FFmpeg  4.4.5
g723_1enc.c
Go to the documentation of this file.
1 /*
2  * G.723.1 compatible encoder
3  * Copyright (c) Mohamed Naufal <naufal22@gmail.com>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21 
22 /**
23  * @file
24  * G.723.1 compatible encoder
25  */
26 
27 #include <stdint.h>
28 #include <string.h>
29 
31 #include "libavutil/common.h"
32 #include "libavutil/mem.h"
33 #include "libavutil/opt.h"
34 
35 #include "avcodec.h"
36 #include "celp_math.h"
37 #include "g723_1.h"
38 #include "internal.h"
39 
40 #define BITSTREAM_WRITER_LE
41 #include "put_bits.h"
42 
43 /**
44  * Hamming window coefficients scaled by 2^15
45  */
46 static const int16_t hamming_window[LPC_FRAME] = {
47  2621, 2631, 2659, 2705, 2770, 2853, 2955, 3074, 3212, 3367,
48  3541, 3731, 3939, 4164, 4405, 4663, 4937, 5226, 5531, 5851,
49  6186, 6534, 6897, 7273, 7661, 8062, 8475, 8899, 9334, 9780,
50  10235, 10699, 11172, 11653, 12141, 12636, 13138, 13645, 14157, 14673,
51  15193, 15716, 16242, 16769, 17298, 17827, 18356, 18884, 19411, 19935,
52  20457, 20975, 21489, 21999, 22503, 23002, 23494, 23978, 24455, 24924,
53  25384, 25834, 26274, 26704, 27122, 27529, 27924, 28306, 28675, 29031,
54  29373, 29700, 30012, 30310, 30592, 30857, 31107, 31340, 31557, 31756,
55  31938, 32102, 32249, 32377, 32488, 32580, 32654, 32710, 32747, 32766,
56  32766, 32747, 32710, 32654, 32580, 32488, 32377, 32249, 32102, 31938,
57  31756, 31557, 31340, 31107, 30857, 30592, 30310, 30012, 29700, 29373,
58  29031, 28675, 28306, 27924, 27529, 27122, 26704, 26274, 25834, 25384,
59  24924, 24455, 23978, 23494, 23002, 22503, 21999, 21489, 20975, 20457,
60  19935, 19411, 18884, 18356, 17827, 17298, 16769, 16242, 15716, 15193,
61  14673, 14157, 13645, 13138, 12636, 12141, 11653, 11172, 10699, 10235,
62  9780, 9334, 8899, 8475, 8062, 7661, 7273, 6897, 6534, 6186,
63  5851, 5531, 5226, 4937, 4663, 4405, 4164, 3939, 3731, 3541,
64  3367, 3212, 3074, 2955, 2853, 2770, 2705, 2659, 2631, 2621
65 };
66 
67 /**
68  * Binomial window coefficients scaled by 2^15
69  */
70 static const int16_t binomial_window[LPC_ORDER] = {
71  32749, 32695, 32604, 32477, 32315, 32118, 31887, 31622, 31324, 30995
72 };
73 
74 /**
75  * 0.994^i scaled by 2^15
76  */
77 static const int16_t bandwidth_expand[LPC_ORDER] = {
78  32571, 32376, 32182, 31989, 31797, 31606, 31416, 31228, 31040, 30854
79 };
80 
81 /**
82  * 0.5^i scaled by 2^15
83  */
84 static const int16_t percept_flt_tbl[2][LPC_ORDER] = {
85  /* Zero part */
86  {29491, 26542, 23888, 21499, 19349, 17414, 15673, 14106, 12695, 11425},
87  /* Pole part */
88  {16384, 8192, 4096, 2048, 1024, 512, 256, 128, 64, 32}
89 };
90 
92 {
93  G723_1_Context *s = avctx->priv_data;
94  G723_1_ChannelContext *p = &s->ch[0];
95 
96  if (avctx->sample_rate != 8000) {
97  av_log(avctx, AV_LOG_ERROR, "Only 8000Hz sample rate supported\n");
98  return AVERROR(EINVAL);
99  }
100 
101  if (avctx->channels != 1) {
102  av_log(avctx, AV_LOG_ERROR, "Only mono supported\n");
103  return AVERROR(EINVAL);
104  }
105 
106  if (avctx->bit_rate == 6300) {
107  p->cur_rate = RATE_6300;
108  } else if (avctx->bit_rate == 5300) {
109  av_log(avctx, AV_LOG_ERROR, "Use bitrate 6300 instead of 5300.\n");
110  avpriv_report_missing_feature(avctx, "Bitrate 5300");
111  return AVERROR_PATCHWELCOME;
112  } else {
113  av_log(avctx, AV_LOG_ERROR, "Bitrate not supported, use 6300\n");
114  return AVERROR(EINVAL);
115  }
116  avctx->frame_size = 240;
117  memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(int16_t));
118 
119  return 0;
120 }
121 
122 /**
123  * Remove DC component from the input signal.
124  *
125  * @param buf input signal
126  * @param fir zero memory
127  * @param iir pole memory
128  */
129 static void highpass_filter(int16_t *buf, int16_t *fir, int *iir)
130 {
131  int i;
132  for (i = 0; i < FRAME_LEN; i++) {
133  *iir = (buf[i] << 15) + ((-*fir) << 15) + MULL2(*iir, 0x7f00);
134  *fir = buf[i];
135  buf[i] = av_clipl_int32((int64_t)*iir + (1 << 15)) >> 16;
136  }
137 }
138 
139 /**
140  * Estimate autocorrelation of the input vector.
141  *
142  * @param buf input buffer
143  * @param autocorr autocorrelation coefficients vector
144  */
145 static void comp_autocorr(int16_t *buf, int16_t *autocorr)
146 {
147  int i, scale, temp;
148  int16_t vector[LPC_FRAME];
149 
150  ff_g723_1_scale_vector(vector, buf, LPC_FRAME);
151 
152  /* Apply the Hamming window */
153  for (i = 0; i < LPC_FRAME; i++)
154  vector[i] = (vector[i] * hamming_window[i] + (1 << 14)) >> 15;
155 
156  /* Compute the first autocorrelation coefficient */
157  temp = ff_dot_product(vector, vector, LPC_FRAME);
158 
159  /* Apply a white noise correlation factor of (1025/1024) */
160  temp += temp >> 10;
161 
162  /* Normalize */
163  scale = ff_g723_1_normalize_bits(temp, 31);
164  autocorr[0] = av_clipl_int32((int64_t) (temp << scale) +
165  (1 << 15)) >> 16;
166 
167  /* Compute the remaining coefficients */
168  if (!autocorr[0]) {
169  memset(autocorr + 1, 0, LPC_ORDER * sizeof(int16_t));
170  } else {
171  for (i = 1; i <= LPC_ORDER; i++) {
172  temp = ff_dot_product(vector, vector + i, LPC_FRAME - i);
173  temp = MULL2((temp << scale), binomial_window[i - 1]);
174  autocorr[i] = av_clipl_int32((int64_t) temp + (1 << 15)) >> 16;
175  }
176  }
177 }
178 
179 /**
180  * Use Levinson-Durbin recursion to compute LPC coefficients from
181  * autocorrelation values.
182  *
183  * @param lpc LPC coefficients vector
184  * @param autocorr autocorrelation coefficients vector
185  * @param error prediction error
186  */
187 static void levinson_durbin(int16_t *lpc, int16_t *autocorr, int16_t error)
188 {
189  int16_t vector[LPC_ORDER];
190  int16_t partial_corr;
191  int i, j, temp;
192 
193  memset(lpc, 0, LPC_ORDER * sizeof(int16_t));
194 
195  for (i = 0; i < LPC_ORDER; i++) {
196  /* Compute the partial correlation coefficient */
197  temp = 0;
198  for (j = 0; j < i; j++)
199  temp -= lpc[j] * autocorr[i - j - 1];
200  temp = ((autocorr[i] << 13) + temp) << 3;
201 
202  if (FFABS(temp) >= (error << 16))
203  break;
204 
205  partial_corr = temp / (error << 1);
206 
207  lpc[i] = av_clipl_int32((int64_t) (partial_corr << 14) +
208  (1 << 15)) >> 16;
209 
210  /* Update the prediction error */
211  temp = MULL2(temp, partial_corr);
212  error = av_clipl_int32((int64_t) (error << 16) - temp +
213  (1 << 15)) >> 16;
214 
215  memcpy(vector, lpc, i * sizeof(int16_t));
216  for (j = 0; j < i; j++) {
217  temp = partial_corr * vector[i - j - 1] << 1;
218  lpc[j] = av_clipl_int32((int64_t) (lpc[j] << 16) - temp +
219  (1 << 15)) >> 16;
220  }
221  }
222 }
223 
224 /**
225  * Calculate LPC coefficients for the current frame.
