FFmpeg
 All Data Structures 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
29 #include "libavutil/channel_layout.h"
30 #include "libavutil/mem.h"
31 #include "libavutil/opt.h"
32 #include "avcodec.h"
33 #include "internal.h"
34 #include "get_bits.h"
35 #include "acelp_vectors.h"
36 #include "celp_filters.h"
37 #include "celp_math.h"
38 #include "g723_1_data.h"
39 #include "internal.h"
40 
41 #define CNG_RANDOM_SEED 12345
42 
43 typedef struct g723_1_context {
44  AVClass *class;
45 
46  G723_1_Subframe subframe[4];
47  enum FrameType cur_frame_type;
48  enum FrameType past_frame_type;
49  enum Rate cur_rate;
50  uint8_t lsp_index[LSP_BANDS];
51  int pitch_lag[2];
52  int erased_frames;
53 
54  int16_t prev_lsp[LPC_ORDER];
55  int16_t sid_lsp[LPC_ORDER];
56  int16_t prev_excitation[PITCH_MAX];
57  int16_t excitation[PITCH_MAX + FRAME_LEN + 4];
58  int16_t synth_mem[LPC_ORDER];
59  int16_t fir_mem[LPC_ORDER];
60  int iir_mem[LPC_ORDER];
61 
62  int random_seed;
63  int cng_random_seed;
64  int interp_index;
65  int interp_gain;
66  int sid_gain;
67  int cur_gain;
68  int reflection_coef;
69  int pf_gain; ///< formant postfilter
70  ///< gain scaling unit memory
71  int postfilter;
72 
73  int16_t audio[FRAME_LEN + LPC_ORDER + PITCH_MAX + 4];
74  int16_t prev_data[HALF_FRAME_LEN];
75  int16_t prev_weight_sig[PITCH_MAX];
76 
77 
78  int16_t hpf_fir_mem; ///< highpass filter fir
79  int hpf_iir_mem; ///< and iir memories
80  int16_t perf_fir_mem[LPC_ORDER]; ///< perceptual filter fir
81  int16_t perf_iir_mem[LPC_ORDER]; ///< and iir memories
82 
83  int16_t harmonic_mem[PITCH_MAX];
84 } G723_1_Context;
85 
86 static av_cold int g723_1_decode_init(AVCodecContext *avctx)
87 {
88  G723_1_Context *p = avctx->priv_data;
89 
90  avctx->channel_layout = AV_CH_LAYOUT_MONO;
91  avctx->sample_fmt = AV_SAMPLE_FMT_S16;
92  avctx->channels = 1;
93  p->pf_gain = 1 << 12;
94 
95  memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
96  memcpy(p->sid_lsp, dc_lsp, LPC_ORDER * sizeof(*p->sid_lsp));
97 
98  p->cng_random_seed = CNG_RANDOM_SEED;
99  p->past_frame_type = SID_FRAME;
100 
101  return 0;
102 }
103 
104 /**
105  * Unpack the frame into parameters.
106  *
107  * @param p the context
108  * @param buf pointer to the input buffer
109  * @param buf_size size of the input buffer
110  */
111 static int unpack_bitstream(G723_1_Context *p, const uint8_t *buf,
112  int buf_size)
113 {
114  GetBitContext gb;
115  int ad_cb_len;
116  int temp, info_bits, i;
117 
118  init_get_bits(&gb, buf, buf_size * 8);
119 
120  /* Extract frame type and rate info */
121  info_bits = get_bits(&gb, 2);
122 
123  if (info_bits == 3) {
124  p->cur_frame_type = UNTRANSMITTED_FRAME;
125  return 0;
126  }
127 
128  /* Extract 24 bit lsp indices, 8 bit for each band */
129  p->lsp_index[2] = get_bits(&gb, 8);
130  p->lsp_index[1] = get_bits(&gb, 8);
131  p->lsp_index[0] = get_bits(&gb, 8);
132 
133  if (info_bits == 2) {
134  p->cur_frame_type = SID_FRAME;
135  p->subframe[0].amp_index = get_bits(&gb, 6);
136  return 0;
137  }
138 
139  /* Extract the info common to both rates */
140  p->cur_rate = info_bits ? RATE_5300 : RATE_6300;
141  p->cur_frame_type = ACTIVE_FRAME;
142 
143  p->pitch_lag[0] = get_bits(&gb, 7);
144  if (p->pitch_lag[0] > 123) /* test if forbidden code */
145  return -1;
146  p->pitch_lag[0] += PITCH_MIN;
147  p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
148 
149  p->pitch_lag[1] = get_bits(&gb, 7);
150  if (p->pitch_lag[1] > 123)
151  return -1;
152  p->pitch_lag[1] += PITCH_MIN;
153  p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
154  p->subframe[0].ad_cb_lag = 1;
155  p->subframe[2].ad_cb_lag = 1;
156 
157  for (i = 0; i < SUBFRAMES; i++) {
158  /* Extract combined gain */
159  temp = get_bits(&gb, 12);
160  ad_cb_len = 170;
161  p->subframe[i].dirac_train = 0;
162  if (p->cur_rate == RATE_6300 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
163  p->subframe[i].dirac_train = temp >> 11;
164  temp &= 0x7FF;
165  ad_cb_len = 85;
166  }
167  p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
168  if (p->subframe[i].ad_cb_gain < ad_cb_len) {
169  p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
170  GAIN_LEVELS;
171  } else {
172  return -1;
173  }
174  }
175 
176  p->subframe[0].grid_index = get_bits1(&gb);
177  p->subframe[1].grid_index = get_bits1(&gb);
178  p->subframe[2].grid_index = get_bits1(&gb);
179  p->subframe[3].grid_index = get_bits1(&gb);
180 
181  if (p->cur_rate == RATE_6300) {
182  skip_bits1(&gb); /* skip reserved bit */
183 
184  /* Compute pulse_pos index using the 13-bit combined position index */
185  temp = get_bits(&gb, 13);
186  p->subframe[0].pulse_pos = temp / 810;
187 
188  temp -= p->subframe[0].pulse_pos * 810;
189  p->subframe[1].pulse_pos = FASTDIV(temp, 90);
190 
191  temp -= p->subframe[1].pulse_pos * 90;
192  p->subframe[2].pulse_pos = FASTDIV(temp, 9);
193  p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
194 
195  p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
196  get_bits(&gb, 16);
197  p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
198  get_bits(&gb, 14);
199  p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
200  get_bits(&gb, 16);
201  p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
202  get_bits(&gb, 14);
203 
204  p->subframe[0].pulse_sign = get_bits(&gb, 6);
205  p->subframe[1].pulse_sign = get_bits(&gb, 5);
206  p->subframe[2].pulse_sign = get_bits(&gb, 6);
207  p->subframe[3].pulse_sign = get_bits(&gb, 5);
208  } else { /* 5300 bps */
209  p->subframe[0].pulse_pos = get_bits(&gb, 12);
210  p->subframe[1].pulse_pos = get_bits(&gb, 12);
211  p->subframe[2].pulse_pos = get_bits(&gb, 12);
212  p->subframe[3].pulse_pos = get_bits(&gb, 12);
213 
214  p->subframe[0].pulse_sign = get_bits(&gb, 4);
215  p->subframe[1].pulse_sign = get_bits(&gb, 4);
216  p->subframe[2].pulse_sign = get_bits(&gb, 4);
217  p->subframe[3].pulse_sign = get_bits(&gb, 4);
218  }
219 
220  return 0;
221 }
222 
223 /**
224  * Bitexact implementation of sqrt(val/2).
225  */
226 static int16_t square_root(unsigned val)
227 {
228  av_assert2(!(val & 0x80000000));
229 
230  return (ff_sqrt(val << 1) >> 1) & (~1);
231 }
232 
233 /**
234  * Calculate the number of left-shifts required for normalizing the input.
235  *
236  * @param num input number
237  * @param width width of the input, 15 or 31 bits
238  */
239 static int normalize_bits(int num, int width)
240 {
241  return width - av_log2(num) - 1;
242 }
243 
244 #define normalize_bits_int16(num) normalize_bits(num, 15)
245 #define normalize_bits_int32(num) normalize_bits(num, 31)
246 
247 /**
248  * Scale vector contents based on the largest of their absolutes.
249  */
250 static int scale_vector(int16_t *dst, const int16_t *vector, int length)
251 {
252  int bits, max = 0;
253  int i;
254 
255  for (i = 0; i < length; i++)
256  max |= FFABS(vector[i]);
257 
258  bits= 14 - av_log2_16bit(max);
259  bits= FFMAX(bits, 0);
260 
261  for (i = 0; i < length; i++)
262  dst[i] = vector[i] << bits >> 3;
263 
264  return bits - 3;
265 }
266 
267 /**
268  * Perform inverse quantization of LSP frequencies.
269  *
270  * @param cur_lsp the current LSP vector
271  * @param prev_lsp the previous LSP vector
272  * @param lsp_index VQ indices
273  * @param bad_frame bad frame flag
274  */
275 static void inverse_quant(int16_t *cur_lsp, int16_t *prev_lsp,
276  uint8_t *lsp_index, int bad_frame)
277 {
278  int min_dist, pred;
279  int i, j, temp, stable;
280 
281  /* Check for frame erasure */
282  if (!bad_frame) {
283  min_dist = 0x100;
284  pred = 12288;
285  } else {
286  min_dist = 0x200;
287  pred = 23552;
288  lsp_index[0] = lsp_index[1] = lsp_index[2] = 0;
289  }
290 
291  /* Get the VQ table entry corresponding to the transmitted index */
292  cur_lsp[0] = lsp_band0[lsp_index[0]][0];
293  cur_lsp[1] = lsp_band0[lsp_index[0]][1];
294  cur_lsp[2] = lsp_band0[lsp_index[0]][2];
295  cur_lsp[3] = lsp_band1[lsp_index[1]][0];
296  cur_lsp[4] = lsp_band1[lsp_index[1]][1];
297  cur_lsp[5] = lsp_band1[lsp_index[1]][2];
298  cur_lsp[6] = lsp_band2[lsp_index[2]][0];
299  cur_lsp[7] = lsp_band2[lsp_index[2]][1];
300  cur_lsp[8] = lsp_band2[lsp_index[2]][2];
301  cur_lsp[9] = lsp_band2[lsp_index[2]][3];
302 
303  /* Add predicted vector & DC component to the previously quantized vector */
304  for (i = 0; i < LPC_ORDER; i++) {
305  temp = ((prev_lsp[i] - dc_lsp[i]) * pred + (1 << 14)) >> 15;
306  cur_lsp[i] += dc_lsp[i] + temp;
307  }
308 
309  for (i = 0; i < LPC_ORDER; i++) {
310  cur_lsp[0] = FFMAX(cur_lsp[0], 0x180);
311  cur_lsp[LPC_ORDER - 1] = FFMIN(cur_lsp[LPC_ORDER - 1], 0x7e00);
312 
313  /* Stability check */
314  for (j = 1; j < LPC_ORDER; j++) {
315  temp = min_dist + cur_lsp[j - 1] - cur_lsp[j];
316  if (temp > 0) {
317  temp >>= 1;
318  cur_lsp[j - 1] -= temp;
319  cur_lsp[j] += temp;
320  }
321  }
322  stable = 1;
323  for (j = 1; j < LPC_ORDER; j++) {
324  temp = cur_lsp[j - 1] + min_dist - cur_lsp[j] - 4;
325  if (temp > 0) {
326  stable = 0;
327  break;
328  }
329  }
330  if (stable)
331  break;
332  }
333  if (!stable)
334  memcpy(cur_lsp, prev_lsp, LPC_ORDER * sizeof(*cur_lsp));
335 }
336 
337 /**
338  * Bitexact implementation of 2ab scaled by 1/2^16.
339  *
340  * @param a 32 bit multiplicand
341  * @param b 16 bit multiplier
342  */
343 #define MULL2(a, b) \
344  MULL(a,b,15)
345 
346 /**
347  * Convert LSP frequencies to LPC coefficients.
