = 0; i < 5; i++)
712  energy[i] = (energy[i] << scale) >> 16;
713 
714  if (fwd_lag && !back_lag) { /* Case 1 */
715  comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
716  energy[2]);
717  } else if (!fwd_lag) { /* Case 2 */
718  comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
719  energy[4]);
720  } else { /* Case 3 */
721 
722  /*
723  * Select the largest of energy[1]^2/energy[2]
724  * and energy[3]^2/energy[4]
725  */
726  temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
727  temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
728  if (temp1 >= temp2) {
729  comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
730  energy[2]);
731  } else {
732  comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
733  energy[4]);
734  }
735  }
736 }
737 
738 /**
739  * Classify frames as voiced/unvoiced.
740  *
741  * @param p the context
742  * @param pitch_lag decoded pitch_lag
743  * @param exc_eng excitation energy estimation
744  * @param scale scaling factor of exc_eng
745  *
746  * @return residual interpolation index if voiced, 0 otherwise
747  */
748 static int comp_interp_index(G723_1_Context *p, int pitch_lag,
749  int *exc_eng, int *scale)
750 {
751  int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
752  int16_t *buf = p->audio + LPC_ORDER;
753 
754  int index, ccr, tgt_eng, best_eng, temp;
755 
756  *scale = scale_vector(buf, p->excitation, FRAME_LEN + PITCH_MAX);
757  buf += offset;
758 
759  /* Compute maximum backward cross-correlation */
760  ccr = 0;
761  index = autocorr_max(buf, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
762  ccr = av_sat_add32(ccr, 1 << 15) >> 16;
763 
764  /* Compute target energy */
765  tgt_eng = dot_product(buf, buf, SUBFRAME_LEN * 2);
766  *exc_eng = av_sat_add32(tgt_eng, 1 << 15) >> 16;
767 
768  if (ccr <= 0)
769  return 0;
770 
771  /* Compute best energy */
772  best_eng = dot_product(buf - index, buf - index, SUBFRAME_LEN * 2);
773  best_eng = av_sat_add32(best_eng, 1 << 15) >> 16;
774 
775  temp = best_eng * *exc_eng >> 3;
776 
777  if (temp < ccr * ccr) {
778  return index;
779  } else
780  return 0;
781 }
782 
783 /**
784  * Peform residual interpolation based on frame classification.
785  *
786  * @param buf decoded excitation vector
787  * @param out output vector
788  * @param lag decoded pitch lag
789  * @param gain interpolated gain
790  * @param rseed seed for random number generator
791  */
792 static void residual_interp(int16_t *buf, int16_t *out, int lag,
793  int gain, int *rseed)
794 {
795  int i;
796  if (lag) { /* Voiced */
797  int16_t *vector_ptr = buf + PITCH_MAX;
798  /* Attenuate */
799  for (i = 0; i < lag; i++)
800  out[i] = vector_ptr[i - lag] * 3 >> 2;
801  av_memcpy_backptr((uint8_t*)(out + lag), lag * sizeof(*out),
802  (FRAME_LEN - lag) * sizeof(*out));
803  } else { /* Unvoiced */
804  for (i = 0; i < FRAME_LEN; i++) {
805  *rseed = *rseed * 521 + 259;
806  out[i] = gain * *rseed >> 15;
807  }
808  memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(*buf));
809  }
810 }
811 
812 /**
813  * Perform IIR filtering.
814  *
815  * @param fir_coef FIR coefficients
816  * @param iir_coef IIR coefficients
817  * @param src source vector
818  * @param dest destination vector
819  * @param width width of the output, 16 bits(0) / 32 bits(1)
820  */
821 #define iir_filter(fir_coef, iir_coef, src, dest, width)\
822 {\
823  int m, n;\
824  int res_shift = 16 & ~-(width);\
825  int in_shift = 16 - res_shift;\
826 \
827  for (m = 0; m < SUBFRAME_LEN; m++) {\
828  int64_t filter = 0;\
829  for (n = 1; n <= LPC_ORDER; n++) {\
830  filter -= (fir_coef)[n - 1] * (src)[m - n] -\
831  (iir_coef)[n - 1] * ((dest)[m - n] >> in_shift);\
832  }\
833 \
834  (dest)[m] = av_clipl_int32(((src)[m] << 16) + (filter << 3) +\
835  (1 << 15)) >> res_shift;\
836  }\
837 }
838 
839 /**
840  * Adjust gain of postfiltered signal.
841  *
842  * @param p the context
843  * @param buf postfiltered output vector
844  * @param energy input energy coefficient
845  */
846 static void gain_scale(G723_1_Context *p, int16_t * buf, int energy)
847 {
848  int num, denom, gain, bits1, bits2;
849  int i;
850 
851  num = energy;
852  denom = 0;
853  for (i = 0; i < SUBFRAME_LEN; i++) {
854  int temp = buf[i] >> 2;
855  temp *= temp;
856  denom = av_sat_dadd32(denom, temp);
857  }
858 
859  if (num && denom) {
860  bits1 = normalize_bits(num, 31);
861  bits2 = normalize_bits(denom, 31);
862  num = num << bits1 >> 1;
863  denom <<= bits2;
864 
865  bits2 = 5 + bits1 - bits2;
866  bits2 = FFMAX(0, bits2);
867 
868  gain = (num >> 1) / (denom >> 16);
869  gain = square_root(gain << 16 >> bits2);
870  } else {
871  gain = 1 << 12;
872  }
873 
874  for (i = 0; i < SUBFRAME_LEN; i++) {
875  p->pf_gain = (15 * p->pf_gain + gain + (1 << 3)) >> 4;
876  buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
877  (1 << 10)) >> 11);
878  }
879 }
880 
881 /**
882  * Perform formant filtering.
883  *
884  * @param p the context
885  * @param lpc quantized lpc coefficients
886  * @param buf input buffer
887  * @param dst output buffer
888  */
889 static void formant_postfilter(G723_1_Context *p, int16_t *lpc,
890  int16_t *buf, int16_t *dst)
891 {
892  int16_t filter_coef[2][LPC_ORDER];
893  int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
894  int i, j, k;
895 
896  memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(*buf));
897  memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(*filter_signal));
898 
899  for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
900  for (k = 0; k < LPC_ORDER; k++) {
901  filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
902  (1 << 14)) >> 15;
903  filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
904  (1 << 14)) >> 15;
905  }
906  iir_filter(filter_coef[0], filter_coef[1], buf + i,
907  filter_signal + i, 1);
908  lpc += LPC_ORDER;
909  }
910 
911  memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(int16_t));
912  memcpy(p->iir_mem, filter_signal + FRAME_LEN, LPC_ORDER * sizeof(int));
913 
914  buf += LPC_ORDER;
915  signal_ptr = filter_signal + LPC_ORDER;
916  for (i = 0; i < SUBFRAMES; i++) {
917  int temp;
918  int auto_corr[2];
919  int scale, energy;
920 
921  /* Normalize */
922  scale = scale_vector(dst, buf, SUBFRAME_LEN);
923 
924  /* Compute auto correlation coefficients */
925  auto_corr[0] = dot_product(dst, dst + 1, SUBFRAME_LEN - 1);
926  auto_corr[1] = dot_product(dst, dst, SUBFRAME_LEN);
927 
928  /* Compute reflection coefficient */
929  temp = auto_corr[1] >> 16;
930  if (temp) {
931  temp = (auto_corr[0] >> 2) / temp;
932  }
933  p->reflection_coef = (3 * p->reflection_coef + temp + 2) >> 2;
934  temp = -p->reflection_coef >> 1 & ~3;
935 
936  /* Compensation filter */
937  for (j = 0; j < SUBFRAME_LEN; j++) {
938  dst[j] = av_sat_dadd32(signal_ptr[j],
939  (signal_ptr[j - 1] >> 16) * temp) >> 16;
940  }
941 
942  /* Compute normalized signal energy */
943  temp = 2 * scale + 4;
944  if (temp < 0) {