226  *
227  * @param buf current frame
228  * @param prev_data 2 trailing subframes of the previous frame
229  * @param lpc LPC coefficients vector
230  */
231 static void comp_lpc_coeff(int16_t *buf, int16_t *lpc)
232 {
233  int16_t autocorr[(LPC_ORDER + 1) * SUBFRAMES];
234  int16_t *autocorr_ptr = autocorr;
235  int16_t *lpc_ptr = lpc;
236  int i, j;
237 
238  for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
239  comp_autocorr(buf + i, autocorr_ptr);
240  levinson_durbin(lpc_ptr, autocorr_ptr + 1, autocorr_ptr[0]);
241 
242  lpc_ptr += LPC_ORDER;
243  autocorr_ptr += LPC_ORDER + 1;
244  }
245 }
246 
247 static void lpc2lsp(int16_t *lpc, int16_t *prev_lsp, int16_t *lsp)
248 {
249  int f[LPC_ORDER + 2]; ///< coefficients of the sum and difference
250  ///< polynomials (F1, F2) ordered as
251  ///< f1[0], f2[0], ...., f1[5], f2[5]
252 
253  int max, shift, cur_val, prev_val, count, p;
254  int i, j;
255  int64_t temp;
256 
257  /* Initialize f1[0] and f2[0] to 1 in Q25 */
258  for (i = 0; i < LPC_ORDER; i++)
259  lsp[i] = (lpc[i] * bandwidth_expand[i] + (1 << 14)) >> 15;
260 
261  /* Apply bandwidth expansion on the LPC coefficients */
262  f[0] = f[1] = 1 << 25;
263 
264  /* Compute the remaining coefficients */
265  for (i = 0; i < LPC_ORDER / 2; i++) {
266  /* f1 */
267  f[2 * i + 2] = -f[2 * i] - ((lsp[i] + lsp[LPC_ORDER - 1 - i]) << 12);
268  /* f2 */
269  f[2 * i + 3] = f[2 * i + 1] - ((lsp[i] - lsp[LPC_ORDER - 1 - i]) << 12);
270  }
271 
272  /* Divide f1[5] and f2[5] by 2 for use in polynomial evaluation */
273  f[LPC_ORDER] >>= 1;
274  f[LPC_ORDER + 1] >>= 1;
275 
276  /* Normalize and shorten */
277  max = FFABS(f[0]);
278  for (i = 1; i < LPC_ORDER + 2; i++)
279  max = FFMAX(max, FFABS(f[i]));
280 
282 
283  for (i = 0; i < LPC_ORDER + 2; i++)
284  f[i] = av_clipl_int32((int64_t) (f[i] << shift) + (1 << 15)) >> 16;
285 
286  /**
287  * Evaluate F1 and F2 at uniform intervals of pi/256 along the
288  * unit circle and check for zero crossings.
289  */
290  p = 0;
291  temp = 0;
292  for (i = 0; i <= LPC_ORDER / 2; i++)
294  prev_val = av_clipl_int32(temp << 1);
295  count = 0;
296  for (i = 1; i < COS_TBL_SIZE / 2; i++) {
297  /* Evaluate */
298  temp = 0;
299  for (j = 0; j <= LPC_ORDER / 2; j++)
300  temp += f[LPC_ORDER - 2 * j + p] * ff_g723_1_cos_tab[i * j % COS_TBL_SIZE];
301  cur_val = av_clipl_int32(temp << 1);
302 
303  /* Check for sign change, indicating a zero crossing */
304  if ((cur_val ^ prev_val) < 0) {
305  int abs_cur = FFABS(cur_val);
306  int abs_prev = FFABS(prev_val);
307  int sum = abs_cur + abs_prev;
308 
309  shift = ff_g723_1_normalize_bits(sum, 31);
310  sum <<= shift;
311  abs_prev = abs_prev << shift >> 8;
312  lsp[count++] = ((i - 1) << 7) + (abs_prev >> 1) / (sum >> 16);
313 
314  if (count == LPC_ORDER)
315  break;
316 
317  /* Switch between sum and difference polynomials */
318  p ^= 1;
319 
320  /* Evaluate */
321  temp = 0;
322  for (j = 0; j <= LPC_ORDER / 2; j++)
323  temp += f[LPC_ORDER - 2 * j + p] *
325  cur_val = av_clipl_int32(temp << 1);
326  }
327  prev_val = cur_val;
328  }
329 
330  if (count != LPC_ORDER)
331  memcpy(lsp, prev_lsp, LPC_ORDER * sizeof(int16_t));
332 }
333 
334 /**
335  * Quantize the current LSP subvector.
336  *
337  * @param num band number
338  * @param offset offset of the current subvector in an LPC_ORDER vector
339  * @param size size of the current subvector
340  */
341 #define get_index(num, offset, size) \
342 { \
343  int error, max = -1; \
344  int16_t temp[4]; \
345  int i, j; \
346  \
347  for (i = 0; i < LSP_CB_SIZE; i++) { \
348  for (j = 0; j < size; j++){ \
349  temp[j] = (weight[j + (offset)] * ff_g723_1_lsp_band##num[i][j] + \
350  (1 << 14)) >> 15; \
351  } \
352  error = ff_g723_1_dot_product(lsp + (offset), temp, size) << 1; \
353  error -= ff_g723_1_dot_product(ff_g723_1_lsp_band##num[i], temp, size); \
354  if (error > max) { \
355  max = error; \
356  lsp_index[num] = i; \
357  } \
358  } \
359 }
360 
361 /**
362  * Vector quantize the LSP frequencies.
363  *
364  * @param lsp the current lsp vector
365  * @param prev_lsp the previous lsp vector
366  */
367 static void lsp_quantize(uint8_t *lsp_index, int16_t *lsp, int16_t *prev_lsp)
368 {
369  int16_t weight[LPC_ORDER];
370  int16_t min, max;
371  int shift, i;
372 
373  /* Calculate the VQ weighting vector */
374  weight[0] = (1 << 20) / (lsp[1] - lsp[0]);
375  weight[LPC_ORDER - 1] = (1 << 20) /
376  (lsp[LPC_ORDER - 1] - lsp[LPC_ORDER - 2]);
377 
378  for (i = 1; i < LPC_ORDER - 1; i++) {
379  min = FFMIN(lsp[i] - lsp[i - 1], lsp[i + 1] - lsp[i]);
380  if (min > 0x20)
381  weight[i] = (1 << 20) / min;
382  else
383  weight[i] = INT16_MAX;
384  }
385 
386  /* Normalize */
387  max = 0;
388  for (i = 0; i < LPC_ORDER; i++)
389  max = FFMAX(weight[i], max);
390 
392  for (i = 0; i < LPC_ORDER; i++) {
393  weight[i] <<= shift;
394  }
395 
396  /* Compute the VQ target vector */
397  for (i = 0; i < LPC_ORDER; i++) {
398  lsp[i] -= dc_lsp[i] +
399  (((prev_lsp[i] - dc_lsp[i]) * 12288 + (1 << 14)) >> 15);
400  }
401 
402  get_index(0, 0, 3);
403  get_index(1, 3, 3);
404  get_index(2, 6, 4);
405 }
406 
407 /**
408  * Perform IIR filtering.
409  *
410  * @param fir_coef FIR coefficients
411  * @param iir_coef IIR coefficients
412  * @param src source vector
413  * @param dest destination vector
414  */
415 static void iir_filter(int16_t *fir_coef, int16_t *iir_coef,
416  int16_t *src, int16_t *dest)
417 {
418  int m, n;
419 
420  for (m = 0; m < SUBFRAME_LEN; m++) {
421  int64_t filter = 0;
422  for (n = 1; n <= LPC_ORDER; n++) {
423  filter -= fir_coef[n - 1] * src[m - n] -
424  iir_coef[n - 1] * dest[m - n];
425  }
426 
427  dest[m] = av_clipl_int32((src[m] << 16) + (filter << 3) +
428  (1 << 15)) >> 16;
429  }
430 }
431 
432 /**
433  * Apply the formant perceptual weighting filter.
434  *
435  * @param flt_coef filter coefficients
436  * @param unq_lpc unquantized lpc vector
437  */
438 static void perceptual_filter(G723_1_ChannelContext *p, int16_t *flt_coef,
439  int16_t *unq_lpc, int16_t *buf)
440 {
441  int16_t vector[FRAME_LEN + LPC_ORDER];
442  int i, j, k, l = 0;
443 
444  memcpy(buf, p->iir_mem, sizeof(int16_t) * LPC_ORDER);
445  memcpy(vector, p->fir_mem, sizeof(int16_t) * LPC_ORDER);
446  memcpy(vector + LPC_ORDER, buf + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
447 
448  for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
449  for (k = 0; k < LPC_ORDER; k++) {
450  flt_coef[k + 2 * l] = (unq_lpc[k + l] * percept_flt_tbl[0][k] +
451  (1 << 14)) >> 15;
452  flt_coef[k + 2 * l + LPC_ORDER] = (unq_lpc[k + l] *
453  percept_flt_tbl[1][k] +
454  (1 << 14)) >> 15;
455  }
456  iir_filter(flt_coef + 2 * l, flt_coef + 2 * l + LPC_ORDER,
457  vector + i, buf + i);
458  l += LPC_ORDER;
459  }
460  memcpy(p->iir_mem, buf + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
461  memcpy(p->fir_mem, vector + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
462 }
463 
464 /**
465  * Estimate the open loop pitch period.
