FFmpeg coverage


Directory: ../../../ffmpeg/
File: src/libavcodec/ilbcdec.c
Date: 2021-09-26 18:22:30
Exec Total Coverage
Lines: 0 664 0.0%
Branches: 0 278 0.0%

Line Branch Exec Source
1 /*
2 * Copyright (c) 2013, The WebRTC project authors. All rights reserved.
3 *
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions are
6 * met:
7 *
8 * * Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 *
11 * * Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in
13 * the documentation and/or other materials provided with the
14 * distribution.
15 *
16 * * Neither the name of Google nor the names of its contributors may
17 * be used to endorse or promote products derived from this software
18 * without specific prior written permission.
19 *
20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
23 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
24 * HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
25 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
26 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
27 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
28 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
29 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
30 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
31 */
32
33 #include "libavutil/channel_layout.h"
34 #include "avcodec.h"
35 #include "internal.h"
36 #include "get_bits.h"
37 #include "ilbcdata.h"
38
39 #define LPC_N_20MS 1
40 #define LPC_N_30MS 2
41 #define LPC_N_MAX 2
42 #define LSF_NSPLIT 3
43 #define NASUB_MAX 4
44 #define LPC_FILTERORDER 10
45 #define NSUB_MAX 6
46 #define SUBL 40
47
48 #define ST_MEM_L_TBL 85
49 #define MEM_LF_TBL 147
50 #define STATE_SHORT_LEN_20MS 57
51 #define STATE_SHORT_LEN_30MS 58
52
53 #define BLOCKL_MAX 240
54 #define CB_MEML 147
55 #define CB_NSTAGES 3
56 #define CB_HALFFILTERLEN 4
57 #define CB_FILTERLEN 8
58
59 #define ENH_NBLOCKS_TOT 8
60 #define ENH_BLOCKL 80
61 #define ENH_BUFL (ENH_NBLOCKS_TOT)*ENH_BLOCKL
62 #define ENH_BUFL_FILTEROVERHEAD 3
63 #define BLOCKL_MAX 240
64 #define NSUB_20MS 4
65 #define NSUB_30MS 6
66 #define NSUB_MAX 6
67 #define NASUB_20MS 2
68 #define NASUB_30MS 4
69 #define NASUB_MAX 4
70 #define STATE_LEN 80
71 #define STATE_SHORT_LEN_30MS 58
72 #define STATE_SHORT_LEN_20MS 57
73
74 #define SPL_MUL_16_16(a, b) ((int32_t) (((int16_t)(a)) * ((int16_t)(b))))
75 #define SPL_MUL_16_16_RSFT(a, b, c) (SPL_MUL_16_16(a, b) >> (c))
76
77 typedef struct ILBCFrame {
78 int16_t lsf[LSF_NSPLIT*LPC_N_MAX];
79 int16_t cb_index[CB_NSTAGES*(NASUB_MAX + 1)];
80 int16_t gain_index[CB_NSTAGES*(NASUB_MAX + 1)];
81 int16_t ifm;
82 int16_t state_first;
83 int16_t idx[STATE_SHORT_LEN_30MS];
84 int16_t firstbits;
85 int16_t start;
86 } ILBCFrame;
87
88 typedef struct ILBCContext {
89 AVClass *class;
90 int enhancer;
91
92 int mode;
93 GetBitContext gb;
94 ILBCFrame frame;
95
96 int prev_enh_pl;
97 int consPLICount;
98 int last_lag;
99 int state_short_len;
100 int lpc_n;
101 int16_t nasub;
102 int16_t nsub;
103 int block_samples;
104 int16_t no_of_words;
105 int16_t no_of_bytes;
106 int16_t lsfdeq[LPC_FILTERORDER*LPC_N_MAX];
107 int16_t lsfold[LPC_FILTERORDER];
108 int16_t syntMem[LPC_FILTERORDER];
109 int16_t lsfdeqold[LPC_FILTERORDER];
110 int16_t weightdenum[(LPC_FILTERORDER + 1) * NSUB_MAX];
111 int16_t syntdenum[NSUB_MAX * (LPC_FILTERORDER + 1)];
112 int16_t old_syntdenum[NSUB_MAX * (LPC_FILTERORDER + 1)];