348  *
349  * @param lpc buffer for LPC coefficients
350  */
351 static void lsp2lpc(int16_t *lpc)
352 {
353  int f1[LPC_ORDER / 2 + 1];
354  int f2[LPC_ORDER / 2 + 1];
355  int i, j;
356 
357  /* Calculate negative cosine */
358  for (j = 0; j < LPC_ORDER; j++) {
359  int index = (lpc[j] >> 7) & 0x1FF;
360  int offset = lpc[j] & 0x7f;
361  int temp1 = cos_tab[index] << 16;
362  int temp2 = (cos_tab[index + 1] - cos_tab[index]) *
363  ((offset << 8) + 0x80) << 1;
364 
365  lpc[j] = -(av_sat_dadd32(1 << 15, temp1 + temp2) >> 16);
366  }
367 
368  /*
369  * Compute sum and difference polynomial coefficients
370  * (bitexact alternative to lsp2poly() in lsp.c)
371  */
372  /* Initialize with values in Q28 */
373  f1[0] = 1 << 28;
374  f1[1] = (lpc[0] << 14) + (lpc[2] << 14);
375  f1[2] = lpc[0] * lpc[2] + (2 << 28);
376 
377  f2[0] = 1 << 28;
378  f2[1] = (lpc[1] << 14) + (lpc[3] << 14);
379  f2[2] = lpc[1] * lpc[3] + (2 << 28);
380 
381  /*
382  * Calculate and scale the coefficients by 1/2 in
383  * each iteration for a final scaling factor of Q25
384  */
385  for (i = 2; i < LPC_ORDER / 2; i++) {
386  f1[i + 1] = f1[i - 1] + MULL2(f1[i], lpc[2 * i]);
387  f2[i + 1] = f2[i - 1] + MULL2(f2[i], lpc[2 * i + 1]);
388 
389  for (j = i; j >= 2; j--) {
390  f1[j] = MULL2(f1[j - 1], lpc[2 * i]) +
391  (f1[j] >> 1) + (f1[j - 2] >> 1);
392  f2[j] = MULL2(f2[j - 1], lpc[2 * i + 1]) +
393  (f2[j] >> 1) + (f2[j - 2] >> 1);
394  }
395 
396  f1[0] >>= 1;
397  f2[0] >>= 1;
398  f1[1] = ((lpc[2 * i] << 16 >> i) + f1[1]) >> 1;
399  f2[1] = ((lpc[2 * i + 1] << 16 >> i) + f2[1]) >> 1;
400  }
401 
402  /* Convert polynomial coefficients to LPC coefficients */
403  for (i = 0; i < LPC_ORDER / 2; i++) {
404  int64_t ff1 = f1[i + 1] + f1[i];
405  int64_t ff2 = f2[i + 1] - f2[i];
406 
407  lpc[i] = av_clipl_int32(((ff1 + ff2) << 3) + (1 << 15)) >> 16;
408  lpc[LPC_ORDER - i - 1] = av_clipl_int32(((ff1 - ff2) << 3) +
409  (1 << 15)) >> 16;
410  }
411 }
412 
413 /**
414  * Quantize LSP frequencies by interpolation and convert them to
415  * the corresponding LPC coefficients.
416  *
417  * @param lpc buffer for LPC coefficients
418  * @param cur_lsp the current LSP vector
419  * @param prev_lsp the previous LSP vector
420  */
421 static void lsp_interpolate(int16_t *lpc, int16_t *cur_lsp, int16_t *prev_lsp)
422 {
423  int i;
424  int16_t *lpc_ptr = lpc;
425 
426  /* cur_lsp * 0.25 + prev_lsp * 0.75 */
427  ff_acelp_weighted_vector_sum(lpc, cur_lsp, prev_lsp,
428  4096, 12288, 1 << 13, 14, LPC_ORDER);
429  ff_acelp_weighted_vector_sum(lpc + LPC_ORDER, cur_lsp, prev_lsp,
430  8192, 8192, 1 << 13, 14, LPC_ORDER);
431  ff_acelp_weighted_vector_sum(lpc + 2 * LPC_ORDER, cur_lsp, prev_lsp,
432  12288, 4096, 1 << 13, 14, LPC_ORDER);
433  memcpy(lpc + 3 * LPC_ORDER, cur_lsp, LPC_ORDER * sizeof(*lpc));
434 
435  for (i = 0; i < SUBFRAMES; i++) {
436  lsp2lpc(lpc_ptr);
437  lpc_ptr += LPC_ORDER;
438  }
439 }
440 
441 /**
442  * Generate a train of dirac functions with period as pitch lag.
443  */
444 static void gen_dirac_train(int16_t *buf, int pitch_lag)
445 {
446  int16_t vector[SUBFRAME_LEN];
447  int i, j;
448 
449  memcpy(vector, buf, SUBFRAME_LEN * sizeof(*vector));
450  for (i = pitch_lag; i < SUBFRAME_LEN; i += pitch_lag) {
451  for (j = 0; j < SUBFRAME_LEN - i; j++)
452  buf[i + j] += vector[j];
453  }
454 }
455 
456 /**
457  * Generate fixed codebook excitation vector.
458  *
459  * @param vector decoded excitation vector
460  * @param subfrm current subframe
461  * @param cur_rate current bitrate
462  * @param pitch_lag closed loop pitch lag
463  * @param index current subframe index
464  */
465 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe *subfrm,
466  enum Rate cur_rate, int pitch_lag, int index)
467 {
468  int temp, i, j;
469 
470  memset(vector, 0, SUBFRAME_LEN * sizeof(*vector));
471 
472  if (cur_rate == RATE_6300) {
473  if (subfrm->pulse_pos >= max_pos[index])
474  return;
475 
476  /* Decode amplitudes and positions */
477  j = PULSE_MAX - pulses[index];
478  temp = subfrm->pulse_pos;
479  for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
480  temp -= combinatorial_table[j][i];
481  if (temp >= 0)
482  continue;
483  temp += combinatorial_table[j++][i];
484  if (subfrm->pulse_sign & (1 << (PULSE_MAX - j))) {
485  vector[subfrm->grid_index + GRID_SIZE * i] =
486  -fixed_cb_gain[subfrm->amp_index];
487  } else {
488  vector[subfrm->grid_index + GRID_SIZE * i] =
489  fixed_cb_gain[subfrm->amp_index];
490  }
491  if (j == PULSE_MAX)
492  break;
493  }
494  if (subfrm->dirac_train == 1)
495  gen_dirac_train(vector, pitch_lag);
496  } else { /* 5300 bps */
497  int cb_gain = fixed_cb_gain[subfrm->amp_index];
498  int cb_shift = subfrm->grid_index;
499  int cb_sign = subfrm->pulse_sign;
500  int cb_pos = subfrm->pulse_pos;
501  int offset, beta, lag;
502 
503  for (i = 0; i < 8; i += 2) {
504  offset = ((cb_pos & 7) << 3) + cb_shift + i;
505  vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
506  cb_pos >>= 3;
507  cb_sign >>= 1;
508  }
509 
510  /* Enhance harmonic components */
511  lag = pitch_contrib[subfrm->ad_cb_gain << 1] + pitch_lag +
512  subfrm->ad_cb_lag - 1;
513  beta = pitch_contrib[(subfrm->ad_cb_gain << 1) + 1];
514 
515  if (lag < SUBFRAME_LEN - 2) {
516  for (i = lag; i < SUBFRAME_LEN; i++)
517  vector[i] += beta * vector[i - lag] >> 15;
518  }
519  }
520 }
521 
522 /**
523  * Get delayed contribution from the previous excitation vector.
524  */
525 static void get_residual(int16_t *residual, int16_t *prev_excitation, int lag)
526 {
527  int offset = PITCH_MAX - PITCH_ORDER / 2 - lag;
528  int i;
529 
530  residual[0] = prev_excitation[offset];
531  residual[1] = prev_excitation[offset + 1];
532 
533  offset += 2;
534  for (i = 2; i < SUBFRAME_LEN + PITCH_ORDER - 1; i++)
535  residual[i] = prev_excitation[offset + (i - 2) % lag];
536 }
537 
538 static int dot_product(const int16_t *a, const int16_t *b, int length)
539 {
540  int sum = ff_dot_product(a,b,length);
541  return av_sat_add32(sum, sum);
542 }
543 
544 /**
545  * Generate adaptive codebook excitation.
546  */
547 static void gen_acb_excitation(int16_t *vector, int16_t *prev_excitation,
548  int pitch_lag, G723_1_Subframe *subfrm,
549  enum Rate cur_rate)
550 {
551  int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
552  const int16_t *cb_ptr;
553  int lag = pitch_lag + subfrm->ad_cb_lag - 1;
554 
555  int i;
556  int sum;
557 
558  get_residual(residual, prev_excitation, lag);
559 
560  /* Select quantization table */
561  if (cur_rate == RATE_6300 && pitch_lag < SUBFRAME_LEN - 2) {
562  cb_ptr = adaptive_cb_gain85;
563  } else
564  cb_ptr = adaptive_cb_gain170;
565 
566  /* Calculate adaptive vector */
567  cb_ptr += subfrm->ad_cb_gain * 20;
568  for (i = 0; i < SUBFRAME_LEN; i++) {
569  sum = ff_dot_product(residual + i, cb_ptr, PITCH_ORDER);
570  vector[i] = av_sat_dadd32(1 << 15, av_sat_add32(sum, sum)) >> 16;
571  }
572 }
573 
574 /**
575  * Estimate maximum auto-correlation around pitch lag.
576  *
577  * @param buf buffer with offset applied
578  * @param offset offset of the excitation vector
579  * @param ccr_max pointer to the maximum auto-correlation
580  * @param pitch_lag decoded pitch lag
581  * @param length length of autocorrelation
582  * @param dir forward lag(1) / backward lag(-1)
583  */
584 static int autocorr_max(const int16_t *buf, int offset, int *ccr_max,
585  int pitch_lag, int length, int dir)
586 {
587  int limit, ccr, lag = 0;
588  int i;
589 
590  pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
591  if (dir > 0)
592  limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
593  else
594  limit = pitch_lag + 3;
595 
596  for (i = pitch_lag - 3; i <= limit; i++) {
597  ccr = dot_product(buf, buf + dir * i, length);
598 
599  if (ccr > *ccr_max) {
600  *ccr_max = ccr;
601  lag = i;
602  }
603  }
604  return lag;
605 }
606 
607 /**
608  * Calculate pitch postfilter optimal and scaling gains.
609  *
610  * @param lag pitch postfilter forward/backward lag
611  * @param ppf pitch postfilter parameters
612  * @param cur_rate current bitrate
613  * @param tgt_eng target energy
614  * @param ccr cross-correlation
615  * @param res_eng residual energy
616  */
617 static void comp_ppf_gains(int lag, PPFParam *ppf, enum Rate cur_rate,
618  int tgt_eng, int ccr, int res_eng)
619 {
620  int pf_residual; /* square of postfiltered residual */
621  int temp1, temp2;
622 
623  ppf->index = lag;
624 
625  temp1 = tgt_eng * res_eng >> 1;
626  temp2 = ccr * ccr << 1;
627 
628  if (temp2 > temp1) {
629  if (ccr >= res_eng) {
630  ppf->opt_gain = ppf_gain_weight[cur_rate];
631  } else {
632  ppf->opt_gain = (ccr << 15) / res_eng *
633  ppf_gain_weight[cur_rate] >> 15;
634  }
635  /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
636  temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
637  temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
638  pf_residual = av_sat_add32(temp1, temp2 + (1 << 15)) >> 16;
639 
640  if (tgt_eng >= pf_residual << 1) {
641  temp1 = 0x7fff;
642  } else {
643  temp1 = (tgt_eng << 14) / pf_residual;
644  }
645 
646  /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
647  ppf->sc_gain = square_root(temp1 << 16);
648  } else {
649  ppf->opt_gain = 0;
650  ppf->sc_gain = 0x7fff;
651  }
652 
653  ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
654 }
655 
656 /**
657  * Calculate pitch postfilter parameters.