466  *
467  * @param buf perceptually weighted speech
468  * @param start estimation is carried out from this position
469  */
470 static int estimate_pitch(int16_t *buf, int start)
471 {
472  int max_exp = 32;
473  int max_ccr = 0x4000;
474  int max_eng = 0x7fff;
475  int index = PITCH_MIN;
476  int offset = start - PITCH_MIN + 1;
477 
478  int ccr, eng, orig_eng, ccr_eng, exp;
479  int diff, temp;
480 
481  int i;
482 
483  orig_eng = ff_dot_product(buf + offset, buf + offset, HALF_FRAME_LEN);
484 
485  for (i = PITCH_MIN; i <= PITCH_MAX - 3; i++) {
486  offset--;
487 
488  /* Update energy and compute correlation */
489  orig_eng += buf[offset] * buf[offset] -
490  buf[offset + HALF_FRAME_LEN] * buf[offset + HALF_FRAME_LEN];
491  ccr = ff_dot_product(buf + start, buf + offset, HALF_FRAME_LEN);
492  if (ccr <= 0)
493  continue;
494 
495  /* Split into mantissa and exponent to maintain precision */
496  exp = ff_g723_1_normalize_bits(ccr, 31);
497  ccr = av_clipl_int32((int64_t) (ccr << exp) + (1 << 15)) >> 16;
498  exp <<= 1;
499  ccr *= ccr;
500  temp = ff_g723_1_normalize_bits(ccr, 31);
501  ccr = ccr << temp >> 16;
502  exp += temp;
503 
504  temp = ff_g723_1_normalize_bits(orig_eng, 31);
505  eng = av_clipl_int32((int64_t) (orig_eng << temp) + (1 << 15)) >> 16;
506  exp -= temp;
507 
508  if (ccr >= eng) {
509  exp--;
510  ccr >>= 1;
511  }
512  if (exp > max_exp)
513  continue;
514 
515  if (exp + 1 < max_exp)
516  goto update;
517 
518  /* Equalize exponents before comparison */
519  if (exp + 1 == max_exp)
520  temp = max_ccr >> 1;
521  else
522  temp = max_ccr;
523  ccr_eng = ccr * max_eng;
524  diff = ccr_eng - eng * temp;
525  if (diff > 0 && (i - index < PITCH_MIN || diff > ccr_eng >> 2)) {
526 update:
527  index = i;
528  max_exp = exp;
529  max_ccr = ccr;
530  max_eng = eng;
531  }
532  }
533  return index;
534 }
535 
536 /**
537  * Compute harmonic noise filter parameters.
538  *
539  * @param buf perceptually weighted speech
540  * @param pitch_lag open loop pitch period
541  * @param hf harmonic filter parameters
542  */
543 static void comp_harmonic_coeff(int16_t *buf, int16_t pitch_lag, HFParam *hf)
544 {
545  int ccr, eng, max_ccr, max_eng;
546  int exp, max, diff;
547  int energy[15];
548  int i, j;
549 
550  for (i = 0, j = pitch_lag - 3; j <= pitch_lag + 3; i++, j++) {
551  /* Compute residual energy */
552  energy[i << 1] = ff_dot_product(buf - j, buf - j, SUBFRAME_LEN);
553  /* Compute correlation */
554  energy[(i << 1) + 1] = ff_dot_product(buf, buf - j, SUBFRAME_LEN);
555  }
556 
557  /* Compute target energy */
558  energy[14] = ff_dot_product(buf, buf, SUBFRAME_LEN);
559 
560  /* Normalize */
561  max = 0;
562  for (i = 0; i < 15; i++)
563  max = FFMAX(max, FFABS(energy[i]));
564 
566  for (i = 0; i < 15; i++) {
567  energy[i] = av_clipl_int32((int64_t)(energy[i] << exp) +
568  (1 << 15)) >> 16;
569  }
570 
571  hf->index = -1;
572  hf->gain = 0;
573  max_ccr = 1;
574  max_eng = 0x7fff;
575 
576  for (i = 0; i <= 6; i++) {
577  eng = energy[i << 1];
578  ccr = energy[(i << 1) + 1];
579 
580  if (ccr <= 0)
581  continue;
582 
583  ccr = (ccr * ccr + (1 << 14)) >> 15;
584  diff = ccr * max_eng - eng * max_ccr;
585  if (diff > 0) {
586  max_ccr = ccr;
587  max_eng = eng;
588  hf->index = i;
589  }
590  }
591 
592  if (hf->index == -1) {
593  hf->index = pitch_lag;
594  return;
595  }
596 
597  eng = energy[14] * max_eng;
598  eng = (eng >> 2) + (eng >> 3);
599  ccr = energy[(hf->index << 1) + 1] * energy[(hf->index << 1) + 1];
600  if (eng < ccr) {
601  eng = energy[(hf->index << 1) + 1];
602 
603  if (eng >= max_eng)
604  hf->gain = 0x2800;
605  else
606  hf->gain = ((eng << 15) / max_eng * 0x2800 + (1 << 14)) >> 15;
607  }
608  hf->index += pitch_lag - 3;
609 }
610 
611 /**
612  * Apply the harmonic noise shaping filter.
613  *
614  * @param hf filter parameters
615  */
616 static void harmonic_filter(HFParam *hf, const int16_t *src, int16_t *dest)
617 {
618  int i;
619 
620  for (i = 0; i < SUBFRAME_LEN; i++) {
621  int64_t temp = hf->gain * src[i - hf->index] << 1;
622  dest[i] = av_clipl_int32((src[i] << 16) - temp + (1 << 15)) >> 16;
623  }
624 }
625 
626 static void harmonic_noise_sub(HFParam *hf, const int16_t *src, int16_t *dest)
627 {
628  int i;
629  for (i = 0; i < SUBFRAME_LEN; i++) {
630  int64_t temp = hf->gain * src[i - hf->index] << 1;
631  dest[i] = av_clipl_int32(((dest[i] - src[i]) << 16) + temp +
632  (1 << 15)) >> 16;
633  }
634 }
635 
636 /**
637  * Combined synthesis and formant perceptual weighting filer.
638  *
639  * @param qnt_lpc quantized lpc coefficients
640  * @param perf_lpc perceptual filter coefficients
641  * @param perf_fir perceptual filter fir memory
642  * @param perf_iir perceptual filter iir memory
643  * @param scale the filter output will be scaled by 2^scale
644  */
645 static void synth_percept_filter(int16_t *qnt_lpc, int16_t *perf_lpc,
646  int16_t *perf_fir, int16_t *perf_iir,
647  const int16_t *src, int16_t *dest, int scale)
648 {
649  int i, j;
650  int16_t buf_16[SUBFRAME_LEN + LPC_ORDER];
651  int64_t buf[SUBFRAME_LEN];
652 
653  int16_t *bptr_16 = buf_16 + LPC_ORDER;
654 
655  memcpy(buf_16, perf_fir, sizeof(int16_t) * LPC_ORDER);
656  memcpy(dest - LPC_ORDER, perf_iir, sizeof(int16_t) * LPC_ORDER);
657 
658  for (i = 0; i < SUBFRAME_LEN; i++) {
659  int64_t temp = 0;
660  for (j = 1; j <= LPC_ORDER; j++)
661  temp -= qnt_lpc[j - 1] * bptr_16[i - j];
662 
663  buf[i] = (src[i] << 15) + (temp << 3);
664  bptr_16[i] = av_clipl_int32(buf[i] + (1 << 15)) >> 16;
665  }
666 
667  for (i = 0; i < SUBFRAME_LEN; i++) {
668  int64_t fir = 0, iir = 0;
669  for (j = 1; j <= LPC_ORDER; j++) {
670  fir -= perf_lpc[j - 1] * bptr_16[i - j];
671  iir += perf_lpc[j + LPC_ORDER - 1] * dest[i - j];
672  }
673  dest[i] = av_clipl_int32(((buf[i] + (fir << 3)) << scale) + (iir << 3) +
674  (1 << 15)) >> 16;
675  }
676  memcpy(perf_fir, buf_16 + SUBFRAME_LEN, sizeof(int16_t) * LPC_ORDER);
677  memcpy(perf_iir, dest + SUBFRAME_LEN - LPC_ORDER,
678  sizeof(int16_t) * LPC_ORDER);
679 }
680 
681 /**
682  * Compute the adaptive codebook contribution.