945  energy = av_clipl_int32((int64_t)auto_corr[1] << -temp);
946  } else
947  energy = auto_corr[1] >> temp;
948 
949  gain_scale(p, dst, energy);
950 
951  buf += SUBFRAME_LEN;
952  signal_ptr += SUBFRAME_LEN;
953  dst += SUBFRAME_LEN;
954  }
955 }
956 
957 static int sid_gain_to_lsp_index(int gain)
958 {
959  if (gain < 0x10)
960  return gain << 6;
961  else if (gain < 0x20)
962  return gain - 8 << 7;
963  else
964  return gain - 20 << 8;
965 }
966 
967 static inline int cng_rand(int *state, int base)
968 {
969  *state = (*state * 521 + 259) & 0xFFFF;
970  return (*state & 0x7FFF) * base >> 15;
971 }
972 
974 {
975  int i, shift, seg, seg2, t, val, val_add, x, y;
976 
977  shift = 16 - p->cur_gain * 2;
978  if (shift > 0)
979  t = p->sid_gain << shift;
980  else
981  t = p->sid_gain >> -shift;
982  x = t * cng_filt[0] >> 16;
983 
984  if (x >= cng_bseg[2])
985  return 0x3F;
986 
987  if (x >= cng_bseg[1]) {
988  shift = 4;
989  seg = 3;
990  } else {
991  shift = 3;
992  seg = (x >= cng_bseg[0]);
993  }
994  seg2 = FFMIN(seg, 3);
995 
996  val = 1 << shift;
997  val_add = val >> 1;
998  for (i = 0; i < shift; i++) {
999  t = seg * 32 + (val << seg2);
1000  t *= t;
1001  if (x >= t)
1002  val += val_add;
1003  else
1004  val -= val_add;
1005  val_add >>= 1;
1006  }
1007 
1008  t = seg * 32 + (val << seg2);
1009  y = t * t - x;
1010  if (y <= 0) {
1011  t = seg * 32 + (val + 1 << seg2);
1012  t = t * t - x;
1013  val = (seg2 - 1 << 4) + val;
1014  if (t >= y)
1015  val++;
1016  } else {
1017  t = seg * 32 + (val - 1 << seg2);
1018  t = t * t - x;
1019  val = (seg2 - 1 << 4) + val;
1020  if (t >= y)
1021  val--;
1022  }
1023 
1024  return val;
1025 }
1026 
1028 {
1029  int i, j, idx, t;
1030  int off[SUBFRAMES];
1031  int signs[SUBFRAMES / 2 * 11], pos[SUBFRAMES / 2 * 11];
1032  int tmp[SUBFRAME_LEN * 2];
1033  int16_t *vector_ptr;
1034  int64_t sum;
1035  int b0, c, delta, x, shift;
1036 
1037  p->pitch_lag[0] = cng_rand(&p->cng_random_seed, 21) + 123;
1038  p->pitch_lag[1] = cng_rand(&p->cng_random_seed, 19) + 123;
1039 
1040  for (i = 0; i < SUBFRAMES; i++) {
1041  p->subframe[i].ad_cb_gain = cng_rand(&p->cng_random_seed, 50) + 1;
1043  }
1044 
1045  for (i = 0; i < SUBFRAMES / 2; i++) {
1046  t = cng_rand(&p->cng_random_seed, 1 << 13);
1047  off[i * 2] = t & 1;
1048  off[i * 2 + 1] = ((t >> 1) & 1) + SUBFRAME_LEN;
1049  t >>= 2;
1050  for (j = 0; j < 11; j++) {
1051  signs[i * 11 + j] = (t & 1) * 2 - 1 << 14;
1052  t >>= 1;
1053  }
1054  }
1055 
1056  idx = 0;
1057  for (i = 0; i < SUBFRAMES; i++) {
1058  for (j = 0; j < SUBFRAME_LEN / 2; j++)
1059  tmp[j] = j;
1060  t = SUBFRAME_LEN / 2;
1061  for (j = 0; j < pulses[i]; j++, idx++) {
1062  int idx2 = cng_rand(&p->cng_random_seed, t);
1063 
1064  pos[idx] = tmp[idx2] * 2 + off[i];
1065  tmp[idx2] = tmp[--t];
1066  }
1067  }
1068 
1069  vector_ptr = p->audio + LPC_ORDER;
1070  memcpy(vector_ptr, p->prev_excitation,
1071  PITCH_MAX * sizeof(*p->excitation));
1072  for (i = 0; i < SUBFRAMES; i += 2) {
1073  gen_acb_excitation(vector_ptr, vector_ptr,
1074  p->pitch_lag[i >> 1], &p->subframe[i],
1075  p->cur_rate);
1076  gen_acb_excitation(vector_ptr + SUBFRAME_LEN,
1077  vector_ptr + SUBFRAME_LEN,
1078  p->pitch_lag[i >> 1], &p->subframe[i + 1],
1079  p->cur_rate);
1080 
1081  t = 0;
1082  for (j = 0; j < SUBFRAME_LEN * 2; j++)
1083  t |= FFABS(vector_ptr[j]);
1084  t = FFMIN(t, 0x7FFF);
1085  if (!t) {
1086  shift = 0;
1087  } else {
1088  shift = -10 + av_log2(t);
1089  if (shift < -2)
1090  shift = -2;
1091  }
1092  sum = 0;
1093  if (shift < 0) {
1094  for (j = 0; j < SUBFRAME_LEN * 2; j++) {
1095  t = vector_ptr[j] << -shift;
1096  sum += t * t;
1097  tmp[j] = t;
1098  }
1099  } else {
1100  for (j = 0; j < SUBFRAME_LEN * 2; j++) {
1101  t = vector_ptr[j] >> shift;
1102  sum += t * t;
1103  tmp[j] = t;
1104  }
1105  }
1106 
1107  b0 = 0;
1108  for (j = 0; j < 11; j++)
1109  b0 += tmp[pos[(i / 2) * 11 + j]] * signs[(i / 2) * 11 + j];
1110  b0 = b0 * 2 * 2979LL + (1 << 29) >> 30; // approximated division by 11
1111 
1112  c = p->cur_gain * (p->cur_gain * SUBFRAME_LEN >> 5);
1113  if (shift * 2 + 3 >= 0)
1114  c >>= shift * 2 + 3;
1115  else
1116  c <<= -(shift * 2 + 3);
1117  c = (av_clipl_int32(sum << 1) - c) * 2979LL >> 15;
1118 
1119  delta = b0 * b0 * 2 - c;
1120  if (delta <= 0) {
1121  x = -b0;
1122  } else {
1123  delta = square_root(delta);
1124  x = delta - b0;
1125  t = delta + b0;
1126  if (FFABS(t) < FFABS(x))
1127  x = -t;
1128  }
1129  shift++;
1130  if (shift < 0)
1131  x >>= -shift;
1132  else
1133  x <<= shift;
1134  x = av_clip(x, -10000, 10000);
1135 
1136  for (j = 0; j < 11; j++) {
1137  idx = (i / 2) * 11 + j;
1138  vector_ptr[pos[idx]] = av_clip_int16(vector_ptr[pos[idx]] +
1139  (x * signs[idx] >> 15));
1140  }
1141 
1142  /* copy decoded data to serve as a history for the next decoded subframes */
1143  memcpy(vector_ptr + PITCH_MAX, vector_ptr,
1144  sizeof(*vector_ptr) * SUBFRAME_LEN * 2);
1145  vector_ptr += SUBFRAME_LEN * 2;
1146  }
1147  /* Save the excitation for the next frame */
1148  memcpy(p->prev_excitation, p->audio + LPC_ORDER + FRAME_LEN,
1149  PITCH_MAX * sizeof(*p->excitation));
1150 }
1151 
1152 static int g723_1_decode_frame(AVCodecContext *avctx, void *data,
1153  int *got_frame_ptr, AVPacket *avpkt)
1154 {
1155  G723_1_Context *p = avctx->priv_data;
1156  AVFrame *frame = data;
1157  const uint8_t *buf = avpkt->data;
1158  int buf_size = avpkt->size;
1159  int dec_mode = buf[0] & 3;
1160 
1161  PPFParam ppf[SUBFRAMES];
1162  int16_t cur_lsp[LPC_ORDER];
1163  int16_t lpc[SUBFRAMES * LPC_ORDER];
1164  int16_t acb_vector[SUBFRAME_LEN];
1165  int16_t *out;
1166  int bad_frame = 0, i, j, ret;
1167  int16_t *audio = p->audio;
1168 
1169  if (buf_size < frame_size[dec_mode]) {
1170  if (buf_size)
1171  av_log(avctx, AV_LOG_WARNING,
1172  "Expected %d bytes, got %d - skipping packet\n",
1173  frame_size[dec_mode], buf_size);
1174  *got_frame_ptr = 0;
1175  return buf_size;
1176  }
1177 
1178  if (unpack_bitstream(p, buf, buf_size) < 0) {
1179  bad_frame = 1;
1180  if (p->past_frame_type == ACTIVE_FRAME)
1182  else
1184  }
1185 
1186  frame->nb_samples = FRAME_LEN;
1187  if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
1188  return ret;
1189 
1190  out = (int16_t *)frame->data[0];
1191 
1192  if (p->cur_frame_type == ACTIVE_FRAME) {
1193  if (!bad_frame)
1194  p->erased_frames = 0;
1195  else if (p->erased_frames != 3)
1196  p->erased_frames++;
1197 
1198  inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
1199  lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
1200 
1201  /* Save the lsp_vector for the next frame */
1202  memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1203 
1204  /* Generate the excitation for the frame */
1205  memcpy(p->excitation, p->prev_excitation,
1206  PITCH_MAX * sizeof(*p->excitation));
1207  if (!p->erased_frames) {
1208  int16_t *vector_ptr = p->excitation + PITCH_MAX;
1209 
1210  /* Update interpolation gain memory */
1212  p->subframe[3].amp_index) >> 1];