113 int16_t enh_buf[ENH_BUFL+ENH_BUFL_FILTEROVERHEAD];
114 int16_t enh_period[ENH_NBLOCKS_TOT];
115 int16_t prevResidual[NSUB_MAX*SUBL];
116 int16_t decresidual[BLOCKL_MAX];
117 int16_t plc_residual[BLOCKL_MAX + LPC_FILTERORDER];
118 int16_t seed;
119 int16_t prevPLI;
120 int16_t prevScale;
121 int16_t prevLag;
122 int16_t per_square;
123 int16_t prev_lpc[LPC_FILTERORDER + 1];
124 int16_t plc_lpc[LPC_FILTERORDER + 1];
125 int16_t hpimemx[2];
126 int16_t hpimemy[4];
127 } ILBCContext;
128
129 static int unpack_frame(ILBCContext *s)
130 {
131 ILBCFrame *frame = &s->frame;
132 GetBitContext *gb = &s->gb;
133 int j;
134
135 frame->lsf[0] = get_bits(gb, 6);
136 frame->lsf[1] = get_bits(gb, 7);
137 frame->lsf[2] = get_bits(gb, 7);
138
139 if (s->mode == 20) {
140 frame->start = get_bits(gb, 2);
141 frame->state_first = get_bits1(gb);
142 frame->ifm = get_bits(gb, 6);
143 frame->cb_index[0] = get_bits(gb, 6) << 1;
144 frame->gain_index[0] = get_bits(gb, 2) << 3;
145 frame->gain_index[1] = get_bits1(gb) << 3;
146 frame->cb_index[3] = get_bits(gb, 7) << 1;
147 frame->gain_index[3] = get_bits1(gb) << 4;
148 frame->gain_index[4] = get_bits1(gb) << 3;
149 frame->gain_index[6] = get_bits1(gb) << 4;
150 } else {
151 frame->lsf[3] = get_bits(gb, 6);
152 frame->lsf[4] = get_bits(gb, 7);
153 frame->lsf[5] = get_bits(gb, 7);
154 frame->start = get_bits(gb, 3);
155 frame->state_first = get_bits1(gb);
156 frame->ifm = get_bits(gb, 6);
157 frame->cb_index[0] = get_bits(gb, 4) << 3;
158 frame->gain_index[0] = get_bits1(gb) << 4;
159 frame->gain_index[1] = get_bits1(gb) << 3;
160 frame->cb_index[3] = get_bits(gb, 6) << 2;
161 frame->gain_index[3] = get_bits1(gb) << 4;
162 frame->gain_index[4] = get_bits1(gb) << 3;
163 }
164
165 for (j = 0; j < 48; j++)
166 frame->idx[j] = get_bits1(gb) << 2;
167
168 if (s->mode == 20) {
169 for (; j < 57; j++)
170 frame->idx[j] = get_bits1(gb) << 2;
171
172 frame->gain_index[1] |= get_bits1(gb) << 2;
173 frame->gain_index[3] |= get_bits(gb, 2) << 2;
174 frame->gain_index[4] |= get_bits1(gb) << 2;
175 frame->gain_index[6] |= get_bits1(gb) << 3;
176 frame->gain_index[7] = get_bits(gb, 2) << 2;
177 } else {
178 for (; j < 58; j++)
179 frame->idx[j] = get_bits1(gb) << 2;
180
181 frame->cb_index[0] |= get_bits(gb, 2) << 1;
182 frame->gain_index[0] |= get_bits1(gb) << 3;
183 frame->gain_index[1] |= get_bits1(gb) << 2;
184 frame->cb_index[3] |= get_bits1(gb) << 1;
185 frame->cb_index[6] = get_bits1(gb) << 7;
186 frame->cb_index[6] |= get_bits(gb, 6) << 1;
187 frame->cb_index[9] = get_bits(gb, 7) << 1;
188 frame->cb_index[12] = get_bits(gb, 3) << 5;
189 frame->cb_index[12] |= get_bits(gb, 4) << 1;
190 frame->gain_index[3] |= get_bits(gb, 2) << 2;
191 frame->gain_index[4] |= get_bits(gb, 2) << 1;
192 frame->gain_index[6] = get_bits(gb, 2) << 3;
193 frame->gain_index[7] = get_bits(gb, 2) << 2;
194 frame->gain_index[9] = get_bits1(gb) << 4;
195 frame->gain_index[10] = get_bits1(gb) << 3;
196 frame->gain_index[12] = get_bits1(gb) << 4;
197 frame->gain_index[13] = get_bits1(gb) << 3;
198 }
199
200 for (j = 0; j < 56; j++)
201 frame->idx[j] |= get_bits(gb, 2);
202
203 if (s->mode == 20) {
204 frame->idx[56] |= get_bits(gb, 2);
205 frame->cb_index[0] |= get_bits1(gb);
206 frame->cb_index[1] = get_bits(gb, 7);
207 frame->cb_index[2] = get_bits(gb, 6) << 1;
208 frame->cb_index[2] |= get_bits1(gb);