683  *
684  * @param buf input signal
685  * @param index the current subframe index
686  */
687 static void acb_search(G723_1_ChannelContext *p, int16_t *residual,
688  int16_t *impulse_resp, const int16_t *buf,
689  int index)
690 {
691  int16_t flt_buf[PITCH_ORDER][SUBFRAME_LEN];
692 
693  const int16_t *cb_tbl = ff_g723_1_adaptive_cb_gain85;
694 
695  int ccr_buf[PITCH_ORDER * SUBFRAMES << 2];
696 
697  int pitch_lag = p->pitch_lag[index >> 1];
698  int acb_lag = 1;
699  int acb_gain = 0;
700  int odd_frame = index & 1;
701  int iter = 3 + odd_frame;
702  int count = 0;
703  int tbl_size = 85;
704 
705  int i, j, k, l, max;
706  int64_t temp;
707 
708  if (!odd_frame) {
709  if (pitch_lag == PITCH_MIN)
710  pitch_lag++;
711  else
712  pitch_lag = FFMIN(pitch_lag, PITCH_MAX - 5);
713  }
714 
715  for (i = 0; i < iter; i++) {
716  ff_g723_1_get_residual(residual, p->prev_excitation, pitch_lag + i - 1);
717 
718  for (j = 0; j < SUBFRAME_LEN; j++) {
719  temp = 0;
720  for (k = 0; k <= j; k++)
721  temp += residual[PITCH_ORDER - 1 + k] * impulse_resp[j - k];
722  flt_buf[PITCH_ORDER - 1][j] = av_clipl_int32((temp << 1) +
723  (1 << 15)) >> 16;
724  }
725 
726  for (j = PITCH_ORDER - 2; j >= 0; j--) {
727  flt_buf[j][0] = ((residual[j] << 13) + (1 << 14)) >> 15;
728  for (k = 1; k < SUBFRAME_LEN; k++) {
729  temp = (flt_buf[j + 1][k - 1] << 15) +
730  residual[j] * impulse_resp[k];
731  flt_buf[j][k] = av_clipl_int32((temp << 1) + (1 << 15)) >> 16;
732  }
733  }
734 
735  /* Compute crosscorrelation with the signal */
736  for (j = 0; j < PITCH_ORDER; j++) {
737  temp = ff_dot_product(buf, flt_buf[j], SUBFRAME_LEN);
738  ccr_buf[count++] = av_clipl_int32(temp << 1);
739  }
740 
741  /* Compute energies */
742  for (j = 0; j < PITCH_ORDER; j++) {
743  ccr_buf[count++] = ff_g723_1_dot_product(flt_buf[j], flt_buf[j],
744  SUBFRAME_LEN);
745  }
746 
747  for (j = 1; j < PITCH_ORDER; j++) {
748  for (k = 0; k < j; k++) {
749  temp = ff_dot_product(flt_buf[j], flt_buf[k], SUBFRAME_LEN);
750  ccr_buf[count++] = av_clipl_int32(temp << 2);
751  }
752  }
753  }
754 
755  /* Normalize and shorten */
756  max = 0;
757  for (i = 0; i < 20 * iter; i++)
758  max = FFMAX(max, FFABS(ccr_buf[i]));
759 
761 
762  for (i = 0; i < 20 * iter; i++)
763  ccr_buf[i] = av_clipl_int32((int64_t) (ccr_buf[i] << temp) +
764  (1 << 15)) >> 16;
765 
766  max = 0;
767  for (i = 0; i < iter; i++) {
768  /* Select quantization table */
769  if (!odd_frame && pitch_lag + i - 1 >= SUBFRAME_LEN - 2 ||
770  odd_frame && pitch_lag >= SUBFRAME_LEN - 2) {
772  tbl_size = 170;
773  }
774 
775  for (j = 0, k = 0; j < tbl_size; j++, k += 20) {
776  temp = 0;
777  for (l = 0; l < 20; l++)
778  temp += ccr_buf[20 * i + l] * cb_tbl[k + l];
780 
781  if (temp > max) {
782  max = temp;
783  acb_gain = j;
784  acb_lag = i;
785  }
786  }
787  }
788 
789  if (!odd_frame) {
790  pitch_lag += acb_lag - 1;
791  acb_lag = 1;
792  }
793 
794  p->pitch_lag[index >> 1] = pitch_lag;
795  p->subframe[index].ad_cb_lag = acb_lag;
796  p->subframe[index].ad_cb_gain = acb_gain;
797 }
798 
799 /**
800  * Subtract the adaptive codebook contribution from the input
801  * to obtain the residual.
802  *
803  * @param buf target vector
804  */
805 static void sub_acb_contrib(const int16_t *residual, const int16_t *impulse_resp,
806  int16_t *buf)
807 {
808  int i, j;
809  /* Subtract adaptive CB contribution to obtain the residual */
810  for (i = 0; i < SUBFRAME_LEN; i++) {
811  int64_t temp = buf[i] << 14;
812  for (j = 0; j <= i; j++)
813  temp -= residual[j] * impulse_resp[i - j];
814 
815  buf[i] = av_clipl_int32((temp << 2) + (1 << 15)) >> 16;
816  }
817 }
818 
819 /**
820  * Quantize the residual signal using the fixed codebook (MP-MLQ).
821  *
822  * @param optim optimized fixed codebook parameters
823  * @param buf excitation vector
824  */
825 static void get_fcb_param(FCBParam *optim, int16_t *impulse_resp,
826  int16_t *buf, int pulse_cnt, int pitch_lag)
827 {
828  FCBParam param;
829  int16_t impulse_r[SUBFRAME_LEN];
830  int16_t temp_corr[SUBFRAME_LEN];
831  int16_t impulse_corr[SUBFRAME_LEN];
832 
833  int ccr1[SUBFRAME_LEN];
834  int ccr2[SUBFRAME_LEN];
835  int amp, err, max, max_amp_index, min, scale, i, j, k, l;
836 
837  int64_t temp;
838 
839  /* Update impulse response */
840  memcpy(impulse_r, impulse_resp, sizeof(int16_t) * SUBFRAME_LEN);
841  param.dirac_train = 0;
842  if (pitch_lag < SUBFRAME_LEN - 2) {
843  param.dirac_train = 1;
844  ff_g723_1_gen_dirac_train(impulse_r, pitch_lag);
845  }
846 
847  for (i = 0; i < SUBFRAME_LEN; i++)
848  temp_corr[i] = impulse_r[i] >> 1;
849 
850  /* Compute impulse response autocorrelation */
851  temp = ff_g723_1_dot_product(temp_corr, temp_corr, SUBFRAME_LEN);
852 
853  scale = ff_g723_1_normalize_bits(temp, 31);
854  impulse_corr[0] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
855 
856  for (i = 1; i < SUBFRAME_LEN; i++) {
857  temp = ff_g723_1_dot_product(temp_corr + i, temp_corr,
858  SUBFRAME_LEN - i);
859  impulse_corr[i] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
860  }
861 
862  /* Compute crosscorrelation of impulse response with residual signal */
863  scale -= 4;
864  for (i = 0; i < SUBFRAME_LEN; i++) {
865  temp = ff_g723_1_dot_product(buf + i, impulse_r, SUBFRAME_LEN - i);
866  if (scale < 0)
867  ccr1[i] = temp >> -scale;
868  else
869  ccr1[i] = av_clipl_int32(temp << scale);
870  }
871 
872  /* Search loop */
873  for (i = 0; i < GRID_SIZE; i++) {
874  /* Maximize the crosscorrelation */
875  max = 0;
876  for (j = i; j < SUBFRAME_LEN; j += GRID_SIZE) {
877  temp = FFABS(ccr1[j]);
878  if (temp >= max) {
879  max = temp;
880  param.pulse_pos[0] = j;
881  }
882  }
883 
884  /* Quantize the gain (max crosscorrelation/impulse_corr[0]) */
885  amp = max;
886  min = 1 << 30;
887  max_amp_index = GAIN_LEVELS - 2;
888  for (j = max_amp_index; j >= 2; j--) {
890  impulse_corr[0] << 1);
891  temp = FFABS(temp - amp);
892  if (temp < min) {
893  min = temp;
894  max_amp_index = j;
895  }
896  }
897 
898  max_amp_index--;
899  /* Select additional gain values */
900  for (j = 1; j < 5; j++) {
901  for (k = i; k < SUBFRAME_LEN; k += GRID_SIZE) {
902  temp_corr[k] = 0;
903  ccr2[k] = ccr1[k];
904  }
905  param.amp_index = max_amp_index + j - 2;
906  amp = ff_g723_1_fixed_cb_gain[param.amp_index];
907 
908  param.pulse_sign[0] = (ccr2[param.pulse_pos[0]] < 0) ? -amp : amp;
909  temp_corr[param.pulse_pos[0]] = 1;
910 
911  for (k = 1; k < pulse_cnt; k++) {
912  max = INT_MIN;
913  for (l = i; l < SUBFRAME_LEN; l += GRID_SIZE) {
914  if (temp_corr[l])
915  continue;
916  temp = impulse_corr[FFABS(l - param.pulse_pos[k - 1])];
918  param.pulse_sign[k - 1] << 1);
919  ccr2[l] -= temp;
920  temp = FFABS(ccr2[l]);
921  if (temp > max) {
922  max = temp;
923  param.pulse_pos[k] = l;
924  }
925  }
926 
927  param.pulse_sign[k] = (ccr2[param.pulse_pos[k]] < 0) ?