1213  for (i = 0; i < SUBFRAMES; i++) {
1214  gen_fcb_excitation(vector_ptr, &p->subframe[i], p->cur_rate,
1215  p->pitch_lag[i >> 1], i);
1216  gen_acb_excitation(acb_vector, &p->excitation[SUBFRAME_LEN * i],
1217  p->pitch_lag[i >> 1], &p->subframe[i],
1218  p->cur_rate);
1219  /* Get the total excitation */
1220  for (j = 0; j < SUBFRAME_LEN; j++) {
1221  int v = av_clip_int16(vector_ptr[j] << 1);
1222  vector_ptr[j] = av_clip_int16(v + acb_vector[j]);
1223  }
1224  vector_ptr += SUBFRAME_LEN;
1225  }
1226 
1227  vector_ptr = p->excitation + PITCH_MAX;
1228 
1230  &p->sid_gain, &p->cur_gain);
1231 
1232  /* Peform pitch postfiltering */
1233  if (p->postfilter) {
1234  i = PITCH_MAX;
1235  for (j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1236  comp_ppf_coeff(p, i, p->pitch_lag[j >> 1],
1237  ppf + j, p->cur_rate);
1238 
1239  for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1241  vector_ptr + i,
1242  vector_ptr + i + ppf[j].index,
1243  ppf[j].sc_gain,
1244  ppf[j].opt_gain,
1245  1 << 14, 15, SUBFRAME_LEN);
1246  } else {
1247  audio = vector_ptr - LPC_ORDER;
1248  }
1249 
1250  /* Save the excitation for the next frame */
1251  memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
1252  PITCH_MAX * sizeof(*p->excitation));
1253  } else {
1254  p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
1255  if (p->erased_frames == 3) {
1256  /* Mute output */
1257  memset(p->excitation, 0,
1258  (FRAME_LEN + PITCH_MAX) * sizeof(*p->excitation));
1259  memset(p->prev_excitation, 0,
1260  PITCH_MAX * sizeof(*p->excitation));
1261  memset(frame->data[0], 0,
1262  (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
1263  } else {
1264  int16_t *buf = p->audio + LPC_ORDER;
1265 
1266  /* Regenerate frame */
1268  p->interp_gain, &p->random_seed);
1269 
1270  /* Save the excitation for the next frame */
1271  memcpy(p->prev_excitation, buf + (FRAME_LEN - PITCH_MAX),
1272  PITCH_MAX * sizeof(*p->excitation));
1273  }
1274  }
1276  } else {
1277  if (p->cur_frame_type == SID_FRAME) {
1279  inverse_quant(p->sid_lsp, p->prev_lsp, p->lsp_index, 0);
1280  } else if (p->past_frame_type == ACTIVE_FRAME) {
1281  p->sid_gain = estimate_sid_gain(p);
1282  }
1283 
1284  if (p->past_frame_type == ACTIVE_FRAME)
1285  p->cur_gain = p->sid_gain;
1286  else
1287  p->cur_gain = (p->cur_gain * 7 + p->sid_gain) >> 3;
1288  generate_noise(p);
1289  lsp_interpolate(lpc, p->sid_lsp, p->prev_lsp);
1290  /* Save the lsp_vector for the next frame */
1291  memcpy(p->prev_lsp, p->sid_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1292  }
1293 
1295 
1296  memcpy(p->audio, p->synth_mem, LPC_ORDER * sizeof(*p->audio));
1297  for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1298  ff_celp_lp_synthesis_filter(p->audio + i, &lpc[j * LPC_ORDER],
1299  audio + i, SUBFRAME_LEN, LPC_ORDER,
1300  0, 1, 1 << 12);
1301  memcpy(p->synth_mem, p->audio + FRAME_LEN, LPC_ORDER * sizeof(*p->audio));
1302 
1303  if (p->postfilter) {
1304  formant_postfilter(p, lpc, p->audio, out);
1305  } else { // if output is not postfiltered it should be scaled by 2
1306  for (i = 0; i < FRAME_LEN; i++)
1307  out[i] = av_clip_int16(p->audio[LPC_ORDER + i] << 1);
1308  }
1309 
1310  *got_frame_ptr = 1;
1311 
1312  return frame_size[dec_mode];
1313 }
1314 
1315 #define OFFSET(x) offsetof(G723_1_Context, x)
1316 #define AD AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1317 
1318 static const AVOption options[] = {
1319  { "postfilter", "postfilter on/off", OFFSET(postfilter), AV_OPT_TYPE_INT,
1320  { .i64 = 1 }, 0, 1, AD },
1321  { NULL }
1322 };
1323 
1324 
1325 static const AVClass g723_1dec_class = {
1326  .class_name = "G.723.1 decoder",
1327  .item_name = av_default_item_name,
1328  .option = options,
1329  .version = LIBAVUTIL_VERSION_INT,
1330 };
1331 
1333  .name = "g723_1",
1334  .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1335  .type = AVMEDIA_TYPE_AUDIO,
1336  .id = AV_CODEC_ID_G723_1,
1337  .priv_data_size = sizeof(G723_1_Context),
1340  .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
1341  .priv_class = &g723_1dec_class,
1342 };
1343 
1344 #if CONFIG_G723_1_ENCODER
1345 #define BITSTREAM_WRITER_LE
1346 #include "put_bits.h"
1347 
1348 static av_cold int g723_1_encode_init(AVCodecContext *avctx)
1349 {
1350  G723_1_Context *p = avctx->priv_data;
1351 
1352  if (avctx->sample_rate != 8000) {
1353  av_log(avctx, AV_LOG_ERROR, "Only 8000Hz sample rate supported\n");
1354  return -1;
1355  }
1356 
1357  if (avctx->channels != 1) {
1358  av_log(avctx, AV_LOG_ERROR, "Only mono supported\n");
1359  return AVERROR(EINVAL);
1360  }
1361 
1362  if (avctx->bit_rate == 6300) {
1363  p->cur_rate = RATE_6300;
1364  } else if (avctx->bit_rate == 5300) {
1365  av_log(avctx, AV_LOG_ERROR, "Bitrate not supported yet, use 6.3k\n");
1366  return AVERROR_PATCHWELCOME;
1367  } else {
1368  av_log(avctx, AV_LOG_ERROR,
1369  "Bitrate not supported, use 6.3k\n");
1370  return AVERROR(EINVAL);
1371  }
1372  avctx->frame_size = 240;
1373  memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(int16_t));
1374 
1375  return 0;
1376 }
1377 
1378 /**
1379  * Remove DC component from the input signal.
1380  *
1381  * @param buf input signal
1382  * @param fir zero memory
1383  * @param iir pole memory
1384  */
1385 static void highpass_filter(int16_t *buf, int16_t *fir, int *iir)
1386 {
1387  int i;
1388  for (i = 0; i < FRAME_LEN; i++) {
1389  *iir = (buf[i] << 15) + ((-*fir) << 15) + MULL2(*iir, 0x7f00);
1390  *fir = buf[i];
1391  buf[i] = av_clipl_int32((int64_t)*iir + (1 << 15)) >> 16;
1392  }
1393 }
1394 
1395 /**
1396  * Estimate autocorrelation of the input vector.
1397  *
1398  * @param buf input buffer
1399  * @param autocorr autocorrelation coefficients vector
1400  */
1401 static void comp_autocorr(int16_t *buf, int16_t *autocorr)
1402 {
1403  int i, scale, temp;
1404  int16_t vector[LPC_FRAME];
1405 
1406  scale_vector(vector, buf, LPC_FRAME);
1407 
1408  /* Apply the Hamming window */
1409  for (i = 0; i < LPC_FRAME; i++)
1410  vector[i] = (vector[i] * hamming_window[i] + (1 << 14)) >> 15;
1411 
1412  /* Compute the first autocorrelation coefficient */
1413  temp = ff_dot_product(vector, vector, LPC_FRAME);
1414 
1415  /* Apply a white noise correlation factor of (1025/1024) */
1416  temp += temp >> 10;
1417 
1418  /* Normalize */
1419  scale = normalize_bits_int32(temp);
1420  autocorr[0] = av_clipl_int32((int64_t)(temp << scale) +
1421  (1 << 15)) >> 16;
1422 
1423  /* Compute the remaining coefficients */
1424  if (!autocorr[0]) {
1425  memset(autocorr + 1, 0, LPC_ORDER * sizeof(int16_t));
1426  } else {
1427  for (i = 1; i <= LPC_ORDER; i++) {
1428  temp = ff_dot_product(vector, vector + i, LPC_FRAME - i);
1429  temp = MULL2((temp << scale), binomial_window[i - 1]);
1430  autocorr[i] = av_clipl_int32((int64_t)temp + (1 << 15)) >> 16;
1431  }
1432  }
1433 }
1434 
1435 /**
1436  * Use Levinson-Durbin recursion to compute LPC coefficients from
1437  * autocorrelation values.