209 frame->gain_index[0] |= get_bits(gb, 3);
210 frame->gain_index[1] |= get_bits(gb, 2);
211 frame->gain_index[2] = get_bits(gb, 3);
212 frame->cb_index[3] |= get_bits1(gb);
213 frame->cb_index[4] = get_bits(gb, 6) << 1;
214 frame->cb_index[4] |= get_bits1(gb);
215 frame->cb_index[5] = get_bits(gb, 7);
216 frame->cb_index[6] = get_bits(gb, 8);
217 frame->cb_index[7] = get_bits(gb, 8);
218 frame->cb_index[8] = get_bits(gb, 8);
219 frame->gain_index[3] |= get_bits(gb, 2);
220 frame->gain_index[4] |= get_bits(gb, 2);
221 frame->gain_index[5] = get_bits(gb, 3);
222 frame->gain_index[6] |= get_bits(gb, 3);
223 frame->gain_index[7] |= get_bits(gb, 2);
224 frame->gain_index[8] = get_bits(gb, 3);
225 } else {
226 frame->idx[56] |= get_bits(gb, 2);
227 frame->idx[57] |= get_bits(gb, 2);
228 frame->cb_index[0] |= get_bits1(gb);
229 frame->cb_index[1] = get_bits(gb, 7);
230 frame->cb_index[2] = get_bits(gb, 4) << 3;
231 frame->cb_index[2] |= get_bits(gb, 3);
232 frame->gain_index[0] |= get_bits(gb, 3);
233 frame->gain_index[1] |= get_bits(gb, 2);
234 frame->gain_index[2] = get_bits(gb, 3);
235 frame->cb_index[3] |= get_bits1(gb);
236 frame->cb_index[4] = get_bits(gb, 4) << 3;
237 frame->cb_index[4] |= get_bits(gb, 3);
238 frame->cb_index[5] = get_bits(gb, 7);
239 frame->cb_index[6] |= get_bits1(gb);
240 frame->cb_index[7] = get_bits(gb, 5) << 3;
241 frame->cb_index[7] |= get_bits(gb, 3);
242 frame->cb_index[8] = get_bits(gb, 8);
243 frame->cb_index[9] |= get_bits1(gb);
244 frame->cb_index[10] = get_bits(gb, 4) << 4;
245 frame->cb_index[10] |= get_bits(gb, 4);
246 frame->cb_index[11] = get_bits(gb, 8);
247 frame->cb_index[12] |= get_bits1(gb);
248 frame->cb_index[13] = get_bits(gb, 3) << 5;
249 frame->cb_index[13] |= get_bits(gb, 5);
250 frame->cb_index[14] = get_bits(gb, 8);
251 frame->gain_index[3] |= get_bits(gb, 2);
252 frame->gain_index[4] |= get_bits1(gb);
253 frame->gain_index[5] = get_bits(gb, 3);
254 frame->gain_index[6] |= get_bits(gb, 3);
255 frame->gain_index[7] |= get_bits(gb, 2);
256 frame->gain_index[8] = get_bits(gb, 3);
257 frame->gain_index[9] |= get_bits(gb, 4);
258 frame->gain_index[10] |= get_bits1(gb) << 2;
259 frame->gain_index[10] |= get_bits(gb, 2);
260 frame->gain_index[11] = get_bits(gb, 3);
261 frame->gain_index[12] |= get_bits(gb, 4);
262 frame->gain_index[13] |= get_bits(gb, 3);
263 frame->gain_index[14] = get_bits(gb, 3);
264 }
265
266 return get_bits1(gb);
267 }
268
269 static void index_conv(int16_t *index)
270 {
271 int k;
272
273 for (k = 4; k < 6; k++) {
274 if (index[k] >= 44 && index[k] < 108) {
275 index[k] += 64;
276 } else if (index[k] >= 108 && index[k] < 128) {
277 index[k] += 128;
278 }
279 }
280 }
281
282 static void lsf_dequantization(int16_t *lsfdeq, int16_t *index, int16_t lpc_n)
283 {
284 int i, j, pos = 0, cb_pos = 0;
285
286 for (i = 0; i < LSF_NSPLIT; i++) {
287 for (j = 0; j < lsf_dim_codebook[i]; j++) {
288 lsfdeq[pos + j] = lsf_codebook[cb_pos + index[i] * lsf_dim_codebook[i] + j];
289 }
290
291 pos += lsf_dim_codebook[i];
292 cb_pos += lsf_size_codebook[i] * lsf_dim_codebook[i];
293 }
294
295 if (lpc_n > 1) {
296 pos = 0;
297 cb_pos = 0;
298 for (i = 0; i < LSF_NSPLIT; i++) {
299 for (j = 0; j < lsf_dim_codebook[i]; j++) {
300 lsfdeq[LPC_FILTERORDER + pos + j] = lsf_codebook[cb_pos +
301 index[LSF_NSPLIT + i] * lsf_dim_codebook[i] + j];
302 }
303
304 pos += lsf_dim_codebook[i];