658  *
659  * @param p the context
660  * @param offset offset of the excitation vector
661  * @param pitch_lag decoded pitch lag
662  * @param ppf pitch postfilter parameters
663  * @param cur_rate current bitrate
664  */
665 static void comp_ppf_coeff(G723_1_Context *p, int offset, int pitch_lag,
666  PPFParam *ppf, enum Rate cur_rate)
667 {
668 
669  int16_t scale;
670  int i;
671  int temp1, temp2;
672 
673  /*
674  * 0 - target energy
675  * 1 - forward cross-correlation
676  * 2 - forward residual energy
677  * 3 - backward cross-correlation
678  * 4 - backward residual energy
679  */
680  int energy[5] = {0, 0, 0, 0, 0};
681  int16_t *buf = p->audio + LPC_ORDER + offset;
682  int fwd_lag = autocorr_max(buf, offset, &energy[1], pitch_lag,
683  SUBFRAME_LEN, 1);
684  int back_lag = autocorr_max(buf, offset, &energy[3], pitch_lag,
685  SUBFRAME_LEN, -1);
686 
687  ppf->index = 0;
688  ppf->opt_gain = 0;
689  ppf->sc_gain = 0x7fff;
690 
691  /* Case 0, Section 3.6 */
692  if (!back_lag && !fwd_lag)
693  return;
694 
695  /* Compute target energy */
696  energy[0] = dot_product(buf, buf, SUBFRAME_LEN);
697 
698  /* Compute forward residual energy */
699  if (fwd_lag)
700  energy[2] = dot_product(buf + fwd_lag, buf + fwd_lag, SUBFRAME_LEN);
701 
702  /* Compute backward residual energy */
703  if (back_lag)
704  energy[4] = dot_product(buf - back_lag, buf - back_lag, SUBFRAME_LEN);
705 
706  /* Normalize and shorten */
707  temp1 = 0;
708  for (i = 0; i < 5; i++)
709  temp1 = FFMAX(energy[i], temp1);
710 
711  scale = normalize_bits(temp1, 31);
712  for (i = 0; i < 5; i++)
713  energy[i] = (energy[i] << scale) >> 16;
714 
715  if (fwd_lag && !back_lag) { /* Case 1 */
716  comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
717  energy[2]);
718  } else if (!fwd_lag) { /* Case 2 */
719  comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
720  energy[4]);
721  } else { /* Case 3 */
722 
723  /*
724  * Select the largest of energy[1]^2/energy[2]
725  * and energy[3]^2/energy[4]
726  */
727  temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
728  temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
729  if (temp1 >= temp2) {
730  comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
731  energy[2]);
732  } else {
733  comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
734  energy[4]);
735  }
736  }
737 }
738 
739 /**
740  * Classify frames as voiced/unvoiced.
741  *
742  * @param p the context
743  * @param pitch_lag decoded pitch_lag
744  * @param exc_eng excitation energy estimation
745  * @param scale scaling factor of exc_eng
746  *
747  * @return residual interpolation index if voiced, 0 otherwise
748  */
749 static int comp_interp_index(G723_1_Context *p, int pitch_lag,
750  int *exc_eng, int *scale)
751 {
752  int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
753  int16_t *buf = p->audio + LPC_ORDER;
754 
755  int index, ccr, tgt_eng, best_eng, temp;
756 
757  *scale = scale_vector(buf, p->excitation, FRAME_LEN + PITCH_MAX);
758  buf += offset;
759 
760  /* Compute maximum backward cross-correlation */
761  ccr = 0;
762  index = autocorr_max(buf, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
763  ccr = av_sat_add32(ccr, 1 << 15) >> 16;
764 
765  /* Compute target energy */
766  tgt_eng = dot_product(buf, buf, SUBFRAME_LEN * 2);
767  *exc_eng = av_sat_add32(tgt_eng, 1 << 15) >> 16;
768 
769  if (ccr <= 0)
770  return 0;
771 
772  /* Compute best energy */
773  best_eng = dot_product(buf - index, buf - index, SUBFRAME_LEN * 2);
774  best_eng = av_sat_add32(best_eng, 1 << 15) >> 16;
775 
776  temp = best_eng * *exc_eng >> 3;
777 
778  if (temp < ccr * ccr) {
779  return index;
780  } else
781  return 0;
782 }
783 
784 /**
785  * Peform residual interpolation based on frame classification.
786  *
787  * @param buf decoded excitation vector
788  * @param out output vector
789  * @param lag decoded pitch lag
790  * @param gain interpolated gain
791  * @param rseed seed for random number generator
792  */
793 static void residual_interp(int16_t *buf, int16_t *out, int lag,
794  int gain, int *rseed)
795 {
796  int i;
797  if (lag) { /* Voiced */
798  int16_t *vector_ptr = buf + PITCH_MAX;
799  /* Attenuate */
800  for (i = 0; i < lag; i++)
801  out[i] = vector_ptr[i - lag] * 3 >> 2;
802  av_memcpy_backptr((uint8_t*)(out + lag), lag * sizeof(*out),
803  (FRAME_LEN - lag) * sizeof(*out));
804  } else { /* Unvoiced */
805  for (i = 0; i < FRAME_LEN; i++) {
806  *rseed = *rseed * 521 + 259;
807  out[i] = gain * *rseed >> 15;
808  }
809  memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(*buf));
810  }
811 }
812 
813 /**
814  * Perform IIR filtering.
815  *
816  * @param fir_coef FIR coefficients
817  * @param iir_coef IIR coefficients
818  * @param src source vector
819  * @param dest destination vector
820  * @param width width of the output, 16 bits(0) / 32 bits(1)
821  */
822 #define iir_filter(fir_coef, iir_coef, src, dest, width)\
823 {\
824  int m, n;\
825  int res_shift = 16 & ~-(width);\
826  int in_shift = 16 - res_shift;\
827 \
828  for (m = 0; m < SUBFRAME_LEN; m++) {\
829  int64_t filter = 0;\
830  for (n = 1; n <= LPC_ORDER; n++) {\
831  filter -= (fir_coef)[n - 1] * (src)[m - n] -\
832  (iir_coef)[n - 1] * ((dest)[m - n] >> in_shift);\
833  }\
834 \
835  (dest)[m] = av_clipl_int32(((src)[m] << 16) + (filter << 3) +\
836  (1 << 15)) >> res_shift;\
837  }\
838 }
839 
840 /**
841  * Adjust gain of postfiltered signal.
842  *
843  * @param p the context
844  * @param buf postfiltered output vector
845  * @param energy input energy coefficient
846  */
847 static void gain_scale(G723_1_Context *p, int16_t * buf, int energy)
848 {
849  int num, denom, gain, bits1, bits2;
850  int i;
851 
852  num = energy;
853  denom = 0;
854  for (i = 0; i < SUBFRAME_LEN; i++) {
855  int temp = buf[i] >> 2;
856  temp *= temp;
857  denom = av_sat_dadd32(denom, temp);
858  }
859 
860  if (num && denom) {
861  bits1 = normalize_bits(num, 31);
862  bits2 = normalize_bits(denom, 31);
863  num = num << bits1 >> 1;
864  denom <<= bits2;
865 
866  bits2 = 5 + bits1 - bits2;
867  bits2 = FFMAX(0, bits2);
868 
869  gain = (num >> 1) / (denom >> 16);
870  gain = square_root(gain << 16 >> bits2);
871  } else {
872  gain = 1 << 12;
873  }
874 
875  for (i = 0; i < SUBFRAME_LEN; i++) {
876  p->pf_gain = (15 * p->pf_gain + gain + (1 << 3)) >> 4;
877  buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
878  (1 << 10)) >> 11);
879  }
880 }
881 
882 /**
883  * Perform formant filtering.
884  *
885  * @param p the context
886  * @param lpc quantized lpc coefficients
887  * @param buf input buffer
888  * @param dst output buffer
889  */
890 static void formant_postfilter(G723_1_Context *p, int16_t *lpc,
891  int16_t *buf, int16_t *dst)
892 {
893  int16_t filter_coef[2][LPC_ORDER];
894  int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
895  int i, j, k;
896 
897  memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(*buf));
898  memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(*filter_signal));
899 
900  for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
901  for (k = 0; k < LPC_ORDER; k++) {
902  filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
903  (1 << 14)) >> 15;
904  filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
905  (1 << 14)) >> 15;
906  }
907  iir_filter(filter_coef[0], filter_coef[1], buf + i,
908  filter_signal + i, 1);
909  lpc += LPC_ORDER;
910  }
911 
912  memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(int16_t));
913  memcpy(p->iir_mem, filter_signal + FRAME_LEN, LPC_ORDER * sizeof(int));
914 
915  buf += LPC_ORDER;
916  signal_ptr = filter_signal + LPC_ORDER;
917  for (i = 0; i < SUBFRAMES; i++) {
918  int temp;
919  int auto_corr[2];
920  int scale, energy;
921 
922  /* Normalize */
923  scale = scale_vector(dst, buf, SUBFRAME_LEN);
924 
925  /* Compute auto correlation coefficients */
926  auto_corr[0] = dot_product(dst, dst + 1, SUBFRAME_LEN - 1);
927  auto_corr[1] = dot_product(dst, dst, SUBFRAME_LEN);
928 
929  /* Compute reflection coefficient */
930  temp = auto_corr[1] >> 16;
931  if (temp) {
932  temp = (auto_corr[0] >> 2) / temp;
933  }
934  p->reflection_coef = (3 * p->reflection_coef + temp + 2) >> 2;
935  temp = -p->reflection_coef >> 1 & ~3;
936 
937  /* Compensation filter */
938  for (j = 0; j < SUBFRAME_LEN; j++) {
939  dst[j] = av_sat_dadd32(signal_ptr[j],
940  (signal_ptr[j - 1] >> 16) * temp) >> 16;
941  }
942 
943  /* Compute normalized signal energy */
944  temp = 2 * scale + 4;
945  if (temp < 0) {
946  energy = av_clipl_int32((int64_t)auto_corr[1] << -temp);
947  } else
948  energy = auto_corr[1] >> temp;
949 
950  gain_scale(p, dst, energy);
951 
952  buf += SUBFRAME_LEN;
953  signal_ptr += SUBFRAME_LEN;
954  dst += SUBFRAME_LEN;
955  }
956 }
957 
958 static int sid_gain_to_lsp_index(int gain)
959 {
960  if (gain < 0x10)
961  return gain << 6;
962  else if (gain < 0x20)
963  return gain - 8 << 7;
964  else
965  return gain - 20 << 8;
966 }
967 
968 static inline int cng_rand(int *state, int base)
969 {
970  *state = (*state * 521 + 259) & 0xFFFF;
971  return (*state & 0x7FFF) * base >> 15;
972 }
973 
974 static int estimate_sid_gain(G723_1_Context *p)
975 {
976  int i, shift, seg, seg2, t, val, val_add, x, y;
977 
978  shift = 16 - p->cur_gain * 2;
979  if (shift > 0)
980  t = p->sid_gain << shift;
981  else
982  t = p->sid_gain >> -shift;
983  x = t * cng_filt[0] >> 16;
984 
985  if (x >= cng_bseg[2])
986  return 0x3F;
987 
988  if (x >= cng_bseg[1]) {
989  shift = 4;
990  seg = 3;
991  } else {
992  shift = 3;
993  seg = (x >= cng_bseg[0]);
994  }
995  seg2 = FFMIN(seg, 3);
996 
997  val = 1 << shift;
998  val_add = val >> 1;
999  for (i = 0; i < shift; i++) {
1000  t = seg * 32 + (val << seg2);
1001  t *= t;
1002  if (x >= t)
1003  val += val_add;
1004  else
1005  val -= val_add;
1006  val_add >>= 1;
1007  }
1008 
1009  t = seg * 32 + (val << seg2);
1010  y = t * t - x;
1011  if (y <= 0) {
1012  t = seg * 32 + (val + 1 << seg2);
1013  t = t * t - x;
1014  val = (seg2 - 1 << 4) + val;
1015  if (t >= y)
1016  val++;
1017  } else {
1018  t = seg * 32 + (val - 1 << seg2);
1019  t = t * t - x;
1020  val = (seg2 - 1 << 4) + val;
1021  if (t >= y)
1022  val--;
1023  }
1024 
1025  return val;
1026 }
1027 
1028 static void generate_noise(G723_1_Context *p)
1029 {
1030  int i, j, idx, t;
1031  int off[SUBFRAMES];
1032  int signs[SUBFRAMES / 2 * 11], pos[SUBFRAMES / 2 * 11];
1033  int tmp[SUBFRAME_LEN * 2];
1034  int16_t *vector_ptr;
1035  int64_t sum;
1036  int b0, c, delta, x, shift;
1037 
1038  p->pitch_lag[0] = cng_rand(&p->cng_random_seed, 21) + 123;
1039  p->pitch_lag[1] = cng_rand(&p->cng_random_seed, 19) + 123;
1040 
1041  for (i = 0; i < SUBFRAMES; i++) {