1438  *
1439  * @param lpc LPC coefficients vector
1440  * @param autocorr autocorrelation coefficients vector
1441  * @param error prediction error
1442  */
1443 static void levinson_durbin(int16_t *lpc, int16_t *autocorr, int16_t error)
1444 {
1445  int16_t vector[LPC_ORDER];
1446  int16_t partial_corr;
1447  int i, j, temp;
1448 
1449  memset(lpc, 0, LPC_ORDER * sizeof(int16_t));
1450 
1451  for (i = 0; i < LPC_ORDER; i++) {
1452  /* Compute the partial correlation coefficient */
1453  temp = 0;
1454  for (j = 0; j < i; j++)
1455  temp -= lpc[j] * autocorr[i - j - 1];
1456  temp = ((autocorr[i] << 13) + temp) << 3;
1457 
1458  if (FFABS(temp) >= (error << 16))
1459  break;
1460 
1461  partial_corr = temp / (error << 1);
1462 
1463  lpc[i] = av_clipl_int32((int64_t)(partial_corr << 14) +
1464  (1 << 15)) >> 16;
1465 
1466  /* Update the prediction error */
1467  temp = MULL2(temp, partial_corr);
1468  error = av_clipl_int32((int64_t)(error << 16) - temp +
1469  (1 << 15)) >> 16;
1470 
1471  memcpy(vector, lpc, i * sizeof(int16_t));
1472  for (j = 0; j < i; j++) {
1473  temp = partial_corr * vector[i - j - 1] << 1;
1474  lpc[j] = av_clipl_int32((int64_t)(lpc[j] << 16) - temp +
1475  (1 << 15)) >> 16;
1476  }
1477  }
1478 }
1479 
1480 /**
1481  * Calculate LPC coefficients for the current frame.
1482  *
1483  * @param buf current frame
1484  * @param prev_data 2 trailing subframes of the previous frame
1485  * @param lpc LPC coefficients vector
1486  */
1487 static void comp_lpc_coeff(int16_t *buf, int16_t *lpc)
1488 {
1489  int16_t autocorr[(LPC_ORDER + 1) * SUBFRAMES];
1490  int16_t *autocorr_ptr = autocorr;
1491  int16_t *lpc_ptr = lpc;
1492  int i, j;
1493 
1494  for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1495  comp_autocorr(buf + i, autocorr_ptr);
1496  levinson_durbin(lpc_ptr, autocorr_ptr + 1, autocorr_ptr[0]);
1497 
1498  lpc_ptr += LPC_ORDER;
1499  autocorr_ptr += LPC_ORDER + 1;
1500  }
1501 }
1502 
1503 static void lpc2lsp(int16_t *lpc, int16_t *prev_lsp, int16_t *lsp)
1504 {
1505  int f[LPC_ORDER + 2]; ///< coefficients of the sum and difference
1506  ///< polynomials (F1, F2) ordered as
1507  ///< f1[0], f2[0], ...., f1[5], f2[5]
1508 
1509  int max, shift, cur_val, prev_val, count, p;
1510  int i, j;
1511  int64_t temp;
1512 
1513  /* Initialize f1[0] and f2[0] to 1 in Q25 */
1514  for (i = 0; i < LPC_ORDER; i++)
1515  lsp[i] = (lpc[i] * bandwidth_expand[i] + (1 << 14)) >> 15;
1516 
1517  /* Apply bandwidth expansion on the LPC coefficients */
1518  f[0] = f[1] = 1 << 25;
1519 
1520  /* Compute the remaining coefficients */
1521  for (i = 0; i < LPC_ORDER / 2; i++) {
1522  /* f1 */
1523  f[2 * i + 2] = -f[2 * i] - ((lsp[i] + lsp[LPC_ORDER - 1 - i]) << 12);
1524  /* f2 */
1525  f[2 * i + 3] = f[2 * i + 1] - ((lsp[i] - lsp[LPC_ORDER - 1 - i]) << 12);
1526  }
1527 
1528  /* Divide f1[5] and f2[5] by 2 for use in polynomial evaluation */
1529  f[LPC_ORDER] >>= 1;
1530  f[LPC_ORDER + 1] >>= 1;
1531 
1532  /* Normalize and shorten */
1533  max = FFABS(f[0]);
1534  for (i = 1; i < LPC_ORDER + 2; i++)
1535  max = FFMAX(max, FFABS(f[i]));
1536 
1537  shift = normalize_bits_int32(max);
1538 
1539  for (i = 0; i < LPC_ORDER + 2; i++)
1540  f[i] = av_clipl_int32((int64_t)(f[i] << shift) + (1 << 15)) >> 16;
1541 
1542  /**
1543  * Evaluate F1 and F2 at uniform intervals of pi/256 along the
1544  * unit circle and check for zero crossings.
1545  */
1546  p = 0;
1547  temp = 0;
1548  for (i = 0; i <= LPC_ORDER / 2; i++)
1549  temp += f[2 * i] * cos_tab[0];
1550  prev_val = av_clipl_int32(temp << 1);
1551  count = 0;
1552  for ( i = 1; i < COS_TBL_SIZE / 2; i++) {
1553  /* Evaluate */
1554  temp = 0;
1555  for (j = 0; j <= LPC_ORDER / 2; j++)
1556  temp += f[LPC_ORDER - 2 * j + p] * cos_tab[i * j % COS_TBL_SIZE];
1557  cur_val = av_clipl_int32(temp << 1);
1558 
1559  /* Check for sign change, indicating a zero crossing */
1560  if ((cur_val ^ prev_val) < 0) {
1561  int abs_cur = FFABS(cur_val);
1562  int abs_prev = FFABS(prev_val);
1563  int sum = abs_cur + abs_prev;
1564 
1565  shift = normalize_bits_int32(sum);
1566  sum <<= shift;
1567  abs_prev = abs_prev << shift >> 8;
1568  lsp[count++] = ((i - 1) << 7) + (abs_prev >> 1) / (sum >> 16);
1569 
1570  if (count == LPC_ORDER)
1571  break;
1572 
1573  /* Switch between sum and difference polynomials */
1574  p ^= 1;
1575 
1576  /* Evaluate */
1577  temp = 0;
1578  for (j = 0; j <= LPC_ORDER / 2; j++){
1579  temp += f[LPC_ORDER - 2 * j + p] *
1580  cos_tab[i * j % COS_TBL_SIZE];
1581  }
1582  cur_val = av_clipl_int32(temp<<1);
1583  }
1584  prev_val = cur_val;
1585  }
1586 
1587  if (count != LPC_ORDER)
1588  memcpy(lsp, prev_lsp, LPC_ORDER * sizeof(int16_t));
1589 }
1590 
1591 /**
1592  * Quantize the current LSP subvector.
1593  *
1594  * @param num band number
1595  * @param offset offset of the current subvector in an LPC_ORDER vector
1596  * @param size size of the current subvector
1597  */
1598 #define get_index(num, offset, size) \
1599 {\
1600  int error, max = -1;\
1601  int16_t temp[4];\
1602  int i, j;\
1603  for (i = 0; i < LSP_CB_SIZE; i++) {\
1604  for (j = 0; j < size; j++){\
1605  temp[j] = (weight[j + (offset)] * lsp_band##num[i][j] +\
1606  (1 << 14)) >> 15;\
1607  }\
1608  error = dot_product(lsp + (offset), temp, size) << 1;\
1609  error -= dot_product(lsp_band##num[i], temp, size);\
1610  if (error > max) {\
1611  max = error;\
1612  lsp_index[num] = i;\
1613  }\
1614  }\
1615 }
1616 
1617 /**
1618  * Vector quantize the LSP frequencies.
1619  *
1620  * @param lsp the current lsp vector
1621  * @param prev_lsp the previous lsp vector
1622  */
1623 static void lsp_quantize(uint8_t *lsp_index, int16_t *lsp, int16_t *prev_lsp)
1624 {
1625  int16_t weight[LPC_ORDER];
1626  int16_t min, max;
1627  int shift, i;
1628 
1629  /* Calculate the VQ weighting vector */
1630  weight[0] = (1 << 20) / (lsp[1] - lsp[0]);
1631  weight[LPC_ORDER - 1] = (1 << 20) /
1632  (lsp[LPC_ORDER - 1] - lsp[LPC_ORDER - 2]);
1633 
1634  for (i = 1; i < LPC_ORDER - 1; i++) {
1635  min = FFMIN(lsp[i] - lsp[i - 1], lsp[i + 1] - lsp[i]);
1636  if (min > 0x20)
1637  weight[i] = (1 << 20) / min;
1638  else
1639  weight[i] = INT16_MAX;
1640  }
1641 
1642  /* Normalize */
1643  max = 0;
1644  for (i = 0; i < LPC_ORDER; i++)
1645  max = FFMAX(weight[i], max);
1646 
1647  shift = normalize_bits_int16(max);
1648  for (i = 0; i < LPC_ORDER; i++) {
1649  weight[i] <<= shift;
1650  }
1651 
1652  /* Compute the VQ target vector */
1653  for (i = 0; i < LPC_ORDER; i++) {
1654  lsp[i] -= dc_lsp[i] +
1655  (((prev_lsp[i] - dc_lsp[i]) * 12288 + (1 << 14)) >> 15);
1656  }
1657 
1658  get_index(0, 0, 3);
1659  get_index(1, 3, 3);
1660  get_index(2, 6, 4);
1661 }
1662 
1663 /**
1664  * Apply the formant perceptual weighting filter.