928  -amp : amp;
929  temp_corr[param.pulse_pos[k]] = 1;
930  }
931 
932  /* Create the error vector */
933  memset(temp_corr, 0, sizeof(int16_t) * SUBFRAME_LEN);
934 
935  for (k = 0; k < pulse_cnt; k++)
936  temp_corr[param.pulse_pos[k]] = param.pulse_sign[k];
937 
938  for (k = SUBFRAME_LEN - 1; k >= 0; k--) {
939  temp = 0;
940  for (l = 0; l <= k; l++) {
941  int prod = av_clipl_int32((int64_t) temp_corr[l] *
942  impulse_r[k - l] << 1);
943  temp = av_clipl_int32(temp + prod);
944  }
945  temp_corr[k] = temp << 2 >> 16;
946  }
947 
948  /* Compute square of error */
949  err = 0;
950  for (k = 0; k < SUBFRAME_LEN; k++) {
951  int64_t prod;
952  prod = av_clipl_int32((int64_t) buf[k] * temp_corr[k] << 1);
953  err = av_clipl_int32(err - prod);
954  prod = av_clipl_int32((int64_t) temp_corr[k] * temp_corr[k]);
955  err = av_clipl_int32(err + prod);
956  }
957 
958  /* Minimize */
959  if (err < optim->min_err) {
960  optim->min_err = err;
961  optim->grid_index = i;
962  optim->amp_index = param.amp_index;
963  optim->dirac_train = param.dirac_train;
964 
965  for (k = 0; k < pulse_cnt; k++) {
966  optim->pulse_sign[k] = param.pulse_sign[k];
967  optim->pulse_pos[k] = param.pulse_pos[k];
968  }
969  }
970  }
971  }
972 }
973 
974 /**
975  * Encode the pulse position and gain of the current subframe.
976  *
977  * @param optim optimized fixed CB parameters
978  * @param buf excitation vector
979  */
980 static void pack_fcb_param(G723_1_Subframe *subfrm, FCBParam *optim,
981  int16_t *buf, int pulse_cnt)
982 {
983  int i, j;
984 
985  j = PULSE_MAX - pulse_cnt;
986 
987  subfrm->pulse_sign = 0;
988  subfrm->pulse_pos = 0;
989 
990  for (i = 0; i < SUBFRAME_LEN >> 1; i++) {
991  int val = buf[optim->grid_index + (i << 1)];
992  if (!val) {
994  } else {
995  subfrm->pulse_sign <<= 1;
996  if (val < 0)
997  subfrm->pulse_sign++;
998  j++;
999 
1000  if (j == PULSE_MAX)
1001  break;
1002  }
1003  }
1004  subfrm->amp_index = optim->amp_index;
1005  subfrm->grid_index = optim->grid_index;
1006  subfrm->dirac_train = optim->dirac_train;
1007 }
1008 
1009 /**
1010  * Compute the fixed codebook excitation.
1011  *
1012  * @param buf target vector
1013  * @param impulse_resp impulse response of the combined filter
1014  */
1015 static void fcb_search(G723_1_ChannelContext *p, int16_t *impulse_resp,
1016  int16_t *buf, int index)
1017 {
1018  FCBParam optim;
1019  int pulse_cnt = pulses[index];
1020  int i;
1021 
1022  optim.min_err = 1 << 30;
1023  get_fcb_param(&optim, impulse_resp, buf, pulse_cnt, SUBFRAME_LEN);
1024 
1025  if (p->pitch_lag[index >> 1] < SUBFRAME_LEN - 2) {
1026  get_fcb_param(&optim, impulse_resp, buf, pulse_cnt,
1027  p->pitch_lag[index >> 1]);
1028  }
1029 
1030  /* Reconstruct the excitation */
1031  memset(buf, 0, sizeof(int16_t) * SUBFRAME_LEN);
1032  for (i = 0; i < pulse_cnt; i++)
1033  buf[optim.pulse_pos[i]] = optim.pulse_sign[i];
1034 
1035  pack_fcb_param(&p->subframe[index], &optim, buf, pulse_cnt);
1036 
1037  if (optim.dirac_train)
1038  ff_g723_1_gen_dirac_train(buf, p->pitch_lag[index >> 1]);
1039 }
1040 
1041 /**
1042  * Pack the frame parameters into output bitstream.
1043  *
1044  * @param frame output buffer
1045  * @param size size of the buffer
1046  */
1048 {
1049  PutBitContext pb;
1050  int info_bits = 0;
1051  int i, temp;
1052 
1053  init_put_bits(&pb, avpkt->data, avpkt->size);
1054 
1055  put_bits(&pb, 2, info_bits);
1056 
1057  put_bits(&pb, 8, p->lsp_index[2]);
1058  put_bits(&pb, 8, p->lsp_index[1]);
1059  put_bits(&pb, 8, p->lsp_index[0]);
1060 
1061  put_bits(&pb, 7, p->pitch_lag[0] - PITCH_MIN);
1062  put_bits(&pb, 2, p->subframe[1].ad_cb_lag);
1063  put_bits(&pb, 7, p->pitch_lag[1] - PITCH_MIN);
1064  put_bits(&pb, 2, p->subframe[3].ad_cb_lag);
1065 
1066  /* Write 12 bit combined gain */
1067  for (i = 0; i < SUBFRAMES; i++) {
1069  p->subframe[i].amp_index;
1070  if (p->cur_rate == RATE_6300)
1071  temp += p->subframe[i].dirac_train << 11;
1072  put_bits(&pb, 12, temp);
1073  }
1074 
1075  put_bits(&pb, 1, p->subframe[0].grid_index);
1076  put_bits(&pb, 1, p->subframe[1].grid_index);
1077  put_bits(&pb, 1, p->subframe[2].grid_index);
1078  put_bits(&pb, 1, p->subframe[3].grid_index);
1079 
1080  if (p->cur_rate == RATE_6300) {
1081  put_bits(&pb, 1, 0); /* reserved bit */
1082 
1083  /* Write 13 bit combined position index */
1084  temp = (p->subframe[0].pulse_pos >> 16) * 810 +
1085  (p->subframe[1].pulse_pos >> 14) * 90 +
1086  (p->subframe[2].pulse_pos >> 16) * 9 +
1087  (p->subframe[3].pulse_pos >> 14);
1088  put_bits(&pb, 13, temp);
1089 
1090  put_bits(&pb, 16, p->subframe[0].pulse_pos & 0xffff);
1091  put_bits(&pb, 14, p->subframe[1].pulse_pos & 0x3fff);
1092  put_bits(&pb, 16, p->subframe[2].pulse_pos & 0xffff);
1093  put_bits(&pb, 14, p->subframe[3].pulse_pos & 0x3fff);
1094 
1095  put_bits(&pb, 6, p->subframe[0].pulse_sign);
1096  put_bits(&pb, 5, p->subframe[1].pulse_sign);
1097  put_bits(&pb, 6, p->subframe[2].pulse_sign);
1098  put_bits(&pb, 5, p->subframe[3].pulse_sign);
1099  }
1100 
1101  flush_put_bits(&pb);
1102  return frame_size[info_bits];
1103 }
1104 
1105 static int g723_1_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
1106  const AVFrame *frame, int *got_packet_ptr)
1107 {
1108  G723_1_Context *s = avctx->priv_data;
1109  G723_1_ChannelContext *p = &s->ch[0];
1110  int16_t unq_lpc[LPC_ORDER * SUBFRAMES];
1111  int16_t qnt_lpc[LPC_ORDER * SUBFRAMES];
1112  int16_t cur_lsp[LPC_ORDER];
1113  int16_t weighted_lpc[LPC_ORDER * SUBFRAMES << 1];
1114  int16_t vector[FRAME_LEN + PITCH_MAX];
1115  int offset, ret, i, j;
1116  int16_t *in, *start;
1117  HFParam hf[4];
1118 
1119  /* duplicate input */
1120  start = in = av_malloc(frame->nb_samples * sizeof(int16_t));
1121  if (!in)
1122  return AVERROR(ENOMEM);
1123  memcpy(in, frame->data[0], frame->nb_samples * sizeof(int16_t));
1124 
1126 
1127  memcpy(vector, p->prev_data, HALF_FRAME_LEN * sizeof(int16_t));
1128  memcpy(vector + HALF_FRAME_LEN, in, FRAME_LEN * sizeof(int16_t));
1129 
1130  comp_lpc_coeff(vector, unq_lpc);
1131  lpc2lsp(&unq_lpc[LPC_ORDER * 3], p->prev_lsp, cur_lsp);
1132  lsp_quantize(p->lsp_index, cur_lsp, p->prev_lsp);
1133 
1134  /* Update memory */
1135  memcpy(vector + LPC_ORDER, p->prev_data + SUBFRAME_LEN,
1136  sizeof(int16_t) * SUBFRAME_LEN);
1137  memcpy(vector + LPC_ORDER + SUBFRAME_LEN, in,
1138  sizeof(int16_t) * (HALF_FRAME_LEN + SUBFRAME_LEN));
1139  memcpy(p->prev_data, in + HALF_FRAME_LEN,
1140  sizeof(int16_t) * HALF_FRAME_LEN);
1141  memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
1142 
1143  perceptual_filter(p, weighted_lpc, unq_lpc, vector);