1042  p->subframe[i].ad_cb_gain = cng_rand(&p->cng_random_seed, 50) + 1;
1043  p->subframe[i].ad_cb_lag = cng_adaptive_cb_lag[i];
1044  }
1045 
1046  for (i = 0; i < SUBFRAMES / 2; i++) {
1047  t = cng_rand(&p->cng_random_seed, 1 << 13);
1048  off[i * 2] = t & 1;
1049  off[i * 2 + 1] = ((t >> 1) & 1) + SUBFRAME_LEN;
1050  t >>= 2;
1051  for (j = 0; j < 11; j++) {
1052  signs[i * 11 + j] = (t & 1) * 2 - 1 << 14;
1053  t >>= 1;
1054  }
1055  }
1056 
1057  idx = 0;
1058  for (i = 0; i < SUBFRAMES; i++) {
1059  for (j = 0; j < SUBFRAME_LEN / 2; j++)
1060  tmp[j] = j;
1061  t = SUBFRAME_LEN / 2;
1062  for (j = 0; j < pulses[i]; j++, idx++) {
1063  int idx2 = cng_rand(&p->cng_random_seed, t);
1064 
1065  pos[idx] = tmp[idx2] * 2 + off[i];
1066  tmp[idx2] = tmp[--t];
1067  }
1068  }
1069 
1070  vector_ptr = p->audio + LPC_ORDER;
1071  memcpy(vector_ptr, p->prev_excitation,
1072  PITCH_MAX * sizeof(*p->excitation));
1073  for (i = 0; i < SUBFRAMES; i += 2) {
1074  gen_acb_excitation(vector_ptr, vector_ptr,
1075  p->pitch_lag[i >> 1], &p->subframe[i],
1076  p->cur_rate);
1077  gen_acb_excitation(vector_ptr + SUBFRAME_LEN,
1078  vector_ptr + SUBFRAME_LEN,
1079  p->pitch_lag[i >> 1], &p->subframe[i + 1],
1080  p->cur_rate);
1081 
1082  t = 0;
1083  for (j = 0; j < SUBFRAME_LEN * 2; j++)
1084  t |= FFABS(vector_ptr[j]);
1085  t = FFMIN(t, 0x7FFF);
1086  if (!t) {
1087  shift = 0;
1088  } else {
1089  shift = -10 + av_log2(t);
1090  if (shift < -2)
1091  shift = -2;
1092  }
1093  sum = 0;
1094  if (shift < 0) {
1095  for (j = 0; j < SUBFRAME_LEN * 2; j++) {
1096  t = vector_ptr[j] << -shift;
1097  sum += t * t;
1098  tmp[j] = t;
1099  }
1100  } else {
1101  for (j = 0; j < SUBFRAME_LEN * 2; j++) {
1102  t = vector_ptr[j] >> shift;
1103  sum += t * t;
1104  tmp[j] = t;
1105  }
1106  }
1107 
1108  b0 = 0;
1109  for (j = 0; j < 11; j++)
1110  b0 += tmp[pos[(i / 2) * 11 + j]] * signs[(i / 2) * 11 + j];
1111  b0 = b0 * 2 * 2979LL + (1 << 29) >> 30; // approximated division by 11
1112 
1113  c = p->cur_gain * (p->cur_gain * SUBFRAME_LEN >> 5);
1114  if (shift * 2 + 3 >= 0)
1115  c >>= shift * 2 + 3;
1116  else
1117  c <<= -(shift * 2 + 3);
1118  c = (av_clipl_int32(sum << 1) - c) * 2979LL >> 15;
1119 
1120  delta = b0 * b0 * 2 - c;
1121  if (delta <= 0) {
1122  x = -b0;
1123  } else {
1124  delta = square_root(delta);
1125  x = delta - b0;
1126  t = delta + b0;
1127  if (FFABS(t) < FFABS(x))
1128  x = -t;
1129  }
1130  shift++;
1131  if (shift < 0)
1132  x >>= -shift;
1133  else
1134  x <<= shift;
1135  x = av_clip(x, -10000, 10000);
1136 
1137  for (j = 0; j < 11; j++) {
1138  idx = (i / 2) * 11 + j;
1139  vector_ptr[pos[idx]] = av_clip_int16(vector_ptr[pos[idx]] +
1140  (x * signs[idx] >> 15));
1141  }
1142 
1143  /* copy decoded data to serve as a history for the next decoded subframes */
1144  memcpy(vector_ptr + PITCH_MAX, vector_ptr,
1145  sizeof(*vector_ptr) * SUBFRAME_LEN * 2);
1146  vector_ptr += SUBFRAME_LEN * 2;
1147  }
1148  /* Save the excitation for the next frame */
1149  memcpy(p->prev_excitation, p->audio + LPC_ORDER + FRAME_LEN,
1150  PITCH_MAX * sizeof(*p->excitation));
1151 }
1152 
1153 static int g723_1_decode_frame(AVCodecContext *avctx, void *data,
1154  int *got_frame_ptr, AVPacket *avpkt)
1155 {
1156  G723_1_Context *p = avctx->priv_data;
1157  AVFrame *frame = data;
1158  const uint8_t *buf = avpkt->data;
1159  int buf_size = avpkt->size;
1160  int dec_mode = buf[0] & 3;
1161 
1162  PPFParam ppf[SUBFRAMES];
1163  int16_t cur_lsp[LPC_ORDER];
1164  int16_t lpc[SUBFRAMES * LPC_ORDER];
1165  int16_t acb_vector[SUBFRAME_LEN];
1166  int16_t *out;
1167  int bad_frame = 0, i, j, ret;
1168  int16_t *audio = p->audio;
1169 
1170  if (buf_size < frame_size[dec_mode]) {
1171  if (buf_size)
1172  av_log(avctx, AV_LOG_WARNING,
1173  "Expected %d bytes, got %d - skipping packet\n",
1174  frame_size[dec_mode], buf_size);
1175  *got_frame_ptr = 0;
1176  return buf_size;
1177  }
1178 
1179  if (unpack_bitstream(p, buf, buf_size) < 0) {
1180  bad_frame = 1;
1181  if (p->past_frame_type == ACTIVE_FRAME)
1182  p->cur_frame_type = ACTIVE_FRAME;
1183  else
1184  p->cur_frame_type = UNTRANSMITTED_FRAME;
1185  }
1186 
1187  frame->nb_samples = FRAME_LEN;
1188  if ((ret = ff_get_buffer(avctx, frame)) < 0) {
1189  av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1190  return ret;
1191  }
1192 
1193  out = (int16_t *)frame->data[0];
1194 
1195  if (p->cur_frame_type == ACTIVE_FRAME) {
1196  if (!bad_frame)
1197  p->erased_frames = 0;
1198  else if (p->erased_frames != 3)
1199  p->erased_frames++;
1200 
1201  inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
1202  lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
1203 
1204  /* Save the lsp_vector for the next frame */
1205  memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1206 
1207  /* Generate the excitation for the frame */
1208  memcpy(p->excitation, p->prev_excitation,
1209  PITCH_MAX * sizeof(*p->excitation));
1210  if (!p->erased_frames) {
1211  int16_t *vector_ptr = p->excitation + PITCH_MAX;
1212 
1213  /* Update interpolation gain memory */
1214  p->interp_gain = fixed_cb_gain[(p->subframe[2].amp_index +
1215  p->subframe[3].amp_index) >> 1];
1216  for (i = 0; i < SUBFRAMES; i++) {
1217  gen_fcb_excitation(vector_ptr, &p->subframe[i], p->cur_rate,
1218  p->pitch_lag[i >> 1], i);
1219  gen_acb_excitation(acb_vector, &p->excitation[SUBFRAME_LEN * i],
1220  p->pitch_lag[i >> 1], &p->subframe[i],
1221  p->cur_rate);
1222  /* Get the total excitation */
1223  for (j = 0; j < SUBFRAME_LEN; j++) {
1224  int v = av_clip_int16(vector_ptr[j] << 1);
1225  vector_ptr[j] = av_clip_int16(v + acb_vector[j]);
1226  }
1227  vector_ptr += SUBFRAME_LEN;
1228  }
1229 
1230  vector_ptr = p->excitation + PITCH_MAX;
1231 
1232  p->interp_index = comp_interp_index(p, p->pitch_lag[1],
1233  &p->sid_gain, &p->cur_gain);
1234 
1235  /* Peform pitch postfiltering */
1236  if (p->postfilter) {
1237  i = PITCH_MAX;
1238  for (j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1239  comp_ppf_coeff(p, i, p->pitch_lag[j >> 1],
1240  ppf + j, p->cur_rate);
1241 
1242  for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1243  ff_acelp_weighted_vector_sum(p->audio + LPC_ORDER + i,
1244  vector_ptr + i,
1245  vector_ptr + i + ppf[j].index,
1246  ppf[j].sc_gain,
1247  ppf[j].opt_gain,
1248  1 << 14, 15, SUBFRAME_LEN);
1249  } else {
1250  audio = vector_ptr - LPC_ORDER;
1251  }
1252 
1253  /* Save the excitation for the next frame */
1254  memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
1255  PITCH_MAX * sizeof(*p->excitation));
1256  } else {
1257  p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
1258  if (p->erased_frames == 3) {
1259  /* Mute output */
1260  memset(p->excitation, 0,
1261  (FRAME_LEN + PITCH_MAX) * sizeof(*p->excitation));
1262  memset(p->prev_excitation, 0,
1263  PITCH_MAX * sizeof(*p->excitation));
1264  memset(frame->data[0], 0,
1265  (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
1266  } else {
1267  int16_t *buf = p->audio + LPC_ORDER;
1268 
1269  /* Regenerate frame */
1270  residual_interp(p->excitation, buf, p->interp_index,
1271  p->interp_gain, &p->random_seed);
1272 
1273  /* Save the excitation for the next frame */
1274  memcpy(p->prev_excitation, buf + (FRAME_LEN - PITCH_MAX),
1275  PITCH_MAX * sizeof(*p->excitation));
1276  }
1277  }
1278  p->cng_random_seed = CNG_RANDOM_SEED;
1279  } else {
1280  if (p->cur_frame_type == SID_FRAME) {
1281  p->sid_gain = sid_gain_to_lsp_index(p->subframe[0].amp_index);
1282  inverse_quant(p->sid_lsp, p->prev_lsp, p->lsp_index, 0);
1283  } else if (p->past_frame_type == ACTIVE_FRAME) {
1284  p->sid_gain = estimate_sid_gain(p);
1285  }
1286 
1287  if (p->past_frame_type == ACTIVE_FRAME)
1288  p->cur_gain = p->sid_gain;
1289  else
1290  p->cur_gain = (p->cur_gain * 7 + p->sid_gain) >> 3;
1291  generate_noise(p);
1292  lsp_interpolate(lpc, p->sid_lsp, p->prev_lsp);
1293  /* Save the lsp_vector for the next frame */
1294  memcpy(p->prev_lsp, p->sid_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1295  }
1296 
1297  p->past_frame_type = p->cur_frame_type;
1298 
1299  memcpy(p->audio, p->synth_mem, LPC_ORDER * sizeof(*p->audio));
1300  for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1301  ff_celp_lp_synthesis_filter(p->audio + i, &lpc[j * LPC_ORDER],
1302  audio + i, SUBFRAME_LEN, LPC_ORDER,
1303  0, 1, 1 << 12);
1304  memcpy(p->synth_mem, p->audio + FRAME_LEN, LPC_ORDER * sizeof(*p->audio));
1305 
1306  if (p->postfilter) {
1307  formant_postfilter(p, lpc, p->audio, out);
1308  } else { // if output is not postfiltered it should be scaled by 2
1309  for (i = 0; i < FRAME_LEN; i++)
1310  out[i] = av_clip_int16(p->audio[LPC_ORDER + i] << 1);
1311  }
1312 
1313  *got_frame_ptr = 1;
1314 
1315  return frame_size[dec_mode];
1316 }
1317 
1318 #define OFFSET(x) offsetof(G723_1_Context, x)
1319 #define AD AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1320 
1321 static const AVOption options[] = {
1322  { "postfilter", "postfilter on/off", OFFSET(postfilter), AV_OPT_TYPE_INT,
1323  { .i64 = 1 }, 0, 1, AD },
1324  { NULL }
1325 };
1326 
1327 
1328 static const AVClass g723_1dec_class = {
1329  .class_name = "G.723.1 decoder",
1330  .item_name = av_default_item_name,
1331  .option = options,
1332  .version = LIBAVUTIL_VERSION_INT,
1333 };
1334 
1335 AVCodec ff_g723_1_decoder = {
1336  .name = "g723_1",
1337  .type = AVMEDIA_TYPE_AUDIO,
1338  .id = AV_CODEC_ID_G723_1,
1339  .priv_data_size = sizeof(G723_1_Context),
1340  .init = g723_1_decode_init,
1341  .decode = g723_1_decode_frame,
1342  .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1343  .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
1344  .priv_class = &g723_1dec_class,
1345 };
1346 
1347 #if CONFIG_G723_1_ENCODER
1348 #define BITSTREAM_WRITER_LE
1349 #include "put_bits.h"
1350 
1351 static av_cold int g723_1_encode_init(AVCodecContext *avctx)
1352 {
1353  G723_1_Context *p = avctx->priv_data;
1354 
1355  if (avctx->sample_rate != 8000) {
1356  av_log(avctx, AV_LOG_ERROR, "Only 8000Hz sample rate supported\n");
1357  return -1;
1358  }
1359 
1360  if (avctx->channels != 1) {
1361  av_log(avctx, AV_LOG_ERROR, "Only mono supported\n");
1362  return AVERROR(EINVAL);
1363  }
1364 
1365  if (avctx->bit_rate == 6300) {
1366  p->cur_rate = RATE_6300;
1367  } else if (avctx->bit_rate == 5300) {
1368  av_log(avctx, AV_LOG_ERROR, "Bitrate not supported yet, use 6.3k\n");
1369  return AVERROR_PATCHWELCOME;
1370  } else {
1371  av_log(avctx, AV_LOG_ERROR,
1372  "Bitrate not supported, use 6.3k\n");
1373  return AVERROR(EINVAL);
1374  }
1375  avctx->frame_size = 240;
1376  memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(int16_t));
1377 
1378  return 0;
1379 }
1380 
1381 /**
1382  * Remove DC component from the input signal.