305 cb_pos += lsf_size_codebook[i] * lsf_dim_codebook[i];
306 }
307 }
308 }
309
310 static void lsf_check_stability(int16_t *lsf, int dim, int nb_vectors)
311 {
312 for (int n = 0; n < 2; n++) {
313 for (int m = 0; m < nb_vectors; m++) {
314 for (int k = 0; k < dim - 1; k++) {
315 int i = m * dim + k;
316
317 if ((lsf[i + 1] - lsf[i]) < 319) {
318 if (lsf[i + 1] < lsf[i]) {
319 lsf[i + 1] = lsf[i] + 160;
320 lsf[i] = lsf[i + 1] - 160;
321 } else {
322 lsf[i] -= 160;
323 lsf[i + 1] += 160;
324 }
325 }
326
327 lsf[i] = av_clip(lsf[i], 82, 25723);
328 }
329 }
330 }
331 }
332
333 static void lsf_interpolate(int16_t *out, int16_t *in1,
334 int16_t *in2, int16_t coef,
335 int size)
336 {
337 int invcoef = 16384 - coef, i;
338
339 for (i = 0; i < size; i++)
340 out[i] = (coef * in1[i] + invcoef * in2[i] + 8192) >> 14;
341 }
342
343 static void lsf2lsp(int16_t *lsf, int16_t *lsp, int order)
344 {
345 int16_t diff, freq;
346 int32_t tmp;
347 int i, k;
348
349 for (i = 0; i < order; i++) {
350 freq = (lsf[i] * 20861) >> 15;
351 /* 20861: 1.0/(2.0*PI) in Q17 */
352 /*
353 Upper 8 bits give the index k and
354 Lower 8 bits give the difference, which needs
355 to be approximated linearly
356 */
357 k = FFMIN(freq >> 8, 63);
358 diff = freq & 0xFF;
359
360 /* Calculate linear approximation */
361 tmp = cos_derivative_tbl[k] * diff;
362 lsp[i] = cos_tbl[k] + (tmp >> 12);
363 }
364 }
365
366 static void get_lsp_poly(int16_t *lsp, int32_t *f)
367 {
368 int16_t high, low;
369 int i, j, k, l;
370 int32_t tmp;
371
372 f[0] = 16777216;
373 f[1] = lsp[0] * -1024;
374
375 for (i = 2, k = 2, l = 2; i <= 5; i++, k += 2) {
376 f[l] = f[l - 2];
377
378 for (j = i; j > 1; j--, l--) {
379 high = f[l - 1] >> 16;
380 low = (f[l - 1] - (high * (1 << 16))) >> 1;
381
382 tmp = ((high * lsp[k]) * 4) + (((low * lsp[k]) >> 15) * 4);
383
384 f[l] += f[l - 2];
385 f[l] -= (unsigned)tmp;
386 }
387
388 f[l] -= lsp[k] * (1 << 10);
389 l += i;
390 }
391 }
392
393 static void lsf2poly(int16_t *a, int16_t *lsf)
394 {
395 int32_t f[2][6];
396 int16_t lsp[10];
397 int32_t tmp;
398 int i;
399
400 lsf2lsp(lsf, lsp, LPC_FILTERORDER);
401
402 get_lsp_poly(&lsp[0], f[0]);
403 get_lsp_poly(&lsp[1], f[1]);
404
405 for (i = 5; i > 0; i--) {
406 f[0][i] += (unsigned)f[0][i - 1];
407 f[1][i] -= (unsigned)f[1][i - 1];
408 }
409
410 a[0] = 4096;
411 for (i = 5; i > 0; i--) {
412 tmp = f[0][6 - i] + (unsigned)f[1][6 - i] + 4096;
413 a[6 - i] = tmp >> 13;
414
415 tmp = f[0][6 - i] - (unsigned)f[1][6 - i] + 4096;
416 a[5 + i] = tmp >> 13;
417 }
418 }
419
420 static void lsp_interpolate2polydec(int16_t *a, int16_t *lsf1,
421 int16_t *lsf2, int coef, int length)
422 {
423 int16_t lsftmp[LPC_FILTERORDER];
424
425 lsf_interpolate(lsftmp, lsf1, lsf2, coef, length);
426 lsf2poly(a, lsftmp);
427 }
428
429 static void bw_expand(int16_t *out, const int16_t *in, const int16_t *coef, int length)
430 {
431 int i;
432
433 out[0] = in[0];
434 for (i = 1; i < length; i++)
435 out[i] = (coef[i] * in[i] + 16384) >> 15;
436 }
437
438 static void lsp_interpolate(int16_t *syntdenum, int16_t *weightdenum,
439 int16_t *lsfdeq, int16_t length,
440 ILBCContext *s)
441 {
442 int16_t lp[LPC_FILTERORDER + 1], *lsfdeq2;
443 int i, pos, lp_length;
444
445 lsfdeq2 = lsfdeq + length;
446 lp_length = length + 1;
447
448 if (s->mode == 30) {
449 lsp_interpolate2polydec(lp, (*s).lsfdeqold, lsfdeq, lsf_weight_30ms[0], length);