1665  *
1666  * @param flt_coef filter coefficients
1667  * @param unq_lpc unquantized lpc vector
1668  */
1669 static void perceptual_filter(G723_1_Context *p, int16_t *flt_coef,
1670  int16_t *unq_lpc, int16_t *buf)
1671 {
1672  int16_t vector[FRAME_LEN + LPC_ORDER];
1673  int i, j, k, l = 0;
1674 
1675  memcpy(buf, p->iir_mem, sizeof(int16_t) * LPC_ORDER);
1676  memcpy(vector, p->fir_mem, sizeof(int16_t) * LPC_ORDER);
1677  memcpy(vector + LPC_ORDER, buf + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
1678 
1679  for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1680  for (k = 0; k < LPC_ORDER; k++) {
1681  flt_coef[k + 2 * l] = (unq_lpc[k + l] * percept_flt_tbl[0][k] +
1682  (1 << 14)) >> 15;
1683  flt_coef[k + 2 * l + LPC_ORDER] = (unq_lpc[k + l] *
1684  percept_flt_tbl[1][k] +
1685  (1 << 14)) >> 15;
1686  }
1687  iir_filter(flt_coef + 2 * l, flt_coef + 2 * l + LPC_ORDER, vector + i,
1688  buf + i, 0);
1689  l += LPC_ORDER;
1690  }
1691  memcpy(p->iir_mem, buf + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1692  memcpy(p->fir_mem, vector + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1693 }
1694 
1695 /**
1696  * Estimate the open loop pitch period.
1697  *
1698  * @param buf perceptually weighted speech
1699  * @param start estimation is carried out from this position
1700  */
1701 static int estimate_pitch(int16_t *buf, int start)
1702 {
1703  int max_exp = 32;
1704  int max_ccr = 0x4000;
1705  int max_eng = 0x7fff;
1706  int index = PITCH_MIN;
1707  int offset = start - PITCH_MIN + 1;
1708 
1709  int ccr, eng, orig_eng, ccr_eng, exp;
1710  int diff, temp;
1711 
1712  int i;
1713 
1714  orig_eng = ff_dot_product(buf + offset, buf + offset, HALF_FRAME_LEN);
1715 
1716  for (i = PITCH_MIN; i <= PITCH_MAX - 3; i++) {
1717  offset--;
1718 
1719  /* Update energy and compute correlation */
1720  orig_eng += buf[offset] * buf[offset] -
1721  buf[offset + HALF_FRAME_LEN] * buf[offset + HALF_FRAME_LEN];
1722  ccr = ff_dot_product(buf + start, buf + offset, HALF_FRAME_LEN);
1723  if (ccr <= 0)
1724  continue;
1725 
1726  /* Split into mantissa and exponent to maintain precision */
1727  exp = normalize_bits_int32(ccr);
1728  ccr = av_clipl_int32((int64_t)(ccr << exp) + (1 << 15)) >> 16;
1729  exp <<= 1;
1730  ccr *= ccr;
1731  temp = normalize_bits_int32(ccr);
1732  ccr = ccr << temp >> 16;
1733  exp += temp;
1734 
1735  temp = normalize_bits_int32(orig_eng);
1736  eng = av_clipl_int32((int64_t)(orig_eng << temp) + (1 << 15)) >> 16;
1737  exp -= temp;
1738 
1739  if (ccr >= eng) {
1740  exp--;
1741  ccr >>= 1;
1742  }
1743  if (exp > max_exp)
1744  continue;
1745 
1746  if (exp + 1 < max_exp)
1747  goto update;
1748 
1749  /* Equalize exponents before comparison */
1750  if (exp + 1 == max_exp)
1751  temp = max_ccr >> 1;
1752  else
1753  temp = max_ccr;
1754  ccr_eng = ccr * max_eng;
1755  diff = ccr_eng - eng * temp;
1756  if (diff > 0 && (i - index < PITCH_MIN || diff > ccr_eng >> 2)) {
1757 update:
1758  index = i;
1759  max_exp = exp;
1760  max_ccr = ccr;
1761  max_eng = eng;
1762  }
1763  }
1764  return index;
1765 }
1766 
1767 /**
1768  * Compute harmonic noise filter parameters.
1769  *
1770  * @param buf perceptually weighted speech
1771  * @param pitch_lag open loop pitch period
1772  * @param hf harmonic filter parameters
1773  */
1774 static void comp_harmonic_coeff(int16_t *buf, int16_t pitch_lag, HFParam *hf)
1775 {
1776  int ccr, eng, max_ccr, max_eng;
1777  int exp, max, diff;
1778  int energy[15];
1779  int i, j;
1780 
1781  for (i = 0, j = pitch_lag - 3; j <= pitch_lag + 3; i++, j++) {
1782  /* Compute residual energy */
1783  energy[i << 1] = ff_dot_product(buf - j, buf - j, SUBFRAME_LEN);
1784  /* Compute correlation */
1785  energy[(i << 1) + 1] = ff_dot_product(buf, buf - j, SUBFRAME_LEN);
1786  }
1787 
1788  /* Compute target energy */
1789  energy[14] = ff_dot_product(buf, buf, SUBFRAME_LEN);
1790 
1791  /* Normalize */
1792  max = 0;
1793  for (i = 0; i < 15; i++)
1794  max = FFMAX(max, FFABS(energy[i]));
1795 
1796  exp = normalize_bits_int32(max);
1797  for (i = 0; i < 15; i++) {
1798  energy[i] = av_clipl_int32((int64_t)(energy[i] << exp) +
1799  (1 << 15)) >> 16;
1800  }
1801 
1802  hf->index = -1;
1803  hf->gain = 0;
1804  max_ccr = 1;
1805  max_eng = 0x7fff;
1806 
1807  for (i = 0; i <= 6; i++) {
1808  eng = energy[i << 1];
1809  ccr = energy[(i << 1) + 1];
1810 
1811  if (ccr <= 0)
1812  continue;
1813 
1814  ccr = (ccr * ccr + (1 << 14)) >> 15;
1815  diff = ccr * max_eng - eng * max_ccr;
1816  if (diff > 0) {
1817  max_ccr = ccr;
1818  max_eng = eng;
1819  hf->index = i;
1820  }
1821  }
1822 
1823  if (hf->index == -1) {
1824  hf->index = pitch_lag;
1825  return;
1826  }
1827 
1828  eng = energy[14] * max_eng;
1829  eng = (eng >> 2) + (eng >> 3);
1830  ccr = energy[(hf->index << 1) + 1] * energy[(hf->index << 1) + 1];
1831  if (eng < ccr) {
1832  eng = energy[(hf->index << 1) + 1];
1833 
1834  if (eng >= max_eng)
1835  hf->gain = 0x2800;
1836  else
1837  hf->gain = ((eng << 15) / max_eng * 0x2800 + (1 << 14)) >> 15;
1838  }
1839  hf->index += pitch_lag - 3;
1840 }
1841 
1842 /**
1843  * Apply the harmonic noise shaping filter.
1844  *
1845  * @param hf filter parameters
1846  */
1847 static void harmonic_filter(HFParam *hf, const int16_t *src, int16_t *dest)
1848 {
1849  int i;
1850 
1851  for (i = 0; i < SUBFRAME_LEN; i++) {
1852  int64_t temp = hf->gain * src[i - hf->index] << 1;
1853  dest[i] = av_clipl_int32((src[i] << 16) - temp + (1 << 15)) >> 16;
1854  }
1855 }
1856 
1857 static void harmonic_noise_sub(HFParam *hf, const int16_t *src, int16_t *dest)
1858 {
1859  int i;
1860  for (i = 0; i < SUBFRAME_LEN; i++) {
1861  int64_t temp = hf->gain * src[i - hf->index] << 1;
1862  dest[i] = av_clipl_int32(((dest[i] - src[i]) << 16) + temp +
1863  (1 << 15)) >> 16;
1864 
1865  }
1866 }
1867 
1868 /**
1869  * Combined synthesis and formant perceptual weighting filer.