1383  *
1384  * @param buf input signal
1385  * @param fir zero memory
1386  * @param iir pole memory
1387  */
1388 static void highpass_filter(int16_t *buf, int16_t *fir, int *iir)
1389 {
1390  int i;
1391  for (i = 0; i < FRAME_LEN; i++) {
1392  *iir = (buf[i] << 15) + ((-*fir) << 15) + MULL2(*iir, 0x7f00);
1393  *fir = buf[i];
1394  buf[i] = av_clipl_int32((int64_t)*iir + (1 << 15)) >> 16;
1395  }
1396 }
1397 
1398 /**
1399  * Estimate autocorrelation of the input vector.
1400  *
1401  * @param buf input buffer
1402  * @param autocorr autocorrelation coefficients vector
1403  */
1404 static void comp_autocorr(int16_t *buf, int16_t *autocorr)
1405 {
1406  int i, scale, temp;
1407  int16_t vector[LPC_FRAME];
1408 
1409  scale_vector(vector, buf, LPC_FRAME);
1410 
1411  /* Apply the Hamming window */
1412  for (i = 0; i < LPC_FRAME; i++)
1413  vector[i] = (vector[i] * hamming_window[i] + (1 << 14)) >> 15;
1414 
1415  /* Compute the first autocorrelation coefficient */
1416  temp = ff_dot_product(vector, vector, LPC_FRAME);
1417 
1418  /* Apply a white noise correlation factor of (1025/1024) */
1419  temp += temp >> 10;
1420 
1421  /* Normalize */
1422  scale = normalize_bits_int32(temp);
1423  autocorr[0] = av_clipl_int32((int64_t)(temp << scale) +
1424  (1 << 15)) >> 16;
1425 
1426  /* Compute the remaining coefficients */
1427  if (!autocorr[0]) {
1428  memset(autocorr + 1, 0, LPC_ORDER * sizeof(int16_t));
1429  } else {
1430  for (i = 1; i <= LPC_ORDER; i++) {
1431  temp = ff_dot_product(vector, vector + i, LPC_FRAME - i);
1432  temp = MULL2((temp << scale), binomial_window[i - 1]);
1433  autocorr[i] = av_clipl_int32((int64_t)temp + (1 << 15)) >> 16;
1434  }
1435  }
1436 }
1437 
1438 /**
1439  * Use Levinson-Durbin recursion to compute LPC coefficients from
1440  * autocorrelation values.
1441  *
1442  * @param lpc LPC coefficients vector
1443  * @param autocorr autocorrelation coefficients vector
1444  * @param error prediction error
1445  */
1446 static void levinson_durbin(int16_t *lpc, int16_t *autocorr, int16_t error)
1447 {
1448  int16_t vector[LPC_ORDER];
1449  int16_t partial_corr;
1450  int i, j, temp;
1451 
1452  memset(lpc, 0, LPC_ORDER * sizeof(int16_t));
1453 
1454  for (i = 0; i < LPC_ORDER; i++) {
1455  /* Compute the partial correlation coefficient */
1456  temp = 0;
1457  for (j = 0; j < i; j++)
1458  temp -= lpc[j] * autocorr[i - j - 1];
1459  temp = ((autocorr[i] << 13) + temp) << 3;
1460 
1461  if (FFABS(temp) >= (error << 16))
1462  break;
1463 
1464  partial_corr = temp / (error << 1);
1465 
1466  lpc[i] = av_clipl_int32((int64_t)(partial_corr << 14) +
1467  (1 << 15)) >> 16;
1468 
1469  /* Update the prediction error */
1470  temp = MULL2(temp, partial_corr);
1471  error = av_clipl_int32((int64_t)(error << 16) - temp +
1472  (1 << 15)) >> 16;
1473 
1474  memcpy(vector, lpc, i * sizeof(int16_t));
1475  for (j = 0; j < i; j++) {
1476  temp = partial_corr * vector[i - j - 1] << 1;
1477  lpc[j] = av_clipl_int32((int64_t)(lpc[j] << 16) - temp +
1478  (1 << 15)) >> 16;
1479  }
1480  }
1481 }
1482 
1483 /**
1484  * Calculate LPC coefficients for the current frame.
1485  *
1486  * @param buf current frame
1487  * @param prev_data 2 trailing subframes of the previous frame
1488  * @param lpc LPC coefficients vector
1489  */
1490 static void comp_lpc_coeff(int16_t *buf, int16_t *lpc)
1491 {
1492  int16_t autocorr[(LPC_ORDER + 1) * SUBFRAMES];
1493  int16_t *autocorr_ptr = autocorr;
1494  int16_t *lpc_ptr = lpc;
1495  int i, j;
1496 
1497  for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1498  comp_autocorr(buf + i, autocorr_ptr);
1499  levinson_durbin(lpc_ptr, autocorr_ptr + 1, autocorr_ptr[0]);
1500 
1501  lpc_ptr += LPC_ORDER;
1502  autocorr_ptr += LPC_ORDER + 1;
1503  }
1504 }
1505 
1506 static void lpc2lsp(int16_t *lpc, int16_t *prev_lsp, int16_t *lsp)
1507 {
1508  int f[LPC_ORDER + 2]; ///< coefficients of the sum and difference
1509  ///< polynomials (F1, F2) ordered as
1510  ///< f1[0], f2[0], ...., f1[5], f2[5]
1511 
1512  int max, shift, cur_val, prev_val, count, p;
1513  int i, j;
1514  int64_t temp;
1515 
1516  /* Initialize f1[0] and f2[0] to 1 in Q25 */
1517  for (i = 0; i < LPC_ORDER; i++)
1518  lsp[i] = (lpc[i] * bandwidth_expand[i] + (1 << 14)) >> 15;
1519 
1520  /* Apply bandwidth expansion on the LPC coefficients */
1521  f[0] = f[1] = 1 << 25;
1522 
1523  /* Compute the remaining coefficients */
1524  for (i = 0; i < LPC_ORDER / 2; i++) {
1525  /* f1 */
1526  f[2 * i + 2] = -f[2 * i] - ((lsp[i] + lsp[LPC_ORDER - 1 - i]) << 12);
1527  /* f2 */
1528  f[2 * i + 3] = f[2 * i + 1] - ((lsp[i] - lsp[LPC_ORDER - 1 - i]) << 12);
1529  }
1530 
1531  /* Divide f1[5] and f2[5] by 2 for use in polynomial evaluation */
1532  f[LPC_ORDER] >>= 1;
1533  f[LPC_ORDER + 1] >>= 1;
1534 
1535  /* Normalize and shorten */
1536  max = FFABS(f[0]);
1537  for (i = 1; i < LPC_ORDER + 2; i++)
1538  max = FFMAX(max, FFABS(f[i]));
1539 
1540  shift = normalize_bits_int32(max);
1541 
1542  for (i = 0; i < LPC_ORDER + 2; i++)
1543  f[i] = av_clipl_int32((int64_t)(f[i] << shift) + (1 << 15)) >> 16;
1544 
1545  /**
1546  * Evaluate F1 and F2 at uniform intervals of pi/256 along the
1547  * unit circle and check for zero crossings.
1548  */
1549  p = 0;
1550  temp = 0;
1551  for (i = 0; i <= LPC_ORDER / 2; i++)
1552  temp += f[2 * i] * cos_tab[0];
1553  prev_val = av_clipl_int32(temp << 1);
1554  count = 0;
1555  for ( i = 1; i < COS_TBL_SIZE / 2; i++) {
1556  /* Evaluate */
1557  temp = 0;
1558  for (j = 0; j <= LPC_ORDER / 2; j++)
1559  temp += f[LPC_ORDER - 2 * j + p] * cos_tab[i * j % COS_TBL_SIZE];
1560  cur_val = av_clipl_int32(temp << 1);
1561 
1562  /* Check for sign change, indicating a zero crossing */
1563  if ((cur_val ^ prev_val) < 0) {
1564  int abs_cur = FFABS(cur_val);
1565  int abs_prev = FFABS(prev_val);
1566  int sum = abs_cur + abs_prev;
1567 
1568  shift = normalize_bits_int32(sum);
1569  sum <<= shift;
1570  abs_prev = abs_prev << shift >> 8;
1571  lsp[count++] = ((i - 1) << 7) + (abs_prev >> 1) / (sum >> 16);
1572 
1573  if (count == LPC_ORDER)
1574  break;
1575 
1576  /* Switch between sum and difference polynomials */
1577  p ^= 1;
1578 
1579  /* Evaluate */
1580  temp = 0;
1581  for (j = 0; j <= LPC_ORDER / 2; j++){
1582  temp += f[LPC_ORDER - 2 * j + p] *
1583  cos_tab[i * j % COS_TBL_SIZE];
1584  }
1585  cur_val = av_clipl_int32(temp<<1);
1586  }
1587  prev_val = cur_val;
1588  }
1589 
1590  if (count != LPC_ORDER)
1591  memcpy(lsp, prev_lsp, LPC_ORDER * sizeof(int16_t));
1592 }
1593 
1594 /**
1595  * Quantize the current LSP subvector.
1596  *
1597  * @param num band number
1598  * @param offset offset of the current subvector in an LPC_ORDER vector
1599  * @param size size of the current subvector
1600  */
1601 #define get_index(num, offset, size) \
1602 {\
1603  int error, max = -1;\
1604  int16_t temp[4];\
1605  int i, j;\
1606  for (i = 0; i < LSP_CB_SIZE; i++) {\
1607  for (j = 0; j < size; j++){\
1608  temp[j] = (weight[j + (offset)] * lsp_band##num[i][j] +\
1609  (1 << 14)) >> 15;\
1610  }\
1611  error = dot_product(lsp + (offset), temp, size) << 1;\
1612  error -= dot_product(lsp_band##num[i], temp, size);\
1613  if (error > max) {\
1614  max = error;\
1615  lsp_index[num] = i;\
1616  }\
1617  }\
1618 }
1619 
1620 /**
1621  * Vector quantize the LSP frequencies.