1144 
1145  memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
1146  memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
1147  memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
1148 
1149  ff_g723_1_scale_vector(vector, vector, FRAME_LEN + PITCH_MAX);
1150 
1151  p->pitch_lag[0] = estimate_pitch(vector, PITCH_MAX);
1152  p->pitch_lag[1] = estimate_pitch(vector, PITCH_MAX + HALF_FRAME_LEN);
1153 
1154  for (i = PITCH_MAX, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1155  comp_harmonic_coeff(vector + i, p->pitch_lag[j >> 1], hf + j);
1156 
1157  memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
1158  memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
1159  memcpy(p->prev_weight_sig, vector + FRAME_LEN, sizeof(int16_t) * PITCH_MAX);
1160 
1161  for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1162  harmonic_filter(hf + j, vector + PITCH_MAX + i, in + i);
1163 
1164  ff_g723_1_inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, 0);
1165  ff_g723_1_lsp_interpolate(qnt_lpc, cur_lsp, p->prev_lsp);
1166 
1167  memcpy(p->prev_lsp, cur_lsp, sizeof(int16_t) * LPC_ORDER);
1168 
1169  offset = 0;
1170  for (i = 0; i < SUBFRAMES; i++) {
1171  int16_t impulse_resp[SUBFRAME_LEN];
1172  int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
1173  int16_t flt_in[SUBFRAME_LEN];
1174  int16_t zero[LPC_ORDER], fir[LPC_ORDER], iir[LPC_ORDER];
1175 
1176  /**
1177  * Compute the combined impulse response of the synthesis filter,
1178  * formant perceptual weighting filter and harmonic noise shaping filter
1179  */
1180  memset(zero, 0, sizeof(int16_t) * LPC_ORDER);
1181  memset(vector, 0, sizeof(int16_t) * PITCH_MAX);
1182  memset(flt_in, 0, sizeof(int16_t) * SUBFRAME_LEN);
1183 
1184  flt_in[0] = 1 << 13; /* Unit impulse */
1185  synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
1186  zero, zero, flt_in, vector + PITCH_MAX, 1);
1187  harmonic_filter(hf + i, vector + PITCH_MAX, impulse_resp);
1188 
1189  /* Compute the combined zero input response */
1190  flt_in[0] = 0;
1191  memcpy(fir, p->perf_fir_mem, sizeof(int16_t) * LPC_ORDER);
1192  memcpy(iir, p->perf_iir_mem, sizeof(int16_t) * LPC_ORDER);
1193 
1194  synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
1195  fir, iir, flt_in, vector + PITCH_MAX, 0);
1196  memcpy(vector, p->harmonic_mem, sizeof(int16_t) * PITCH_MAX);
1197  harmonic_noise_sub(hf + i, vector + PITCH_MAX, in);
1198 
1199  acb_search(p, residual, impulse_resp, in, i);
1201  p->pitch_lag[i >> 1], &p->subframe[i],
1202  p->cur_rate);
1203  sub_acb_contrib(residual, impulse_resp, in);
1204 
1205  fcb_search(p, impulse_resp, in, i);
1206 
1207  /* Reconstruct the excitation */
1209  p->pitch_lag[i >> 1], &p->subframe[i],
1210  RATE_6300);
1211 
1212  memmove(p->prev_excitation, p->prev_excitation + SUBFRAME_LEN,
1213  sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
1214  for (j = 0; j < SUBFRAME_LEN; j++)
1215  in[j] = av_clip_int16((in[j] << 1) + impulse_resp[j]);
1216  memcpy(p->prev_excitation + PITCH_MAX - SUBFRAME_LEN, in,
1217  sizeof(int16_t) * SUBFRAME_LEN);
1218 
1219  /* Update filter memories */
1220  synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
1221  p->perf_fir_mem, p->perf_iir_mem,
1222  in, vector + PITCH_MAX, 0);
1223  memmove(p->harmonic_mem, p->harmonic_mem + SUBFRAME_LEN,
1224  sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
1225  memcpy(p->harmonic_mem + PITCH_MAX - SUBFRAME_LEN, vector + PITCH_MAX,
1226  sizeof(int16_t) * SUBFRAME_LEN);
1227 
1228  in += SUBFRAME_LEN;
1229  offset += LPC_ORDER;
1230  }
1231 
1232  av_free(start);
1233 
1234  if ((ret = ff_alloc_packet2(avctx, avpkt, 24, 0)) < 0)
1235  return ret;
1236 
1237  *got_packet_ptr = 1;
1238  avpkt->size = pack_bitstream(p, avpkt);
1239  return 0;
1240 }
1241 
1242 static const AVCodecDefault defaults[] = {
1243  { "b", "6300" },
1244  { NULL },
1245 };
1246 
1248  .name = "g723_1",
1249  .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1250  .type = AVMEDIA_TYPE_AUDIO,
1251  .id = AV_CODEC_ID_G723_1,
1252  .priv_data_size = sizeof(G723_1_Context),
1254  .encode2 = g723_1_encode_frame,
1255  .defaults = defaults,
1256  .sample_fmts = (const enum AVSampleFormat[]) {
1258  },
1259 };
static enum AVSampleFormat sample_fmts[]
Definition: adpcmenc.c:925
static double val(void *priv, double ch)
Definition: aeval.c:76
static av_always_inline void update(SilenceDetectContext *s, AVFrame *insamples, int is_silence, int current_sample, int64_t nb_samples_notify, AVRational time_base)
#define av_cold
Definition: attributes.h:88
uint8_t pi<< 24) CONV_FUNC_GROUP(AV_SAMPLE_FMT_FLT, float, AV_SAMPLE_FMT_U8, uint8_t,(*(const uint8_t *) pi - 0x80) *(1.0f/(1<< 7))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_DBL, double, AV_SAMPLE_FMT_U8, uint8_t,(*(const uint8_t *) pi - 0x80) *(1.0/(1<< 7))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_U8, uint8_t, AV_SAMPLE_FMT_S16, int16_t,(*(const int16_t *) pi >> 8)+0x80) CONV_FUNC_GROUP(AV_SAMPLE_FMT_FLT, float, AV_SAMPLE_FMT_S16, int16_t, *(const int16_t *) pi *(1.0f/(1<< 15))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_DBL, double, AV_SAMPLE_FMT_S16, int16_t, *(const int16_t *) pi *(1.0/(1<< 15))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_U8, uint8_t, AV_SAMPLE_FMT_S32, int32_t,(*(const int32_t *) pi >> 24)+0x80) CONV_FUNC_GROUP(AV_SAMPLE_FMT_FLT, float, AV_SAMPLE_FMT_S32, int32_t, *(const int32_t *) pi *(1.0f/(1U<< 31))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_DBL, double, AV_SAMPLE_FMT_S32, int32_t, *(const int32_t *) pi *(1.0/(1U<< 31))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_U8, uint8_t, AV_SAMPLE_FMT_FLT, float, av_clip_uint8(lrintf(*(const float *) pi *(1<< 7))+0x80)) CONV_FUNC_GROUP(AV_SAMPLE_FMT_S16, int16_t, AV_SAMPLE_FMT_FLT, float, av_clip_int16(lrintf(*(const float *) pi *(1<< 15)))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_S32, int32_t, AV_SAMPLE_FMT_FLT, float, av_clipl_int32(llrintf(*(const float *) pi *(1U<< 31)))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_U8, uint8_t, AV_SAMPLE_FMT_DBL, double, av_clip_uint8(lrint(*(const double *) pi *(1<< 7))+0x80)) CONV_FUNC_GROUP(AV_SAMPLE_FMT_S16, int16_t, AV_SAMPLE_FMT_DBL, double, av_clip_int16(lrint(*(const double *) pi *(1<< 15)))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_S32, int32_t, AV_SAMPLE_FMT_DBL, double, av_clipl_int32(llrint(*(const double *) pi *(1U<< 31)))) #define SET_CONV_FUNC_GROUP(ofmt, ifmt) static void set_generic_function(AudioConvert *ac) { } void ff_audio_convert_free(AudioConvert **ac) { if(! *ac) return;ff_dither_free(&(*ac) ->dc);av_freep(ac);} AudioConvert *ff_audio_convert_alloc(AVAudioResampleContext *avr, enum AVSampleFormat out_fmt, enum AVSampleFormat in_fmt, int channels, int sample_rate, int apply_map) { AudioConvert *ac;int in_planar, out_planar;ac=av_mallocz(sizeof(*ac));if(!ac) return NULL;ac->avr=avr;ac->out_fmt=out_fmt;ac->in_fmt=in_fmt;ac->channels=channels;ac->apply_map=apply_map;if(avr->dither_method !