450 memcpy(syntdenum, lp, lp_length * 2);
451 bw_expand(weightdenum, lp, kLpcChirpSyntDenum, lp_length);
452
453 pos = lp_length;
454 for (i = 1; i < 6; i++) {
455 lsp_interpolate2polydec(lp, lsfdeq, lsfdeq2,
456 lsf_weight_30ms[i],
457 length);
458 memcpy(syntdenum + pos, lp, lp_length * 2);
459 bw_expand(weightdenum + pos, lp, kLpcChirpSyntDenum, lp_length);
460 pos += lp_length;
461 }
462 } else {
463 pos = 0;
464 for (i = 0; i < s->nsub; i++) {
465 lsp_interpolate2polydec(lp, s->lsfdeqold, lsfdeq,
466 lsf_weight_20ms[i], length);
467 memcpy(syntdenum + pos, lp, lp_length * 2);
468 bw_expand(weightdenum + pos, lp, kLpcChirpSyntDenum, lp_length);
469 pos += lp_length;
470 }
471 }
472
473 if (s->mode == 30) {
474 memcpy(s->lsfdeqold, lsfdeq2, length * 2);
475 } else {
476 memcpy(s->lsfdeqold, lsfdeq, length * 2);
477 }
478 }
479
480 static void filter_mafq12(int16_t *in_ptr, int16_t *out_ptr,
481 int16_t *B, int16_t B_length,
482 int16_t length)
483 {
484 int o, i, j;
485
486 for (i = 0; i < length; i++) {
487 const int16_t *b_ptr = &B[0];
488 const int16_t *x_ptr = &in_ptr[i];
489
490 o = 0;
491 for (j = 0; j < B_length; j++)
492 o += b_ptr[j] * *x_ptr--;
493
494 o = av_clip(o, -134217728, 134215679);
495
496 out_ptr[i] = ((o + 2048) >> 12);
497 }
498 }
499
500 static void filter_arfq12(const int16_t *data_in,
501 int16_t *data_out,
502 const int16_t *coefficients,
503 int coefficients_length,
504 int data_length)
505 {
506 int i, j;
507
508 for (i = 0; i < data_length; i++) {
509 int output = 0, sum = 0;
510
511 for (j = coefficients_length - 1; j > 0; j--) {
512 sum += (unsigned)(coefficients[j] * data_out[i - j]);
513 }
514
515 output = coefficients[0] * data_in[i] - (unsigned)sum;
516 output = av_clip(output, -134217728, 134215679);
517
518 data_out[i] = (output + 2048) >> 12;
519 }
520 }
521
522 static void state_construct(int16_t ifm, int16_t *idx,
523 int16_t *synt_denum, int16_t *Out_fix,
524 int16_t len)
525 {
526 int k;
527 int16_t maxVal;
528 int16_t *tmp1, *tmp2, *tmp3;
529 /* Stack based */
530 int16_t numerator[1 + LPC_FILTERORDER];
531 int16_t sampleValVec[2 * STATE_SHORT_LEN_30MS + LPC_FILTERORDER];
532 int16_t sampleMaVec[2 * STATE_SHORT_LEN_30MS + LPC_FILTERORDER];
533 int16_t *sampleVal = &sampleValVec[LPC_FILTERORDER];
534 int16_t *sampleMa = &sampleMaVec[LPC_FILTERORDER];
535 int16_t *sampleAr = &sampleValVec[LPC_FILTERORDER];
536
537 /* initialization of coefficients */
538
539 for (k = 0; k < LPC_FILTERORDER + 1; k++) {
540 numerator[k] = synt_denum[LPC_FILTERORDER - k];
541 }
542
543 /* decoding of the maximum value */
544
545 maxVal = frg_quant_mod[ifm];
546
547 /* decoding of the sample values */
548 tmp1 = sampleVal;
549 tmp2 = &idx[len - 1];
550
551 if (ifm < 37) {
552 for (k = 0; k < len; k++) {
553 /*the shifting is due to the Q13 in sq4_fixQ13[i], also the adding of 2097152 (= 0.5 << 22)
554 maxVal is in Q8 and result is in Q(-1) */
555 (*tmp1) = (int16_t) ((SPL_MUL_16_16(maxVal, ilbc_state[(*tmp2)]) + 2097152) >> 22);
556 tmp1++;
557 tmp2--;
558 }
559 } else if (ifm < 59) {
560 for (k = 0; k < len; k++) {
561 /*the shifting is due to the Q13 in sq4_fixQ13[i], also the adding of 262144 (= 0.5 << 19)
562 maxVal is in Q5 and result is in Q(-1) */
563 (*tmp1) = (int16_t) ((SPL_MUL_16_16(maxVal, ilbc_state[(*tmp2)]) + 262144) >> 19);
564 tmp1++;
565 tmp2--;
566 }
567 } else {