1622  *
1623  * @param lsp the current lsp vector
1624  * @param prev_lsp the previous lsp vector
1625  */
1626 static void lsp_quantize(uint8_t *lsp_index, int16_t *lsp, int16_t *prev_lsp)
1627 {
1628  int16_t weight[LPC_ORDER];
1629  int16_t min, max;
1630  int shift, i;
1631 
1632  /* Calculate the VQ weighting vector */
1633  weight[0] = (1 << 20) / (lsp[1] - lsp[0]);
1634  weight[LPC_ORDER - 1] = (1 << 20) /
1635  (lsp[LPC_ORDER - 1] - lsp[LPC_ORDER - 2]);
1636 
1637  for (i = 1; i < LPC_ORDER - 1; i++) {
1638  min = FFMIN(lsp[i] - lsp[i - 1], lsp[i + 1] - lsp[i]);
1639  if (min > 0x20)
1640  weight[i] = (1 << 20) / min;
1641  else
1642  weight[i] = INT16_MAX;
1643  }
1644 
1645  /* Normalize */
1646  max = 0;
1647  for (i = 0; i < LPC_ORDER; i++)
1648  max = FFMAX(weight[i], max);
1649 
1650  shift = normalize_bits_int16(max);
1651  for (i = 0; i < LPC_ORDER; i++) {
1652  weight[i] <<= shift;
1653  }
1654 
1655  /* Compute the VQ target vector */
1656  for (i = 0; i < LPC_ORDER; i++) {
1657  lsp[i] -= dc_lsp[i] +
1658  (((prev_lsp[i] - dc_lsp[i]) * 12288 + (1 << 14)) >> 15);
1659  }
1660 
1661  get_index(0, 0, 3);
1662  get_index(1, 3, 3);
1663  get_index(2, 6, 4);
1664 }
1665 
1666 /**
1667  * Apply the formant perceptual weighting filter.
1668  *
1669  * @param flt_coef filter coefficients
1670  * @param unq_lpc unquantized lpc vector
1671  */
1672 static void perceptual_filter(G723_1_Context *p, int16_t *flt_coef,
1673  int16_t *unq_lpc, int16_t *buf)
1674 {
1675  int16_t vector[FRAME_LEN + LPC_ORDER];
1676  int i, j, k, l = 0;
1677 
1678  memcpy(buf, p->iir_mem, sizeof(int16_t) * LPC_ORDER);
1679  memcpy(vector, p->fir_mem, sizeof(int16_t) * LPC_ORDER);
1680  memcpy(vector + LPC_ORDER, buf + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
1681 
1682  for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1683  for (k = 0; k < LPC_ORDER; k++) {
1684  flt_coef[k + 2 * l] = (unq_lpc[k + l] * percept_flt_tbl[0][k] +
1685  (1 << 14)) >> 15;
1686  flt_coef[k + 2 * l + LPC_ORDER] = (unq_lpc[k + l] *
1687  percept_flt_tbl[1][k] +
1688  (1 << 14)) >> 15;
1689  }
1690  iir_filter(flt_coef + 2 * l, flt_coef + 2 * l + LPC_ORDER, vector + i,
1691  buf + i, 0);
1692  l += LPC_ORDER;
1693  }
1694  memcpy(p->iir_mem, buf + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1695  memcpy(p->fir_mem, vector + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1696 }
1697 
1698 /**
1699  * Estimate the open loop pitch period.
1700  *
1701  * @param buf perceptually weighted speech
1702  * @param start estimation is carried out from this position
1703  */
1704 static int estimate_pitch(int16_t *buf, int start)
1705 {
1706  int max_exp = 32;
1707  int max_ccr = 0x4000;
1708  int max_eng = 0x7fff;
1709  int index = PITCH_MIN;
1710  int offset = start - PITCH_MIN + 1;
1711 
1712  int ccr, eng, orig_eng, ccr_eng, exp;
1713  int diff, temp;
1714 
1715  int i;
1716 
1717  orig_eng = ff_dot_product(buf + offset, buf + offset, HALF_FRAME_LEN);
1718 
1719  for (i = PITCH_MIN; i <= PITCH_MAX - 3; i++) {
1720  offset--;
1721 
1722  /* Update energy and compute correlation */
1723  orig_eng += buf[offset] * buf[offset] -
1724  buf[offset + HALF_FRAME_LEN] * buf[offset + HALF_FRAME_LEN];
1725  ccr = ff_dot_product(buf + start, buf + offset, HALF_FRAME_LEN);
1726  if (ccr <= 0)
1727  continue;
1728 
1729  /* Split into mantissa and exponent to maintain precision */
1730  exp = normalize_bits_int32(ccr);
1731  ccr = av_clipl_int32((int64_t)(ccr << exp) + (1 << 15)) >> 16;
1732  exp <<= 1;
1733  ccr *= ccr;
1734  temp = normalize_bits_int32(ccr);
1735  ccr = ccr << temp >> 16;
1736  exp += temp;
1737 
1738  temp = normalize_bits_int32(orig_eng);
1739  eng = av_clipl_int32((int64_t)(orig_eng << temp) + (1 << 15)) >> 16;
1740  exp -= temp;
1741 
1742  if (ccr >= eng) {
1743  exp--;
1744  ccr >>= 1;
1745  }
1746  if (exp > max_exp)
1747  continue;
1748 
1749  if (exp + 1 < max_exp)
1750  goto update;
1751 
1752  /* Equalize exponents before comparison */
1753  if (exp + 1 == max_exp)
1754  temp = max_ccr >> 1;
1755  else
1756  temp = max_ccr;
1757  ccr_eng = ccr * max_eng;
1758  diff = ccr_eng - eng * temp;
1759  if (diff > 0 && (i - index < PITCH_MIN || diff > ccr_eng >> 2)) {
1760 update:
1761  index = i;
1762  max_exp = exp;
1763  max_ccr = ccr;
1764  max_eng = eng;
1765  }
1766  }
1767  return index;
1768 }
1769 
1770 /**
1771  * Compute harmonic noise filter parameters.
1772  *
1773  * @param buf perceptually weighted speech
1774  * @param pitch_lag open loop pitch period
1775  * @param hf harmonic filter parameters
1776  */
1777 static void comp_harmonic_coeff(int16_t *buf, int16_t pitch_lag, HFParam *hf)
1778 {
1779  int ccr, eng, max_ccr, max_eng;
1780  int exp, max, diff;
1781  int energy[15];
1782  int i, j;
1783 
1784  for (i = 0, j = pitch_lag - 3; j <= pitch_lag + 3; i++, j++) {
1785  /* Compute residual energy */
1786  energy[i << 1] = ff_dot_product(buf - j, buf - j, SUBFRAME_LEN);
1787  /* Compute correlation */
1788  energy[(i << 1) + 1] = ff_dot_product(buf, buf - j, SUBFRAME_LEN);
1789  }
1790 
1791  /* Compute target energy */
1792  energy[14] = ff_dot_product(buf, buf, SUBFRAME_LEN);
1793 
1794  /* Normalize */
1795  max = 0;
1796  for (i = 0; i < 15; i++)
1797  max = FFMAX(max, FFABS(energy[i]));
1798 
1799  exp = normalize_bits_int32(max);
1800  for (i = 0; i < 15; i++) {
1801  energy[i] = av_clipl_int32((int64_t)(energy[i] << exp) +
1802  (1 << 15)) >> 16;
1803  }
1804 
1805  hf->index = -1;
1806  hf->gain = 0;
1807  max_ccr = 1;
1808  max_eng = 0x7fff;
1809 
1810  for (i = 0; i <= 6; i++) {
1811  eng = energy[i << 1];
1812  ccr = energy[(i << 1) + 1];
1813 
1814  if (ccr <= 0)
1815  continue;
1816 
1817  ccr = (ccr * ccr + (1 << 14)) >> 15;
1818  diff = ccr * max_eng - eng * max_ccr;
1819  if (diff > 0) {
1820  max_ccr = ccr;
1821  max_eng = eng;
1822  hf->index = i;
1823  }
1824  }
1825 
1826  if (hf->index == -1) {
1827  hf->index = pitch_lag;
1828  return;
1829  }
1830 
1831  eng = energy[14] * max_eng;
1832  eng = (eng >> 2) + (eng >> 3);
1833  ccr = energy[(hf->index << 1) + 1] * energy[(hf->index << 1) + 1];
1834  if (eng < ccr) {
1835  eng = energy[(hf->index << 1) + 1];
1836 
1837  if (eng >= max_eng)
1838  hf->gain = 0x2800;
1839  else
1840  hf->gain = ((eng << 15) / max_eng * 0x2800 + (1 << 14)) >> 15;
1841  }
1842  hf->index += pitch_lag - 3;
1843 }
1844 
1845 /**
1846  * Apply the harmonic noise shaping filter.
1847  *
1848  * @param hf filter parameters
1849  */
1850 static void harmonic_filter(HFParam *hf, const int16_t *src, int16_t *dest)
1851 {
1852  int i;
1853 
1854  for (i = 0; i < SUBFRAME_LEN; i++) {
1855  int64_t temp = hf->gain * src[i - hf->index] << 1;
1856  dest[i] = av_clipl_int32((src[i] << 16) - temp + (1 << 15)) >> 16;
1857  }
1858 }
1859 
1860 static void harmonic_noise_sub(HFParam *hf, const int16_t *src, int16_t *dest)
1861 {
1862  int i;
1863  for (i = 0; i < SUBFRAME_LEN; i++) {
1864  int64_t temp = hf->gain * src[i - hf->index] << 1;
1865  dest[i] = av_clipl_int32(((dest[i] - src[i]) << 16) + temp +
1866  (1 << 15)) >> 16;
1867 
1868  }
1869 }
1870 
1871 /**
1872  * Combined synthesis and formant perceptual weighting filer.
1873  *
1874  * @param qnt_lpc quantized lpc coefficients
1875  * @param perf_lpc perceptual filter coefficients
1876  * @param perf_fir perceptual filter fir memory
1877  * @param perf_iir perceptual filter iir memory
1878  * @param scale the filter output will be scaled by 2^scale
1879  */
1880 static void synth_percept_filter(int16_t *qnt_lpc, int16_t *perf_lpc,
1881  int16_t *perf_fir, int16_t *perf_iir,
1882  const int16_t *src, int16_t *dest, int scale)
1883 {
1884  int i, j;
1885  int16_t buf_16[SUBFRAME_LEN + LPC_ORDER];
1886  int64_t buf[SUBFRAME_LEN];
1887 
1888  int16_t *bptr_16 = buf_16 + LPC_ORDER;
1889 
1890  memcpy(buf_16, perf_fir, sizeof(int16_t) * LPC_ORDER);
1891  memcpy(dest - LPC_ORDER, perf_iir, sizeof(int16_t) * LPC_ORDER);
1892 
1893  for (i = 0; i < SUBFRAME_LEN; i++) {
1894  int64_t temp = 0;
1895  for (j = 1; j <= LPC_ORDER; j++)
1896  temp -= qnt_lpc[j - 1] * bptr_16[i - j];
1897 
1898  buf[i] = (src[i] << 15) + (temp << 3);
1899  bptr_16[i] = av_clipl_int32(buf[i] + (1 << 15)) >> 16;
1900  }
1901 
1902  for (i = 0; i < SUBFRAME_LEN; i++) {
1903  int64_t fir = 0, iir = 0;
1904  for (j = 1; j <= LPC_ORDER; j++) {
1905  fir -= perf_lpc[j - 1] * bptr_16[i - j];
1906  iir += perf_lpc[j + LPC_ORDER - 1] * dest[i - j];
1907  }
1908  dest[i] = av_clipl_int32(((buf[i] + (fir << 3)) << scale) + (iir << 3) +
1909  (1 << 15)) >> 16;
1910  }
1911  memcpy(perf_fir, buf_16 + SUBFRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1912  memcpy(perf_iir, dest + SUBFRAME_LEN - LPC_ORDER,
1913  sizeof(int16_t) * LPC_ORDER);
1914 }
1915 
1916 /**
1917  * Compute the adaptive codebook contribution.