1870  *
1871  * @param qnt_lpc quantized lpc coefficients
1872  * @param perf_lpc perceptual filter coefficients
1873  * @param perf_fir perceptual filter fir memory
1874  * @param perf_iir perceptual filter iir memory
1875  * @param scale the filter output will be scaled by 2^scale
1876  */
1877 static void synth_percept_filter(int16_t *qnt_lpc, int16_t *perf_lpc,
1878  int16_t *perf_fir, int16_t *perf_iir,
1879  const int16_t *src, int16_t *dest, int scale)
1880 {
1881  int i, j;
1882  int16_t buf_16[SUBFRAME_LEN + LPC_ORDER];
1883  int64_t buf[SUBFRAME_LEN];
1884 
1885  int16_t *bptr_16 = buf_16 + LPC_ORDER;
1886 
1887  memcpy(buf_16, perf_fir, sizeof(int16_t) * LPC_ORDER);
1888  memcpy(dest - LPC_ORDER, perf_iir, sizeof(int16_t) * LPC_ORDER);
1889 
1890  for (i = 0; i < SUBFRAME_LEN; i++) {
1891  int64_t temp = 0;
1892  for (j = 1; j <= LPC_ORDER; j++)
1893  temp -= qnt_lpc[j - 1] * bptr_16[i - j];
1894 
1895  buf[i] = (src[i] << 15) + (temp << 3);
1896  bptr_16[i] = av_clipl_int32(buf[i] + (1 << 15)) >> 16;
1897  }
1898 
1899  for (i = 0; i < SUBFRAME_LEN; i++) {
1900  int64_t fir = 0, iir = 0;
1901  for (j = 1; j <= LPC_ORDER; j++) {
1902  fir -= perf_lpc[j - 1] * bptr_16[i - j];
1903  iir += perf_lpc[j + LPC_ORDER - 1] * dest[i - j];
1904  }
1905  dest[i] = av_clipl_int32(((buf[i] + (fir << 3)) << scale) + (iir << 3) +
1906  (1 << 15)) >> 16;
1907  }
1908  memcpy(perf_fir, buf_16 + SUBFRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1909  memcpy(perf_iir, dest + SUBFRAME_LEN - LPC_ORDER,
1910  sizeof(int16_t) * LPC_ORDER);
1911 }
1912 
1913 /**
1914  * Compute the adaptive codebook contribution.
1915  *
1916  * @param buf input signal
1917  * @param index the current subframe index
1918  */
1919 static void acb_search(G723_1_Context *p, int16_t *residual,
1920  int16_t *impulse_resp, const int16_t *buf,
1921  int index)
1922 {
1923 
1924  int16_t flt_buf[PITCH_ORDER][SUBFRAME_LEN];
1925 
1926  const int16_t *cb_tbl = adaptive_cb_gain85;
1927 
1928  int ccr_buf[PITCH_ORDER * SUBFRAMES << 2];
1929 
1930  int pitch_lag = p->pitch_lag[index >> 1];
1931  int acb_lag = 1;
1932  int acb_gain = 0;
1933  int odd_frame = index & 1;
1934  int iter = 3 + odd_frame;
1935  int count = 0;
1936  int tbl_size = 85;
1937 
1938  int i, j, k, l, max;
1939  int64_t temp;
1940 
1941  if (!odd_frame) {
1942  if (pitch_lag == PITCH_MIN)
1943  pitch_lag++;
1944  else
1945  pitch_lag = FFMIN(pitch_lag, PITCH_MAX - 5);
1946  }
1947 
1948  for (i = 0; i < iter; i++) {
1949  get_residual(residual, p->prev_excitation, pitch_lag + i - 1);
1950 
1951  for (j = 0; j < SUBFRAME_LEN; j++) {
1952  temp = 0;
1953  for (k = 0; k <= j; k++)
1954  temp += residual[PITCH_ORDER - 1 + k] * impulse_resp[j - k];
1955  flt_buf[PITCH_ORDER - 1][j] = av_clipl_int32((temp << 1) +
1956  (1 << 15)) >> 16;
1957  }
1958 
1959  for (j = PITCH_ORDER - 2; j >= 0; j--) {
1960  flt_buf[j][0] = ((residual[j] << 13) + (1 << 14)) >> 15;
1961  for (k = 1; k < SUBFRAME_LEN; k++) {
1962  temp = (flt_buf[j + 1][k - 1] << 15) +
1963  residual[j] * impulse_resp[k];
1964  flt_buf[j][k] = av_clipl_int32((temp << 1) + (1 << 15)) >> 16;
1965  }
1966  }
1967 
1968  /* Compute crosscorrelation with the signal */
1969  for (j = 0; j < PITCH_ORDER; j++) {
1970  temp = ff_dot_product(buf, flt_buf[j], SUBFRAME_LEN);
1971  ccr_buf[count++] = av_clipl_int32(temp << 1);
1972  }
1973 
1974  /* Compute energies */
1975  for (j = 0; j < PITCH_ORDER; j++) {
1976  ccr_buf[count++] = dot_product(flt_buf[j], flt_buf[j],
1977  SUBFRAME_LEN);
1978  }
1979 
1980  for (j = 1; j < PITCH_ORDER; j++) {
1981  for (k = 0; k < j; k++) {
1982  temp = ff_dot_product(flt_buf[j], flt_buf[k], SUBFRAME_LEN);
1983  ccr_buf[count++] = av_clipl_int32(temp<<2);
1984  }
1985  }
1986  }
1987 
1988  /* Normalize and shorten */
1989  max = 0;
1990  for (i = 0; i < 20 * iter; i++)
1991  max = FFMAX(max, FFABS(ccr_buf[i]));
1992 
1993  temp = normalize_bits_int32(max);
1994 
1995  for (i = 0; i < 20 * iter; i++){
1996  ccr_buf[i] = av_clipl_int32((int64_t)(ccr_buf[i] << temp) +
1997  (1 << 15)) >> 16;
1998  }
1999 
2000  max = 0;
2001  for (i = 0; i < iter; i++) {
2002  /* Select quantization table */
2003  if (!odd_frame && pitch_lag + i - 1 >= SUBFRAME_LEN - 2 ||
2004  odd_frame && pitch_lag >= SUBFRAME_LEN - 2) {
2005  cb_tbl = adaptive_cb_gain170;
2006  tbl_size = 170;
2007  }
2008 
2009  for (j = 0, k = 0; j < tbl_size; j++, k += 20) {
2010  temp = 0;
2011  for (l = 0; l < 20; l++)
2012  temp += ccr_buf[20 * i + l] * cb_tbl[k + l];
2013  temp = av_clipl_int32(temp);
2014 
2015  if (temp > max) {
2016  max = temp;
2017  acb_gain = j;
2018  acb_lag = i;
2019  }
2020  }
2021  }
2022 
2023  if (!odd_frame) {
2024  pitch_lag += acb_lag - 1;
2025  acb_lag = 1;
2026  }
2027 
2028  p->pitch_lag[index >> 1] = pitch_lag;
2029  p->subframe[index].ad_cb_lag = acb_lag;
2030  p->subframe[index].ad_cb_gain = acb_gain;
2031 }
2032 
2033 /**
2034  * Subtract the adaptive codebook contribution from the input
2035  * to obtain the residual.
2036  *
2037  * @param buf target vector
2038  */
2039 static void sub_acb_contrib(const int16_t *residual, const int16_t *impulse_resp,
2040  int16_t *buf)
2041 {
2042  int i, j;
2043  /* Subtract adaptive CB contribution to obtain the residual */
2044  for (i = 0; i < SUBFRAME_LEN; i++) {
2045  int64_t temp = buf[i] << 14;
2046  for (j = 0; j <= i; j++)
2047  temp -= residual[j] * impulse_resp[i - j];
2048 
2049  buf[i] = av_clipl_int32((temp << 2) + (1 << 15)) >> 16;
2050  }
2051 }
2052 
2053 /**
2054  * Quantize the residual signal using the fixed codebook (MP-MLQ).