=AV_RESAMPLE_DITHER_NONE &&av_get_packed_sample_fmt(out_fmt)==AV_SAMPLE_FMT_S16 &&av_get_bytes_per_sample(in_fmt) > 2) { ac->dc=ff_dither_alloc(avr, out_fmt, in_fmt, channels, sample_rate, apply_map);if(!ac->dc) { av_free(ac);return NULL;} return ac;} in_planar=ff_sample_fmt_is_planar(in_fmt, channels);out_planar=ff_sample_fmt_is_planar(out_fmt, channels);if(in_planar==out_planar) { ac->func_type=CONV_FUNC_TYPE_FLAT;ac->planes=in_planar ? ac->channels :1;} else if(in_planar) ac->func_type=CONV_FUNC_TYPE_INTERLEAVE;else ac->func_type=CONV_FUNC_TYPE_DEINTERLEAVE;set_generic_function(ac);if(ARCH_AARCH64) ff_audio_convert_init_aarch64(ac);if(ARCH_ARM) ff_audio_convert_init_arm(ac);if(ARCH_X86) ff_audio_convert_init_x86(ac);return ac;} int ff_audio_convert(AudioConvert *ac, AudioData *out, AudioData *in) { int use_generic=1;int len=in->nb_samples;int p;if(ac->dc) { av_log(ac->avr, AV_LOG_TRACE, "%d samples - audio_convert: %s to %s (dithered)\n", len, av_get_sample_fmt_name(ac->in_fmt), av_get_sample_fmt_name(ac->out_fmt));return ff_convert_dither(ac-> in
uint8_t
Libavcodec external API header.
static av_cold int init(AVCodecContext *avctx)
Definition: avrndec.c:31
#define s(width, name)
Definition: cbs_vp9.c:257
#define f(width, name)
Definition: cbs_vp9.c:255
int64_t ff_dot_product(const int16_t *a, const int16_t *b, int length)
Calculate the dot product of 2 int16_t vectors.
Definition: celp_math.c:100
static av_always_inline void filter(int16_t *output, ptrdiff_t out_stride, const int16_t *low, ptrdiff_t low_stride, const int16_t *high, ptrdiff_t high_stride, int len, int clip)
Definition: cfhddsp.c:27
audio channel layout utility functions
common internal and external API header
#define FFMIN(a, b)
Definition: common.h:105
#define av_clipl_int32
Definition: common.h:140
#define av_clip_int16
Definition: common.h:137
#define FFMAX(a, b)
Definition: common.h:103
#define FFABS(a)
Absolute value, Note, INT_MIN / INT64_MIN result in undefined behavior as they are not representable ...
Definition: common.h:72
#define NULL
Definition: coverity.c:32
long long int64_t
Definition: coverity.c:34
#define max(a, b)
Definition: cuda_runtime.h:33
#define SUBFRAMES
Definition: dcaenc.c:51
static AVFrame * frame
#define PULSE_MAX
Definition: dss_sp.c:33
int ff_alloc_packet2(AVCodecContext *avctx, AVPacket *avpkt, int64_t size, int64_t min_size)
Check AVPacket size and/or allocate data.
Definition: encode.c:33
int8_t exp
Definition: eval.c:72
G.723.1 types, functions and data tables.
#define GRID_SIZE
Definition: g723_1.h:46
@ RATE_6300
Definition: g723_1.h:73
#define PITCH_ORDER
Definition: g723_1.h:45
#define G723_1_COS_TAB_FIRST_ELEMENT
Definition: g723_1.h:242
#define MULL2(a, b)
Bitexact implementation of 2ab scaled by 1/2^16.
Definition: g723_1.h:57
#define LPC_FRAME
Definition: g723_1.h:39
#define PITCH_MAX
Definition: g723_1.h:44
static const int16_t dc_lsp[LPC_ORDER]
LSP DC component.
Definition: g723_1.h:227
#define FRAME_LEN
Definition: g723_1.h:37
#define HALF_FRAME_LEN
Definition: g723_1.h:38
#define SUBFRAME_LEN
Definition: g723_1.h:36
#define LPC_ORDER
Definition: g723_1.h:40
#define PITCH_MIN
Definition: g723_1.h:43
static const int8_t pulses[4]
Number of non-zero pulses in the MP-MLQ excitation.
Definition: g723_1.h:260
#define COS_TBL_SIZE
Definition: g723_1.h:49
#define GAIN_LEVELS
Definition: g723_1.h:48
static void sub_acb_contrib(const int16_t *residual, const int16_t *impulse_resp, int16_t *buf)
Subtract the adaptive codebook contribution from the input to obtain the residual.
Definition: g723_1enc.c:805
static const AVCodecDefault defaults[]
Definition: g723_1enc.c:1242
static void iir_filter(int16_t *fir_coef, int16_t *iir_coef, int16_t *src, int16_t *dest)
Perform IIR filtering.
Definition: g723_1enc.c:415
static av_cold int g723_1_encode_init(AVCodecContext *avctx)
Definition: g723_1enc.c:91
static int estimate_pitch(int16_t *buf, int start)
Estimate the open loop pitch period.
Definition: g723_1enc.c:470
static int pack_bitstream(G723_1_ChannelContext *p, AVPacket *avpkt)
Pack the frame parameters into output bitstream.
Definition: g723_1enc.c:1047
static const int16_t binomial_window[LPC_ORDER]
Binomial window coefficients scaled by 2^15.
Definition: g723_1enc.c:70
static void fcb_search(G723_1_ChannelContext *p, int16_t *impulse_resp, int16_t *buf, int index)
Compute the fixed codebook excitation.
Definition: g723_1enc.c:1015
static void levinson_durbin(int16_t *lpc, int16_t *autocorr, int16_t error)
Use Levinson-Durbin recursion to compute LPC coefficients from autocorrelation values.
Definition: g723_1enc.c:187
AVCodec ff_g723_1_encoder
Definition: g723_1enc.c:1247
static void comp_autocorr(int16_t *buf, int16_t *autocorr)
Estimate autocorrelation of the input vector.
Definition: g723_1enc.c:145
static void highpass_filter(int16_t *buf, int16_t *fir, int *iir)
Remove DC component from the input signal.
Definition: g723_1enc.c:129
static void lsp_quantize(uint8_t *lsp_index, int16_t *lsp, int16_t *prev_lsp)
Vector quantize the LSP frequencies.
Definition: g723_1enc.c:367
static void acb_search(G723_1_ChannelContext *p, int16_t *residual, int16_t *impulse_resp, const int16_t *buf, int index)
Compute the adaptive codebook contribution.
Definition: g723_1enc.c:687
static const int16_t percept_flt_tbl[2][LPC_ORDER]
0.5^i scaled by 2^15
Definition: g723_1enc.c:84
static void get_fcb_param(FCBParam *optim, int16_t *impulse_resp, int16_t *buf, int pulse_cnt, int pitch_lag)
Quantize the residual signal using the fixed codebook (MP-MLQ).
Definition: g723_1enc.c:825
static const int16_t hamming_window[LPC_FRAME]
Hamming window coefficients scaled by 2^15.
Definition: g723_1enc.c:46
static void synth_percept_filter(int16_t *qnt_lpc, int16_t *perf_lpc, int16_t *perf_fir, int16_t *perf_iir, const int16_t *src, int16_t *dest, int scale)
Combined synthesis and formant perceptual weighting filer.
Definition: g723_1enc.c:645
static const int16_t bandwidth_expand[LPC_ORDER]
0.994^i scaled by 2^15
Definition: g723_1enc.c:77
#define get_index(num, offset, size)
Quantize the current LSP subvector.
Definition: g723_1enc.c:341
static void comp_lpc_coeff(int16_t *buf, int16_t *lpc)
Calculate LPC coefficients for the current frame.