1918  *
1919  * @param buf input signal
1920  * @param index the current subframe index
1921  */
1922 static void acb_search(G723_1_Context *p, int16_t *residual,
1923  int16_t *impulse_resp, const int16_t *buf,
1924  int index)
1925 {
1926 
1927  int16_t flt_buf[PITCH_ORDER][SUBFRAME_LEN];
1928 
1929  const int16_t *cb_tbl = adaptive_cb_gain85;
1930 
1931  int ccr_buf[PITCH_ORDER * SUBFRAMES << 2];
1932 
1933  int pitch_lag = p->pitch_lag[index >> 1];
1934  int acb_lag = 1;
1935  int acb_gain = 0;
1936  int odd_frame = index & 1;
1937  int iter = 3 + odd_frame;
1938  int count = 0;
1939  int tbl_size = 85;
1940 
1941  int i, j, k, l, max;
1942  int64_t temp;
1943 
1944  if (!odd_frame) {
1945  if (pitch_lag == PITCH_MIN)
1946  pitch_lag++;
1947  else
1948  pitch_lag = FFMIN(pitch_lag, PITCH_MAX - 5);
1949  }
1950 
1951  for (i = 0; i < iter; i++) {
1952  get_residual(residual, p->prev_excitation, pitch_lag + i - 1);
1953 
1954  for (j = 0; j < SUBFRAME_LEN; j++) {
1955  temp = 0;
1956  for (k = 0; k <= j; k++)
1957  temp += residual[PITCH_ORDER - 1 + k] * impulse_resp[j - k];
1958  flt_buf[PITCH_ORDER - 1][j] = av_clipl_int32((temp << 1) +
1959  (1 << 15)) >> 16;
1960  }
1961 
1962  for (j = PITCH_ORDER - 2; j >= 0; j--) {
1963  flt_buf[j][0] = ((residual[j] << 13) + (1 << 14)) >> 15;
1964  for (k = 1; k < SUBFRAME_LEN; k++) {
1965  temp = (flt_buf[j + 1][k - 1] << 15) +
1966  residual[j] * impulse_resp[k];
1967  flt_buf[j][k] = av_clipl_int32((temp << 1) + (1 << 15)) >> 16;
1968  }
1969  }
1970 
1971  /* Compute crosscorrelation with the signal */
1972  for (j = 0; j < PITCH_ORDER; j++) {
1973  temp = ff_dot_product(buf, flt_buf[j], SUBFRAME_LEN);
1974  ccr_buf[count++] = av_clipl_int32(temp << 1);
1975  }
1976 
1977  /* Compute energies */
1978  for (j = 0; j < PITCH_ORDER; j++) {
1979  ccr_buf[count++] = dot_product(flt_buf[j], flt_buf[j],
1980  SUBFRAME_LEN);
1981  }
1982 
1983  for (j = 1; j < PITCH_ORDER; j++) {
1984  for (k = 0; k < j; k++) {
1985  temp = ff_dot_product(flt_buf[j], flt_buf[k], SUBFRAME_LEN);
1986  ccr_buf[count++] = av_clipl_int32(temp<<2);
1987  }
1988  }
1989  }
1990 
1991  /* Normalize and shorten */
1992  max = 0;
1993  for (i = 0; i < 20 * iter; i++)
1994  max = FFMAX(max, FFABS(ccr_buf[i]));
1995 
1996  temp = normalize_bits_int32(max);
1997 
1998  for (i = 0; i < 20 * iter; i++){
1999  ccr_buf[i] = av_clipl_int32((int64_t)(ccr_buf[i] << temp) +
2000  (1 << 15)) >> 16;
2001  }
2002 
2003  max = 0;
2004  for (i = 0; i < iter; i++) {
2005  /* Select quantization table */
2006  if (!odd_frame && pitch_lag + i - 1 >= SUBFRAME_LEN - 2 ||
2007  odd_frame && pitch_lag >= SUBFRAME_LEN - 2) {
2008  cb_tbl = adaptive_cb_gain170;
2009  tbl_size = 170;
2010  }
2011 
2012  for (j = 0, k = 0; j < tbl_size; j++, k += 20) {
2013  temp = 0;
2014  for (l = 0; l < 20; l++)
2015  temp += ccr_buf[20 * i + l] * cb_tbl[k + l];
2016  temp = av_clipl_int32(temp);
2017 
2018  if (temp > max) {
2019  max = temp;
2020  acb_gain = j;
2021  acb_lag = i;
2022  }
2023  }
2024  }
2025 
2026  if (!odd_frame) {
2027  pitch_lag += acb_lag - 1;
2028  acb_lag = 1;
2029  }
2030 
2031  p->pitch_lag[index >> 1] = pitch_lag;
2032  p->subframe[index].ad_cb_lag = acb_lag;
2033  p->subframe[index].ad_cb_gain = acb_gain;
2034 }
2035 
2036 /**
2037  * Subtract the adaptive codebook contribution from the input
2038  * to obtain the residual.
2039  *
2040  * @param buf target vector
2041  */
2042 static void sub_acb_contrib(const int16_t *residual, const int16_t *impulse_resp,
2043  int16_t *buf)
2044 {
2045  int i, j;
2046  /* Subtract adaptive CB contribution to obtain the residual */
2047  for (i = 0; i < SUBFRAME_LEN; i++) {
2048  int64_t temp = buf[i] << 14;
2049  for (j = 0; j <= i; j++)
2050  temp -= residual[j] * impulse_resp[i - j];
2051 
2052  buf[i] = av_clipl_int32((temp << 2) + (1 << 15)) >> 16;
2053  }
2054 }
2055 
2056 /**
2057  * Quantize the residual signal using the fixed codebook (MP-MLQ).
2058  *
2059  * @param optim optimized fixed codebook parameters
2060  * @param buf excitation vector
2061  */
2062 static void get_fcb_param(FCBParam *optim, int16_t *impulse_resp,
2063  int16_t *buf, int pulse_cnt, int pitch_lag)
2064 {
2065  FCBParam param;
2066  int16_t impulse_r[SUBFRAME_LEN];
2067  int16_t temp_corr[SUBFRAME_LEN];
2068  int16_t impulse_corr[SUBFRAME_LEN];
2069 
2070  int ccr1[SUBFRAME_LEN];
2071  int ccr2[SUBFRAME_LEN];
2072  int amp, err, max, max_amp_index, min, scale, i, j, k, l;
2073 
2074  int64_t temp;
2075 
2076  /* Update impulse response */
2077  memcpy(impulse_r, impulse_resp, sizeof(int16_t) * SUBFRAME_LEN);
2078  param.dirac_train = 0;
2079  if (pitch_lag < SUBFRAME_LEN - 2) {
2080  param.dirac_train = 1;
2081  gen_dirac_train(impulse_r, pitch_lag);
2082  }
2083 
2084  for (i = 0; i < SUBFRAME_LEN; i++)
2085  temp_corr[i] = impulse_r[i] >> 1;
2086 
2087  /* Compute impulse response autocorrelation */
2088  temp = dot_product(temp_corr, temp_corr, SUBFRAME_LEN);
2089 
2090  scale = normalize_bits_int32(temp);
2091  impulse_corr[0] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
2092 
2093  for (i = 1; i < SUBFRAME_LEN; i++) {
2094  temp = dot_product(temp_corr + i, temp_corr, SUBFRAME_LEN - i);
2095  impulse_corr[i] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
2096  }
2097 
2098  /* Compute crosscorrelation of impulse response with residual signal */
2099  scale -= 4;
2100  for (i = 0; i < SUBFRAME_LEN; i++){
2101  temp = dot_product(buf + i, impulse_r, SUBFRAME_LEN - i);
2102  if (scale < 0)
2103  ccr1[i] = temp >> -scale;
2104  else
2105  ccr1[i] = av_clipl_int32(temp << scale);
2106  }
2107 
2108  /* Search loop */
2109  for (i = 0; i < GRID_SIZE; i++) {
2110  /* Maximize the crosscorrelation */
2111  max = 0;
2112  for (j = i; j < SUBFRAME_LEN; j += GRID_SIZE) {
2113  temp = FFABS(ccr1[j]);
2114  if (temp >= max) {
2115  max = temp;
2116  param.pulse_pos[0] = j;
2117  }
2118  }
2119 
2120  /* Quantize the gain (max crosscorrelation/impulse_corr[0]) */
2121  amp = max;
2122  min = 1 << 30;
2123  max_amp_index = GAIN_LEVELS - 2;
2124  for (j = max_amp_index; j >= 2; j--) {
2125  temp = av_clipl_int32((int64_t)fixed_cb_gain[j] *
2126  impulse_corr[0] << 1);
2127  temp = FFABS(temp - amp);
2128  if (temp < min) {
2129  min = temp;
2130  max_amp_index = j;
2131  }
2132  }
2133 
2134  max_amp_index--;
2135  /* Select additional gain values */
2136  for (j = 1; j < 5; j++) {
2137  for (k = i; k < SUBFRAME_LEN; k += GRID_SIZE) {
2138  temp_corr[k] = 0;
2139  ccr2[k] = ccr1[k];
2140  }
2141  param.amp_index = max_amp_index + j - 2;
2142  amp = fixed_cb_gain[param.amp_index];
2143 
2144  param.pulse_sign[0] = (ccr2[param.pulse_pos[0]] < 0) ? -amp : amp;
2145  temp_corr[param.pulse_pos[0]] = 1;
2146 
2147  for (k = 1; k < pulse_cnt; k++) {
2148  max = -1 << 30;
2149  for (l = i; l < SUBFRAME_LEN; l += GRID_SIZE) {
2150  if (temp_corr[l])
2151  continue;
2152  temp = impulse_corr[FFABS(l - param.pulse_pos[k - 1])];
2153  temp = av_clipl_int32((int64_t)temp *
2154  param.pulse_sign[k - 1] << 1);
2155  ccr2[l] -= temp;
2156  temp = FFABS(ccr2[l]);
2157  if (temp > max) {
2158  max = temp;
2159  param.pulse_pos[k] = l;
2160  }
2161  }
2162 
2163  param.pulse_sign[k] = (ccr2[param.pulse_pos[k]] < 0) ?
2164  -amp : amp;
2165  temp_corr[param.pulse_pos[k]] = 1;
2166  }
2167 
2168  /* Create the error vector */
2169  memset(temp_corr, 0, sizeof(int16_t) * SUBFRAME_LEN);
2170 
2171  for (k = 0; k < pulse_cnt; k++)
2172  temp_corr[param.pulse_pos[k]] = param.pulse_sign[k];
2173 
2174  for (k = SUBFRAME_LEN - 1; k >= 0; k--) {
2175  temp = 0;
2176  for (l = 0; l <= k; l++) {
2177  int prod = av_clipl_int32((int64_t)temp_corr[l] *
2178  impulse_r[k - l] << 1);
2179  temp = av_clipl_int32(temp + prod);
2180  }
2181  temp_corr[k] = temp << 2 >> 16;
2182  }
2183 
2184  /* Compute square of error */
2185  err = 0;
2186  for (k = 0; k < SUBFRAME_LEN; k++) {
2187  int64_t prod;
2188  prod = av_clipl_int32((int64_t)buf[k] * temp_corr[k] << 1);
2189  err = av_clipl_int32(err - prod);
2190  prod = av_clipl_int32((int64_t)temp_corr[k] * temp_corr[k]);
2191  err = av_clipl_int32(err + prod);
2192  }
2193 
2194  /* Minimize */
2195  if (err < optim->min_err) {
2196  optim->min_err = err;
2197  optim->grid_index = i;
2198  optim->amp_index = param.amp_index;
2199  optim->dirac_train = param.dirac_train;
2200 
2201  for (k = 0; k < pulse_cnt; k++) {
2202  optim->pulse_sign[k] = param.pulse_sign[k];
2203  optim->pulse_pos[k] = param.pulse_pos[k];
2204  }
2205  }
2206  }
2207  }
2208 }
2209 
2210 /**
2211  * Encode the pulse position and gain of the current subframe.