2055  *
2056  * @param optim optimized fixed codebook parameters
2057  * @param buf excitation vector
2058  */
2059 static void get_fcb_param(FCBParam *optim, int16_t *impulse_resp,
2060  int16_t *buf, int pulse_cnt, int pitch_lag)
2061 {
2062  FCBParam param;
2063  int16_t impulse_r[SUBFRAME_LEN];
2064  int16_t temp_corr[SUBFRAME_LEN];
2065  int16_t impulse_corr[SUBFRAME_LEN];
2066 
2067  int ccr1[SUBFRAME_LEN];
2068  int ccr2[SUBFRAME_LEN];
2069  int amp, err, max, max_amp_index, min, scale, i, j, k, l;
2070 
2071  int64_t temp;
2072 
2073  /* Update impulse response */
2074  memcpy(impulse_r, impulse_resp, sizeof(int16_t) * SUBFRAME_LEN);
2075  param.dirac_train = 0;
2076  if (pitch_lag < SUBFRAME_LEN - 2) {
2077  param.dirac_train = 1;
2078  gen_dirac_train(impulse_r, pitch_lag);
2079  }
2080 
2081  for (i = 0; i < SUBFRAME_LEN; i++)
2082  temp_corr[i] = impulse_r[i] >> 1;
2083 
2084  /* Compute impulse response autocorrelation */
2085  temp = dot_product(temp_corr, temp_corr, SUBFRAME_LEN);
2086 
2087  scale = normalize_bits_int32(temp);
2088  impulse_corr[0] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
2089 
2090  for (i = 1; i < SUBFRAME_LEN; i++) {
2091  temp = dot_product(temp_corr + i, temp_corr, SUBFRAME_LEN - i);
2092  impulse_corr[i] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
2093  }
2094 
2095  /* Compute crosscorrelation of impulse response with residual signal */
2096  scale -= 4;
2097  for (i = 0; i < SUBFRAME_LEN; i++){
2098  temp = dot_product(buf + i, impulse_r, SUBFRAME_LEN - i);
2099  if (scale < 0)
2100  ccr1[i] = temp >> -scale;
2101  else
2102  ccr1[i] = av_clipl_int32(temp << scale);
2103  }
2104 
2105  /* Search loop */
2106  for (i = 0; i < GRID_SIZE; i++) {
2107  /* Maximize the crosscorrelation */
2108  max = 0;
2109  for (j = i; j < SUBFRAME_LEN; j += GRID_SIZE) {
2110  temp = FFABS(ccr1[j]);
2111  if (temp >= max) {
2112  max = temp;
2113  param.pulse_pos[0] = j;
2114  }
2115  }
2116 
2117  /* Quantize the gain (max crosscorrelation/impulse_corr[0]) */
2118  amp = max;
2119  min = 1 << 30;
2120  max_amp_index = GAIN_LEVELS - 2;
2121  for (j = max_amp_index; j >= 2; j--) {
2122  temp = av_clipl_int32((int64_t)fixed_cb_gain[j] *
2123  impulse_corr[0] << 1);
2124  temp = FFABS(temp - amp);
2125  if (temp < min) {
2126  min = temp;
2127  max_amp_index = j;
2128  }
2129  }
2130 
2131  max_amp_index--;
2132  /* Select additional gain values */
2133  for (j = 1; j < 5; j++) {
2134  for (k = i; k < SUBFRAME_LEN; k += GRID_SIZE) {
2135  temp_corr[k] = 0;
2136  ccr2[k] = ccr1[k];
2137  }
2138  param.amp_index = max_amp_index + j - 2;
2139  amp = fixed_cb_gain[param.amp_index];
2140 
2141  param.pulse_sign[0] = (ccr2[param.pulse_pos[0]] < 0) ? -amp : amp;
2142  temp_corr[param.pulse_pos[0]] = 1;
2143 
2144  for (k = 1; k < pulse_cnt; k++) {
2145  max = -1 << 30;
2146  for (l = i; l < SUBFRAME_LEN; l += GRID_SIZE) {
2147  if (temp_corr[l])
2148  continue;
2149  temp = impulse_corr[FFABS(l - param.pulse_pos[k - 1])];
2150  temp = av_clipl_int32((int64_t)temp *
2151  param.pulse_sign[k - 1] << 1);
2152  ccr2[l] -= temp;
2153  temp = FFABS(ccr2[l]);
2154  if (temp > max) {
2155  max = temp;
2156  param.pulse_pos[k] = l;
2157  }
2158  }
2159 
2160  param.pulse_sign[k] = (ccr2[param.pulse_pos[k]] < 0) ?
2161  -amp : amp;
2162  temp_corr[param.pulse_pos[k]] = 1;
2163  }
2164 
2165  /* Create the error vector */
2166  memset(temp_corr, 0, sizeof(int16_t) * SUBFRAME_LEN);
2167 
2168  for (k = 0; k < pulse_cnt; k++)
2169  temp_corr[param.pulse_pos[k]] = param.pulse_sign[k];
2170 
2171  for (k = SUBFRAME_LEN - 1; k >= 0; k--) {
2172  temp = 0;
2173  for (l = 0; l <= k; l++) {
2174  int prod = av_clipl_int32((int64_t)temp_corr[l] *
2175  impulse_r[k - l] << 1);
2176  temp = av_clipl_int32(temp + prod);
2177  }
2178  temp_corr[k] = temp << 2 >> 16;
2179  }
2180 
2181  /* Compute square of error */
2182  err = 0;
2183  for (k = 0; k < SUBFRAME_LEN; k++) {
2184  int64_t prod;
2185  prod = av_clipl_int32((int64_t)buf[k] * temp_corr[k] << 1);
2186  err = av_clipl_int32(err - prod);
2187  prod = av_clipl_int32((int64_t)temp_corr[k] * temp_corr[k]);
2188  err = av_clipl_int32(err + prod);
2189  }
2190 
2191  /* Minimize */
2192  if (err < optim->min_err) {
2193  optim->min_err = err;
2194  optim->grid_index = i;
2195  optim->amp_index = param.amp_index;
2196  optim->dirac_train = param.dirac_train;
2197 
2198  for (k = 0; k < pulse_cnt; k++) {
2199  optim->pulse_sign[k] = param.pulse_sign[k];
2200  optim->pulse_pos[k] = param.pulse_pos[k];
2201  }
2202  }
2203  }
2204  }
2205 }
2206 
2207 /**
2208  * Encode the pulse position and gain of the current subframe.
2209  *
2210  * @param optim optimized fixed CB parameters
2211  * @param buf excitation vector
2212  */
2213 static void pack_fcb_param(G723_1_Subframe *subfrm, FCBParam *optim,
2214  int16_t *buf, int pulse_cnt)
2215 {
2216  int i, j;
2217 
2218  j = PULSE_MAX - pulse_cnt;
2219 
2220  subfrm->pulse_sign = 0;
2221  subfrm->pulse_pos = 0;
2222 
2223  for (i = 0; i < SUBFRAME_LEN >> 1; i++) {
2224  int val = buf[optim->grid_index + (i << 1)];
2225  if (!val) {
2226  subfrm->pulse_pos += combinatorial_table[j][i];
2227  } else {
2228  subfrm->pulse_sign <<= 1;
2229  if (val < 0) subfrm->pulse_sign++;
2230  j++;
2231 
2232  if (j == PULSE_MAX) break;
2233  }
2234  }
2235  subfrm->amp_index = optim->amp_index;
2236  subfrm->grid_index = optim->grid_index;
2237  subfrm->dirac_train = optim->dirac_train;
2238 }
2239 
2240 /**
2241  * Compute the fixed codebook excitation.
2242  *
2243  * @param buf target vector
2244  * @param impulse_resp impulse response of the combined filter
2245  */
2246 static void fcb_search(G723_1_Context *p, int16_t *impulse_resp,
2247  int16_t *buf, int index)
2248 {
2249  FCBParam optim;
2250  int pulse_cnt = pulses[index];
2251  int i;
2252 
2253  optim.min_err = 1 << 30;
2254  get_fcb_param(&optim, impulse_resp, buf, pulse_cnt, SUBFRAME_LEN);
2255 
2256  if (p->pitch_lag[index >> 1] < SUBFRAME_LEN - 2) {
2257  get_fcb_param(&optim, impulse_resp, buf, pulse_cnt,
2258  p->pitch_lag[index >> 1]);
2259  }
2260 
2261  /* Reconstruct the excitation */
2262  memset(buf, 0, sizeof(int16_t) * SUBFRAME_LEN);
2263  for (i = 0; i < pulse_cnt; i++)
2264  buf[optim.pulse_pos[i]] = optim.pulse_sign[i];
2265 
2266  pack_fcb_param(&p->subframe[index], &optim, buf, pulse_cnt);
2267 
2268  if (optim.dirac_train)
2269  gen_dirac_train(buf, p->pitch_lag[index >> 1]);
2270 }
2271 
2272 /**
2273  * Pack the frame parameters into output bitstream.