Definition: g723_1enc.c:231
static void harmonic_noise_sub(HFParam *hf, const int16_t *src, int16_t *dest)
Definition: g723_1enc.c:626
static void perceptual_filter(G723_1_ChannelContext *p, int16_t *flt_coef, int16_t *unq_lpc, int16_t *buf)
Apply the formant perceptual weighting filter.
Definition: g723_1enc.c:438
static void lpc2lsp(int16_t *lpc, int16_t *prev_lsp, int16_t *lsp)
Definition: g723_1enc.c:247
static void comp_harmonic_coeff(int16_t *buf, int16_t pitch_lag, HFParam *hf)
Compute harmonic noise filter parameters.
Definition: g723_1enc.c:543
static void harmonic_filter(HFParam *hf, const int16_t *src, int16_t *dest)
Apply the harmonic noise shaping filter.
Definition: g723_1enc.c:616
static int g723_1_encode_frame(AVCodecContext *avctx, AVPacket *avpkt, const AVFrame *frame, int *got_packet_ptr)
Definition: g723_1enc.c:1105
static void pack_fcb_param(G723_1_Subframe *subfrm, FCBParam *optim, int16_t *buf, int pulse_cnt)
Encode the pulse position and gain of the current subframe.
Definition: g723_1enc.c:980
@ AV_CODEC_ID_G723_1
Definition: codec_id.h:476
#define AVERROR_PATCHWELCOME
Not yet implemented in FFmpeg, patches welcome.
Definition: error.h:62
#define AVERROR(e)
Definition: error.h:43
#define AV_LOG_ERROR
Something went wrong and cannot losslessly be recovered.
Definition: log.h:194
@ AVMEDIA_TYPE_AUDIO
Definition: avutil.h:202
AVSampleFormat
Audio sample formats.
Definition: samplefmt.h:58
@ AV_SAMPLE_FMT_NONE
Definition: samplefmt.h:59
@ AV_SAMPLE_FMT_S16
signed 16 bits
Definition: samplefmt.h:61
int index
Definition: gxfenc.c:89
for(j=16;j >0;--j)
int i
Definition: input.c:407
static void put_bits(Jpeg2000EncoderContext *s, int val, int n)
put n times val bit
Definition: j2kenc.c:218
static int weight(int i, int blen, int offset)
Definition: diracdec.c:1561
const int16_t ff_g723_1_cos_tab[COS_TBL_SIZE+1]
Definition: g723_1.c:32
int ff_g723_1_dot_product(const int16_t *a, const int16_t *b, int length)
Definition: g723_1.c:1125
void ff_g723_1_lsp_interpolate(int16_t *lpc, int16_t *cur_lsp, int16_t *prev_lsp)
Quantize LSP frequencies by interpolation and convert them to the corresponding LPC coefficients.
Definition: g723_1.c:1251
int ff_g723_1_normalize_bits(int num, int width)
Calculate the number of left-shifts required for normalizing the input.
Definition: g723_1.c:1120
void ff_g723_1_get_residual(int16_t *residual, int16_t *prev_excitation, int lag)
Get delayed contribution from the previous excitation vector.
Definition: g723_1.c:1131
void ff_g723_1_gen_acb_excitation(int16_t *vector, int16_t *prev_excitation, int pitch_lag, G723_1_Subframe *subfrm, enum Rate cur_rate)
Generate adaptive codebook excitation.
Definition: g723_1.c:1157
void ff_g723_1_gen_dirac_train(int16_t *buf, int pitch_lag)
Generate a train of dirac functions with period as pitch lag.
Definition: g723_1.c:1145
void ff_g723_1_inverse_quant(int16_t *cur_lsp, int16_t *prev_lsp, uint8_t *lsp_index, int bad_frame)
Perform inverse quantization of LSP frequencies.
Definition: g723_1.c:1272
int ff_g723_1_scale_vector(int16_t *dst, const int16_t *vector, int length)
Scale vector contents based on the largest of their absolutes.
Definition: g723_1.c:1103
const int32_t ff_g723_1_combinatorial_table[PULSE_MAX][SUBFRAME_LEN/GRID_SIZE]
Used for the coding/decoding of the pulses positions for the MP-MLQ codebook.
Definition: g723_1.c:409
const int16_t ff_g723_1_fixed_cb_gain[GAIN_LEVELS]
Definition: g723_1.c:453
const int16_t ff_g723_1_adaptive_cb_gain85[85 *20]
Definition: g723_1.c:459
const int16_t ff_g723_1_adaptive_cb_gain170[170 *20]
Definition: g723_1.c:675
common internal API header
#define NULL_IF_CONFIG_SMALL(x)
Return NULL if CONFIG_SMALL is true, otherwise the argument without modification.
Definition: internal.h:117
void avpriv_report_missing_feature(void *avc, const char *msg,...) av_printf_format(2
Log a generic warning message about a missing feature.
Memory handling functions.
int frame_size
Definition: mxfenc.c:2206
AVOptions.
bitstream writer API
static void init_put_bits(PutBitContext *s, uint8_t *buffer, int buffer_size)
Initialize the PutBitContext s.
Definition: put_bits.h:57
static void flush_put_bits(PutBitContext *s)
Pad the end of the output stream with zeros.
Definition: put_bits.h:110
#define zero
Definition: regdef.h:64
static int shift(int a, int b)
Definition: sonic.c:82
main external API structure.
Definition: avcodec.h:536
int64_t bit_rate
the average bitrate
Definition: avcodec.h:586
int sample_rate
samples per second
Definition: avcodec.h:1196
int channels
number of audio channels
Definition: avcodec.h:1197
int frame_size
Number of samples per channel in an audio frame.
Definition: avcodec.h:1216
void * priv_data
Definition: avcodec.h:563
AVCodec.
Definition: codec.h:197
const char * name
Name of the codec implementation.
Definition: codec.h:204
This structure describes decoded (raw) audio or video data.
Definition: frame.h:318
int nb_samples
number of audio samples (per channel) described by this frame
Definition: frame.h:384
uint8_t * data[AV_NUM_DATA_POINTERS]
pointer to the picture/channel planes.
Definition: frame.h:332
This structure stores compressed data.
Definition: packet.h:346
int size
Definition: packet.h:370
uint8_t * data
Definition: packet.h:369
Optimized fixed codebook excitation parameters.
Definition: g723_1.h:110
int min_err
Definition: g723_1.h:111
int dirac_train
Definition: g723_1.h:114
int pulse_sign[PULSE_MAX]
Definition: g723_1.h:116
int amp_index
Definition: g723_1.h:112
int grid_index
Definition: g723_1.h:113
int pulse_pos[PULSE_MAX]
Definition: g723_1.h:115
int16_t harmonic_mem[PITCH_MAX]
Definition: g723_1.h:156
int16_t prev_weight_sig[PITCH_MAX]
Definition: g723_1.h:149
int16_t perf_iir_mem[LPC_ORDER]
and iir memories
Definition: g723_1.h:154
int16_t hpf_fir_mem
highpass filter fir
Definition: g723_1.h:151
int16_t prev_data[HALF_FRAME_LEN]
Definition: g723_1.h:148
int iir_mem[LPC_ORDER]
Definition: g723_1.h:134
int16_t prev_lsp[LPC_ORDER]
Definition: g723_1.h:128
uint8_t lsp_index[LSP_BANDS]
Definition: g723_1.h:124
int hpf_iir_mem
and iir memories
Definition: g723_1.h:152
enum Rate cur_rate
Definition: g723_1.h:123
int16_t prev_excitation[PITCH_MAX]
Definition: g723_1.h:130
int16_t perf_fir_mem[LPC_ORDER]
perceptual filter fir
Definition: g723_1.h:153
int16_t fir_mem[LPC_ORDER]
Definition: g723_1.h:133
G723_1_Subframe subframe[4]
Definition: g723_1.h:120
G723.1 unpacked data subframe.
Definition: g723_1.h:80
int ad_cb_gain
Definition: g723_1.h:82
int amp_index
Definition: g723_1.h:86
int grid_index
Definition: g723_1.h:85
int dirac_train
Definition: g723_1.h:83
int pulse_sign
Definition: g723_1.h:84
int ad_cb_lag
adaptive codebook lag
Definition: g723_1.h:81
int pulse_pos
Definition: g723_1.h:87
Harmonic filter parameters.
Definition: g723_1.h:102
int gain
Definition: g723_1.h:104
int index
Definition: g723_1.h:103
#define av_free(p)
#define av_malloc(s)
#define av_log(a,...)
static void error(const char *err)
#define src
Definition: vp8dsp.c:255
else temp
Definition: vf_mcdeint.c:259
static av_always_inline int diff(const uint32_t a, const uint32_t b)
static const uint8_t offset[127][2]
Definition: vf_spp.c:107
float min