568 for (k = 0; k < len; k++) {
569 /*the shifting is due to the Q13 in sq4_fixQ13[i], also the adding of 65536 (= 0.5 << 17)
570 maxVal is in Q3 and result is in Q(-1) */
571 (*tmp1) = (int16_t) ((SPL_MUL_16_16(maxVal, ilbc_state[(*tmp2)]) + 65536) >> 17);
572 tmp1++;
573 tmp2--;
574 }
575 }
576
577 /* Set the rest of the data to zero */
578 memset(&sampleVal[len], 0, len * 2);
579
580 /* circular convolution with all-pass filter */
581
582 /* Set the state to zero */
583 memset(sampleValVec, 0, LPC_FILTERORDER * 2);
584
585 /* Run MA filter + AR filter */
586 filter_mafq12(sampleVal, sampleMa, numerator, LPC_FILTERORDER + 1, len + LPC_FILTERORDER);
587 memset(&sampleMa[len + LPC_FILTERORDER], 0, (len - LPC_FILTERORDER) * 2);
588 filter_arfq12(sampleMa, sampleAr, synt_denum, LPC_FILTERORDER + 1, 2 * len);
589
590 tmp1 = &sampleAr[len - 1];
591 tmp2 = &sampleAr[2 * len - 1];
592 tmp3 = Out_fix;
593 for (k = 0; k < len; k++) {
594 (*tmp3) = (*tmp1) + (*tmp2);
595 tmp1--;
596 tmp2--;
597 tmp3++;
598 }
599 }
600
601 static int16_t gain_dequantization(int index, int max_in, int stage)
602 {
603 int16_t scale = FFMAX(1638, FFABS(max_in));
604
605 return ((scale * ilbc_gain[stage][index]) + 8192) >> 14;
606 }
607
608 static void vector_rmultiplication(int16_t *out, const int16_t *in,
609 const int16_t *win,
610 int length, int shift)
611 {
612 for (int i = 0; i < length; i++)
613 out[i] = (in[i] * win[-i]) >> shift;
614 }
615
616 static void vector_multiplication(int16_t *out, const int16_t *in,
617 const int16_t *win, int length,
618 int shift)
619 {
620 for (int i = 0; i < length; i++)
621 out[i] = (in[i] * win[i]) >> shift;
622 }
623
624 static void add_vector_and_shift(int16_t *out, const int16_t *in1,
625 const int16_t *in2, int length,
626 int shift)
627 {
628 for (int i = 0; i < length; i++)
629 out[i] = (in1[i] + in2[i]) >> shift;
630 }
631
632 static void create_augmented_vector(int index, int16_t *buffer, int16_t *cbVec)
633 {
634 int16_t cbVecTmp[4];
635 int interpolation_length = FFMIN(4, index);
636 int16_t ilow = index - interpolation_length;
637
638 memcpy(cbVec, buffer - index, index * 2);
639
640 vector_multiplication(&cbVec[ilow], buffer - index - interpolation_length, alpha, interpolation_length, 15);
641 vector_rmultiplication(cbVecTmp, buffer - interpolation_length, &alpha[interpolation_length - 1], interpolation_length, 15);
642 add_vector_and_shift(&cbVec[ilow], &cbVec[ilow], cbVecTmp, interpolation_length, 0);
643
644 memcpy(cbVec + index, buffer - index, FFMIN(SUBL - index, index) * sizeof(*cbVec));
645 }
646
647 static void get_codebook(int16_t * cbvec, /* (o) Constructed codebook vector */
648 int16_t * mem, /* (i) Codebook buffer */
649 int16_t index, /* (i) Codebook index */
650 int16_t lMem, /* (i) Length of codebook buffer */
651 int16_t cbveclen /* (i) Codebook vector length */
652 )
653 {
654 int16_t k, base_size;
655 int16_t lag;
656 /* Stack based */
657 int16_t tempbuff2[SUBL + 5];
658
659 /* Determine size of codebook sections */
660 base_size = lMem - cbveclen + 1;
661
662 if (cbveclen == SUBL) {
663 base_size += cbveclen / 2;
664 }
665
666 /* No filter -> First codebook section */
667 if (index < lMem - cbveclen + 1) {
668 /* first non-interpolated vectors */
669
670 k = index + cbveclen;
671 /* get vector */
672 memcpy(cbvec, mem + lMem - k, cbveclen * 2);
673 } else if (index < base_size) {