2274  *
2275  * @param frame output buffer
2276  * @param size size of the buffer
2277  */
2278 static int pack_bitstream(G723_1_Context *p, unsigned char *frame, int size)
2279 {
2280  PutBitContext pb;
2281  int info_bits, i, temp;
2282 
2283  init_put_bits(&pb, frame, size);
2284 
2285  if (p->cur_rate == RATE_6300) {
2286  info_bits = 0;
2287  put_bits(&pb, 2, info_bits);
2288  }else
2289  av_assert0(0);
2290 
2291  put_bits(&pb, 8, p->lsp_index[2]);
2292  put_bits(&pb, 8, p->lsp_index[1]);
2293  put_bits(&pb, 8, p->lsp_index[0]);
2294 
2295  put_bits(&pb, 7, p->pitch_lag[0] - PITCH_MIN);
2296  put_bits(&pb, 2, p->subframe[1].ad_cb_lag);
2297  put_bits(&pb, 7, p->pitch_lag[1] - PITCH_MIN);
2298  put_bits(&pb, 2, p->subframe[3].ad_cb_lag);
2299 
2300  /* Write 12 bit combined gain */
2301  for (i = 0; i < SUBFRAMES; i++) {
2302  temp = p->subframe[i].ad_cb_gain * GAIN_LEVELS +
2303  p->subframe[i].amp_index;
2304  if (p->cur_rate == RATE_6300)
2305  temp += p->subframe[i].dirac_train << 11;
2306  put_bits(&pb, 12, temp);
2307  }
2308 
2309  put_bits(&pb, 1, p->subframe[0].grid_index);
2310  put_bits(&pb, 1, p->subframe[1].grid_index);
2311  put_bits(&pb, 1, p->subframe[2].grid_index);
2312  put_bits(&pb, 1, p->subframe[3].grid_index);
2313 
2314  if (p->cur_rate == RATE_6300) {
2315  skip_put_bits(&pb, 1); /* reserved bit */
2316 
2317  /* Write 13 bit combined position index */
2318  temp = (p->subframe[0].pulse_pos >> 16) * 810 +
2319  (p->subframe[1].pulse_pos >> 14) * 90 +
2320  (p->subframe[2].pulse_pos >> 16) * 9 +
2321  (p->subframe[3].pulse_pos >> 14);
2322  put_bits(&pb, 13, temp);
2323 
2324  put_bits(&pb, 16, p->subframe[0].pulse_pos & 0xffff);
2325  put_bits(&pb, 14, p->subframe[1].pulse_pos & 0x3fff);
2326  put_bits(&pb, 16, p->subframe[2].pulse_pos & 0xffff);
2327  put_bits(&pb, 14, p->subframe[3].pulse_pos & 0x3fff);
2328 
2329  put_bits(&pb, 6, p->subframe[0].pulse_sign);
2330  put_bits(&pb, 5, p->subframe[1].pulse_sign);
2331  put_bits(&pb, 6, p->subframe[2].pulse_sign);
2332  put_bits(&pb, 5, p->subframe[3].pulse_sign);
2333  }
2334 
2335  flush_put_bits(&pb);
2336  return frame_size[info_bits];
2337 }
2338 
2339 static int g723_1_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
2340  const AVFrame *frame, int *got_packet_ptr)
2341 {
2342  G723_1_Context *p = avctx->priv_data;
2343  int16_t unq_lpc[LPC_ORDER * SUBFRAMES];
2344  int16_t qnt_lpc[LPC_ORDER * SUBFRAMES];
2345  int16_t cur_lsp[LPC_ORDER];
2346  int16_t weighted_lpc[LPC_ORDER * SUBFRAMES << 1];
2347  int16_t vector[FRAME_LEN + PITCH_MAX];
2348  int offset, ret;
2349  int16_t *in_orig = av_memdup(frame->data[0], frame->nb_samples * sizeof(int16_t));
2350  int16_t *in = in_orig;
2351 
2352  HFParam hf[4];
2353  int i, j;
2354 
2355  if (!in)
2356  return AVERROR(ENOMEM);
2357 
2358  highpass_filter(in, &p->hpf_fir_mem, &p->hpf_iir_mem);
2359 
2360  memcpy(vector, p->prev_data, HALF_FRAME_LEN * sizeof(int16_t));
2361  memcpy(vector + HALF_FRAME_LEN, in, FRAME_LEN * sizeof(int16_t));
2362 
2363  comp_lpc_coeff(vector, unq_lpc);
2364  lpc2lsp(&unq_lpc[LPC_ORDER * 3], p->prev_lsp, cur_lsp);
2365  lsp_quantize(p->lsp_index, cur_lsp, p->prev_lsp);
2366 
2367  /* Update memory */
2368  memcpy(vector + LPC_ORDER, p->prev_data + SUBFRAME_LEN,
2369  sizeof(int16_t) * SUBFRAME_LEN);
2370  memcpy(vector + LPC_ORDER + SUBFRAME_LEN, in,
2371  sizeof(int16_t) * (HALF_FRAME_LEN + SUBFRAME_LEN));
2372  memcpy(p->prev_data, in + HALF_FRAME_LEN,
2373  sizeof(int16_t) * HALF_FRAME_LEN);
2374  memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2375 
2376  perceptual_filter(p, weighted_lpc, unq_lpc, vector);
2377 
2378  memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2379  memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2380  memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2381 
2382  scale_vector(vector, vector, FRAME_LEN + PITCH_MAX);
2383 
2384  p->pitch_lag[0] = estimate_pitch(vector, PITCH_MAX);
2385  p->pitch_lag[1] = estimate_pitch(vector, PITCH_MAX + HALF_FRAME_LEN);
2386 
2387  for (i = PITCH_MAX, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2388  comp_harmonic_coeff(vector + i, p->pitch_lag[j >> 1], hf + j);
2389 
2390  memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2391  memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2392  memcpy(p->prev_weight_sig, vector + FRAME_LEN, sizeof(int16_t) * PITCH_MAX);
2393 
2394  for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2395  harmonic_filter(hf + j, vector + PITCH_MAX + i, in + i);
2396 
2397  inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, 0);
2398  lsp_interpolate(qnt_lpc, cur_lsp, p->prev_lsp);
2399 
2400  memcpy(p->prev_lsp, cur_lsp, sizeof(int16_t) * LPC_ORDER);
2401 
2402  offset = 0;
2403  for (i = 0; i < SUBFRAMES; i++) {
2404  int16_t impulse_resp[SUBFRAME_LEN];
2405  int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
2406  int16_t flt_in[SUBFRAME_LEN];
2407  int16_t zero[LPC_ORDER], fir[LPC_ORDER], iir[LPC_ORDER];
2408 
2409  /**
2410  * Compute the combined impulse response of the synthesis filter,
2411  * formant perceptual weighting filter and harmonic noise shaping filter
2412  */
2413  memset(zero, 0, sizeof(int16_t) * LPC_ORDER);
2414  memset(vector, 0, sizeof(int16_t) * PITCH_MAX);
2415  memset(flt_in, 0, sizeof(int16_t) * SUBFRAME_LEN);
2416 
2417  flt_in[0] = 1 << 13; /* Unit impulse */
2418  synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2419  zero, zero, flt_in, vector + PITCH_MAX, 1);
2420  harmonic_filter(hf + i, vector + PITCH_MAX, impulse_resp);
2421 
2422  /* Compute the combined zero input response */
2423  flt_in[0] = 0;
2424  memcpy(fir, p->perf_fir_mem, sizeof(int16_t) * LPC_ORDER);
2425  memcpy(iir, p->perf_iir_mem, sizeof(int16_t) * LPC_ORDER);
2426 
2427  synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2428  fir, iir, flt_in, vector + PITCH_MAX, 0);
2429  memcpy(vector, p->harmonic_mem, sizeof(int16_t) * PITCH_MAX);
2430  harmonic_noise_sub(hf + i, vector + PITCH_MAX, in);
2431 
2432  acb_search(p, residual, impulse_resp, in, i);
2433  gen_acb_excitation(residual, p->prev_excitation,p->pitch_lag[i >> 1],
2434  &p->subframe[i], p->cur_rate);
2435  sub_acb_contrib(residual, impulse_resp, in);
2436 
2437  fcb_search(p, impulse_resp, in, i);
2438 
2439  /* Reconstruct the excitation */
2440  gen_acb_excitation(impulse_resp, p->prev_excitation, p->pitch_lag[i >> 1],
2441  &p->subframe[i], RATE_6300);
2442 
2443  memmove(p->prev_excitation, p->prev_excitation + SUBFRAME_LEN,
2444  sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2445  for (j = 0; j < SUBFRAME_LEN; j++)
2446  in[j] = av_clip_int16((in[j] << 1) + impulse_resp[j]);
2447  memcpy(p->prev_excitation + PITCH_MAX - SUBFRAME_LEN, in,
2448  sizeof(int16_t) * SUBFRAME_LEN);
2449 
2450  /* Update filter memories */
2451  synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2452  p->perf_fir_mem, p->perf_iir_mem,
2453  in, vector + PITCH_MAX, 0);
2454  memmove(p->harmonic_mem, p->harmonic_mem + SUBFRAME_LEN,
2455  sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2456  memcpy(p->harmonic_mem + PITCH_MAX - SUBFRAME_LEN, vector + PITCH_MAX,
2457  sizeof(int16_t) * SUBFRAME_LEN);
2458 
2459  in += SUBFRAME_LEN;
2460  offset += LPC_ORDER;
2461  }
2462 
2463  av_freep(&in_orig); in = NULL;
2464 
2465  if ((ret = ff_alloc_packet2(avctx, avpkt, 24)) < 0)
2466  return ret;
2467 
2468  *got_packet_ptr = 1;
2469  avpkt->size = pack_bitstream(p, avpkt->data, avpkt->size);
2470  return 0;
2471 }
2472 
2473 AVCodec ff_g723_1_encoder = {
2474  .name = "g723_1",
2475  .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
2476  .type = AVMEDIA_TYPE_AUDIO,
2477  .id = AV_CODEC_ID_G723_1,
2478  .priv_data_size = sizeof(G723_1_Context),
2479  .init = g723_1_encode_init,
2480  .encode2 = g723_1_encode_frame,
2481  .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_S16,
2483 };
2484 #endif