2212  *
2213  * @param optim optimized fixed CB parameters
2214  * @param buf excitation vector
2215  */
2216 static void pack_fcb_param(G723_1_Subframe *subfrm, FCBParam *optim,
2217  int16_t *buf, int pulse_cnt)
2218 {
2219  int i, j;
2220 
2221  j = PULSE_MAX - pulse_cnt;
2222 
2223  subfrm->pulse_sign = 0;
2224  subfrm->pulse_pos = 0;
2225 
2226  for (i = 0; i < SUBFRAME_LEN >> 1; i++) {
2227  int val = buf[optim->grid_index + (i << 1)];
2228  if (!val) {
2229  subfrm->pulse_pos += combinatorial_table[j][i];
2230  } else {
2231  subfrm->pulse_sign <<= 1;
2232  if (val < 0) subfrm->pulse_sign++;
2233  j++;
2234 
2235  if (j == PULSE_MAX) break;
2236  }
2237  }
2238  subfrm->amp_index = optim->amp_index;
2239  subfrm->grid_index = optim->grid_index;
2240  subfrm->dirac_train = optim->dirac_train;
2241 }
2242 
2243 /**
2244  * Compute the fixed codebook excitation.
2245  *
2246  * @param buf target vector
2247  * @param impulse_resp impulse response of the combined filter
2248  */
2249 static void fcb_search(G723_1_Context *p, int16_t *impulse_resp,
2250  int16_t *buf, int index)
2251 {
2252  FCBParam optim;
2253  int pulse_cnt = pulses[index];
2254  int i;
2255 
2256  optim.min_err = 1 << 30;
2257  get_fcb_param(&optim, impulse_resp, buf, pulse_cnt, SUBFRAME_LEN);
2258 
2259  if (p->pitch_lag[index >> 1] < SUBFRAME_LEN - 2) {
2260  get_fcb_param(&optim, impulse_resp, buf, pulse_cnt,
2261  p->pitch_lag[index >> 1]);
2262  }
2263 
2264  /* Reconstruct the excitation */
2265  memset(buf, 0, sizeof(int16_t) * SUBFRAME_LEN);
2266  for (i = 0; i < pulse_cnt; i++)
2267  buf[optim.pulse_pos[i]] = optim.pulse_sign[i];
2268 
2269  pack_fcb_param(&p->subframe[index], &optim, buf, pulse_cnt);
2270 
2271  if (optim.dirac_train)
2272  gen_dirac_train(buf, p->pitch_lag[index >> 1]);
2273 }
2274 
2275 /**
2276  * Pack the frame parameters into output bitstream.
2277  *
2278  * @param frame output buffer
2279  * @param size size of the buffer
2280  */
2281 static int pack_bitstream(G723_1_Context *p, unsigned char *frame, int size)
2282 {
2283  PutBitContext pb;
2284  int info_bits, i, temp;
2285 
2286  init_put_bits(&pb, frame, size);
2287 
2288  if (p->cur_rate == RATE_6300) {
2289  info_bits = 0;
2290  put_bits(&pb, 2, info_bits);
2291  }
2292 
2293  put_bits(&pb, 8, p->lsp_index[2]);
2294  put_bits(&pb, 8, p->lsp_index[1]);
2295  put_bits(&pb, 8, p->lsp_index[0]);
2296 
2297  put_bits(&pb, 7, p->pitch_lag[0] - PITCH_MIN);
2298  put_bits(&pb, 2, p->subframe[1].ad_cb_lag);
2299  put_bits(&pb, 7, p->pitch_lag[1] - PITCH_MIN);
2300  put_bits(&pb, 2, p->subframe[3].ad_cb_lag);
2301 
2302  /* Write 12 bit combined gain */
2303  for (i = 0; i < SUBFRAMES; i++) {
2304  temp = p->subframe[i].ad_cb_gain * GAIN_LEVELS +
2305  p->subframe[i].amp_index;
2306  if (p->cur_rate == RATE_6300)
2307  temp += p->subframe[i].dirac_train << 11;
2308  put_bits(&pb, 12, temp);
2309  }
2310 
2311  put_bits(&pb, 1, p->subframe[0].grid_index);
2312  put_bits(&pb, 1, p->subframe[1].grid_index);
2313  put_bits(&pb, 1, p->subframe[2].grid_index);
2314  put_bits(&pb, 1, p->subframe[3].grid_index);
2315 
2316  if (p->cur_rate == RATE_6300) {
2317  skip_put_bits(&pb, 1); /* reserved bit */
2318 
2319  /* Write 13 bit combined position index */
2320  temp = (p->subframe[0].pulse_pos >> 16) * 810 +
2321  (p->subframe[1].pulse_pos >> 14) * 90 +
2322  (p->subframe[2].pulse_pos >> 16) * 9 +
2323  (p->subframe[3].pulse_pos >> 14);
2324  put_bits(&pb, 13, temp);
2325 
2326  put_bits(&pb, 16, p->subframe[0].pulse_pos & 0xffff);
2327  put_bits(&pb, 14, p->subframe[1].pulse_pos & 0x3fff);
2328  put_bits(&pb, 16, p->subframe[2].pulse_pos & 0xffff);
2329  put_bits(&pb, 14, p->subframe[3].pulse_pos & 0x3fff);
2330 
2331  put_bits(&pb, 6, p->subframe[0].pulse_sign);
2332  put_bits(&pb, 5, p->subframe[1].pulse_sign);
2333  put_bits(&pb, 6, p->subframe[2].pulse_sign);
2334  put_bits(&pb, 5, p->subframe[3].pulse_sign);
2335  }
2336 
2337  flush_put_bits(&pb);
2338  return frame_size[info_bits];
2339 }
2340 
2341 static int g723_1_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
2342  const AVFrame *frame, int *got_packet_ptr)
2343 {
2344  G723_1_Context *p = avctx->priv_data;
2345  int16_t unq_lpc[LPC_ORDER * SUBFRAMES];
2346  int16_t qnt_lpc[LPC_ORDER * SUBFRAMES];
2347  int16_t cur_lsp[LPC_ORDER];
2348  int16_t weighted_lpc[LPC_ORDER * SUBFRAMES << 1];
2349  int16_t vector[FRAME_LEN + PITCH_MAX];
2350  int offset, ret;
2351  int16_t *in = (const int16_t *)frame->data[0];
2352 
2353  HFParam hf[4];
2354  int i, j;
2355 
2356  highpass_filter(in, &p->hpf_fir_mem, &p->hpf_iir_mem);
2357 
2358  memcpy(vector, p->prev_data, HALF_FRAME_LEN * sizeof(int16_t));
2359  memcpy(vector + HALF_FRAME_LEN, in, FRAME_LEN * sizeof(int16_t));
2360 
2361  comp_lpc_coeff(vector, unq_lpc);
2362  lpc2lsp(&unq_lpc[LPC_ORDER * 3], p->prev_lsp, cur_lsp);
2363  lsp_quantize(p->lsp_index, cur_lsp, p->prev_lsp);
2364 
2365  /* Update memory */
2366  memcpy(vector + LPC_ORDER, p->prev_data + SUBFRAME_LEN,
2367  sizeof(int16_t) * SUBFRAME_LEN);
2368  memcpy(vector + LPC_ORDER + SUBFRAME_LEN, in,
2369  sizeof(int16_t) * (HALF_FRAME_LEN + SUBFRAME_LEN));
2370  memcpy(p->prev_data, in + HALF_FRAME_LEN,
2371  sizeof(int16_t) * HALF_FRAME_LEN);
2372  memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2373 
2374  perceptual_filter(p, weighted_lpc, unq_lpc, vector);
2375 
2376  memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2377  memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2378  memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2379 
2380  scale_vector(vector, vector, FRAME_LEN + PITCH_MAX);
2381 
2382  p->pitch_lag[0] = estimate_pitch(vector, PITCH_MAX);
2383  p->pitch_lag[1] = estimate_pitch(vector, PITCH_MAX + HALF_FRAME_LEN);
2384 
2385  for (i = PITCH_MAX, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2386  comp_harmonic_coeff(vector + i, p->pitch_lag[j >> 1], hf + j);
2387 
2388  memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2389  memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2390  memcpy(p->prev_weight_sig, vector + FRAME_LEN, sizeof(int16_t) * PITCH_MAX);
2391 
2392  for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2393  harmonic_filter(hf + j, vector + PITCH_MAX + i, in + i);
2394 
2395  inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, 0);
2396  lsp_interpolate(qnt_lpc, cur_lsp, p->prev_lsp);
2397 
2398  memcpy(p->prev_lsp, cur_lsp, sizeof(int16_t) * LPC_ORDER);
2399 
2400  offset = 0;
2401  for (i = 0; i < SUBFRAMES; i++) {
2402  int16_t impulse_resp[SUBFRAME_LEN];
2403  int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
2404  int16_t flt_in[SUBFRAME_LEN];
2405  int16_t zero[LPC_ORDER], fir[LPC_ORDER], iir[LPC_ORDER];
2406 
2407  /**
2408  * Compute the combined impulse response of the synthesis filter,
2409  * formant perceptual weighting filter and harmonic noise shaping filter
2410  */
2411  memset(zero, 0, sizeof(int16_t) * LPC_ORDER);
2412  memset(vector, 0, sizeof(int16_t) * PITCH_MAX);
2413  memset(flt_in, 0, sizeof(int16_t) * SUBFRAME_LEN);
2414 
2415  flt_in[0] = 1 << 13; /* Unit impulse */
2416  synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2417  zero, zero, flt_in, vector + PITCH_MAX, 1);
2418  harmonic_filter(hf + i, vector + PITCH_MAX, impulse_resp);
2419 
2420  /* Compute the combined zero input response */
2421  flt_in[0] = 0;
2422  memcpy(fir, p->perf_fir_mem, sizeof(int16_t) * LPC_ORDER);
2423  memcpy(iir, p->perf_iir_mem, sizeof(int16_t) * LPC_ORDER);
2424 
2425  synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2426  fir, iir, flt_in, vector + PITCH_MAX, 0);
2427  memcpy(vector, p->harmonic_mem, sizeof(int16_t) * PITCH_MAX);
2428  harmonic_noise_sub(hf + i, vector + PITCH_MAX, in);
2429 
2430  acb_search(p, residual, impulse_resp, in, i);
2431  gen_acb_excitation(residual, p->prev_excitation,p->pitch_lag[i >> 1],
2432  &p->subframe[i], p->cur_rate);
2433  sub_acb_contrib(residual, impulse_resp, in);
2434 
2435  fcb_search(p, impulse_resp, in, i);
2436 
2437  /* Reconstruct the excitation */
2438  gen_acb_excitation(impulse_resp, p->prev_excitation, p->pitch_lag[i >> 1],
2439  &p->subframe[i], RATE_6300);
2440 
2441  memmove(p->prev_excitation, p->prev_excitation + SUBFRAME_LEN,
2442  sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2443  for (j = 0; j < SUBFRAME_LEN; j++)
2444  in[j] = av_clip_int16((in[j] << 1) + impulse_resp[j]);
2445  memcpy(p->prev_excitation + PITCH_MAX - SUBFRAME_LEN, in,
2446  sizeof(int16_t) * SUBFRAME_LEN);
2447 
2448  /* Update filter memories */
2449  synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2450  p->perf_fir_mem, p->perf_iir_mem,
2451  in, vector + PITCH_MAX, 0);
2452  memmove(p->harmonic_mem, p->harmonic_mem + SUBFRAME_LEN,
2453  sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2454  memcpy(p->harmonic_mem + PITCH_MAX - SUBFRAME_LEN, vector + PITCH_MAX,
2455  sizeof(int16_t) * SUBFRAME_LEN);
2456 
2457  in += SUBFRAME_LEN;
2458  offset += LPC_ORDER;
2459  }
2460 
2461  if ((ret = ff_alloc_packet2(avctx, avpkt, 24)) < 0)
2462  return ret;
2463 
2464  *got_packet_ptr = 1;
2465  avpkt->size = pack_bitstream(p, avpkt->data, avpkt->size);
2466  return 0;
2467 }
2468 
2469 AVCodec ff_g723_1_encoder = {
2470  .name = "g723_1",
2471  .type = AVMEDIA_TYPE_AUDIO,
2472  .id = AV_CODEC_ID_G723_1,
2473  .priv_data_size = sizeof(G723_1_Context),
2474  .init = g723_1_encode_init,
2475  .encode2 = g723_1_encode_frame,
2476  .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
2477  .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_S16,
2478  AV_SAMPLE_FMT_NONE},
2479 };
2480 #endif
Generated on Sat May 25 2013 04:01:04 for FFmpeg by   doxygen 1.8.2