674
675 /* Calculate lag */
676
677 k = (int16_t) SPL_MUL_16_16(2, (index - (lMem - cbveclen + 1))) + cbveclen;
678
679 lag = k / 2;
680
681 create_augmented_vector(lag, mem + lMem, cbvec);
682 } else {
683 int16_t memIndTest;
684
685 /* first non-interpolated vectors */
686
687 if (index - base_size < lMem - cbveclen + 1) {
688
689 /* Set up filter memory, stuff zeros outside memory buffer */
690
691 memIndTest = lMem - (index - base_size + cbveclen);
692
693 memset(mem - CB_HALFFILTERLEN, 0, CB_HALFFILTERLEN * 2);
694 memset(mem + lMem, 0, CB_HALFFILTERLEN * 2);
695
696 /* do filtering to get the codebook vector */
697
698 filter_mafq12(&mem[memIndTest + 4], cbvec, (int16_t *) kCbFiltersRev, CB_FILTERLEN, cbveclen);
699 } else {
700 /* interpolated vectors */
701 /* Stuff zeros outside memory buffer */
702 memIndTest = lMem - cbveclen - CB_FILTERLEN;
703 memset(mem + lMem, 0, CB_HALFFILTERLEN * 2);
704
705 /* do filtering */
706 filter_mafq12(&mem[memIndTest + 7], tempbuff2, (int16_t *) kCbFiltersRev, CB_FILTERLEN, (int16_t) (cbveclen + 5));
707
708 /* Calculate lag index */
709 lag = (cbveclen << 1) - 20 + index - base_size - lMem - 1;
710
711 create_augmented_vector(lag, tempbuff2 + SUBL + 5, cbvec);
712 }
713 }
714 }
715
716 static void construct_vector (
717 int16_t *decvector, /* (o) Decoded vector */
718 int16_t *index, /* (i) Codebook indices */
719 int16_t *gain_index, /* (i) Gain quantization indices */
720 int16_t *mem, /* (i) Buffer for codevector construction */
721 int16_t lMem, /* (i) Length of buffer */
722 int16_t veclen)
723 {
724 int16_t gain[CB_NSTAGES];
725 int16_t cbvec0[SUBL];
726 int16_t cbvec1[SUBL];
727 int16_t cbvec2[SUBL];
728 unsigned a32;
729 int16_t *gainPtr;
730 int j;
731
732 /* gain de-quantization */
733
734 gain[0] = gain_dequantization(gain_index[0], 16384, 0);
735 gain[1] = gain_dequantization(gain_index[1], gain[0], 1);
736 gain[2] = gain_dequantization(gain_index[2], gain[1], 2);
737
738 /* codebook vector construction and construction of total vector */
739
740 /* Stack based */
741 get_codebook(cbvec0, mem, index[0], lMem, veclen);
742 get_codebook(cbvec1, mem, index[1], lMem, veclen);
743 get_codebook(cbvec2, mem, index[2], lMem, veclen);
744
745 gainPtr = &gain[0];
746 for (j = 0; j < veclen; j++) {
747 a32 = SPL_MUL_16_16(*gainPtr++, cbvec0[j]);
748 a32 += SPL_MUL_16_16(*gainPtr++, cbvec1[j]);
749 a32 += SPL_MUL_16_16(*gainPtr, cbvec2[j]);
750 gainPtr -= 2;
751 decvector[j] = (int)(a32 + 8192) >> 14;
752 }
753 }
754
755 static void reverse_memcpy(int16_t *dest, int16_t *source, int length)
756 {
757 int16_t* destPtr = dest;
758 int16_t* sourcePtr = source;
759 int j;
760
761 for (j = 0; j < length; j++)
762 *destPtr-- = *sourcePtr++;
763 }
764
765 static void decode_residual(ILBCContext *s,
766 ILBCFrame *encbits,
767 int16_t *decresidual,
768 int16_t *syntdenum)
769 {
770 int16_t meml_gotten, Nfor, Nback, diff, start_pos;
771 int16_t subcount, subframe;
772 int16_t *reverseDecresidual = s->enh_buf; /* Reversed decoded data, used for decoding backwards in time (reuse memory in state) */
773 int16_t *memVec = s->prevResidual;
774 int16_t *mem = &memVec[CB_HALFFILTERLEN]; /* Memory for codebook */
775
776 diff = STATE_LEN - s->state_short_len;
777
778 if (encbits->state_first == 1) {
779 start_pos = (encbits->start - 1) * SUBL;
780 } else {
781 start_pos = (encbits->start - 1) * SUBL + diff;
782 }
783
784 /* decode scalar part of start state */
785
786 state_construct(encbits->ifm, encbits->idx, &syntdenum[(encbits->start - 1) * (LPC_FILTERORDER + 1)], &decresidual[start_pos], s->state_short_len);
787
788 if (encbits->state_first) { /* put adaptive part in the end */
789 /* setup memory */
790 memset(mem, 0, (int16_t) (CB_MEML - s->state_short_len) * 2);
791 memcpy(mem + CB_MEML - s->state_short_len, decresidual + start_pos, s->state_short_len * 2);
792
793 /* construct decoded vector */
794
795 construct_vector(&decresidual[start_pos + s->state_short_len], encbits->cb_index, encbits->gain_index, mem + CB_MEML - ST_MEM_L_TBL, ST_MEM_L_TBL, (int16_t) diff);
796
797 } else { /* put adaptive part in the beginning */
798 /* setup memory */
799 meml_gotten = s->state_short_len;
800 reverse_memcpy(mem + CB_MEML - 1, decresidual + start_pos, meml_gotten);
801 memset(mem, 0, (int16_t) (CB_MEML - meml_gotten) * 2);
802
803 /* construct decoded vector */
804 construct_vector(reverseDecresidual, encbits->cb_index, encbits->gain_index, mem + CB_MEML - ST_MEM_L_TBL, ST_MEM_L_TBL, diff);
805
806 /* get decoded residual from reversed vector */
807 reverse_memcpy(&decresidual[start_pos - 1], reverseDecresidual, diff);
808 }
809
810 /* counter for predicted subframes */
811 subcount = 1;
812
813 /* forward prediction of subframes */
814 Nfor = s->nsub - encbits->start - 1;
815
816 if (Nfor > 0) {
817 /* setup memory */
818 memset(mem, 0, (CB_MEML - STATE_LEN) * 2);
819 memcpy(mem + CB_MEML - STATE_LEN, decresidual + (encbits->start - 1) * SUBL, STATE_LEN * 2);
820
821 /* loop over subframes to encode */
822 for (subframe = 0; subframe < Nfor; subframe++) {
823 /* construct decoded vector */
824 construct_vector(&decresidual[(encbits->start + 1 + subframe) * SUBL], encbits->cb_index + subcount * CB_NSTAGES, encbits->gain_index + subcount * CB_NSTAGES, mem, MEM_LF_TBL, SUBL);
825
826 /* update memory */
827 memmove(mem, mem + SUBL, (CB_MEML - SUBL) * sizeof(*mem));
828 memcpy(mem + CB_MEML - SUBL, &decresidual[(encbits->start + 1 + subframe) * SUBL], SUBL * 2);
829
830 subcount++;
831 }
832
833 }
834
835 /* backward prediction of subframes */
836 Nback = encbits->start - 1;
837
838 if (Nback > 0) {
839 /* setup memory */
840 meml_gotten = SUBL * (s->nsub + 1 - encbits->start);
841 if (meml_gotten > CB_MEML) {
842 meml_gotten = CB_MEML;
843 }
844
845 reverse_memcpy(mem + CB_MEML - 1, decresidual + (encbits->start - 1) * SUBL, meml_gotten);
846 memset(mem, 0, (int16_t) (CB_MEML - meml_gotten) * 2);
847
848 /* loop over subframes to decode */
849 for (subframe = 0; subframe < Nback; subframe++) {
850 /* construct decoded vector */
851 construct_vector(&reverseDecresidual[subframe * SUBL], encbits->cb_index + subcount * CB_NSTAGES,
852 encbits->gain_index + subcount * CB_NSTAGES, mem, MEM_LF_TBL, SUBL);
853
854 /* update memory */
855 memmove(mem, mem + SUBL, (CB_MEML - SUBL) * sizeof(*mem));
856 memcpy(mem + CB_MEML - SUBL, &reverseDecresidual[subframe * SUBL], SUBL * 2);
857
858 subcount++;
859 }
860
861 /* get decoded residual from reversed vector */
862 reverse_memcpy(decresidual + SUBL * Nback - 1, reverseDecresidual, SUBL * Nback);
863 }
864 }
865
866 static int16_t max_abs_value_w16(const int16_t* vector, int length)
867 {
868 int i = 0, absolute = 0, maximum = 0;
869
870 if (vector == NULL || length <= 0) {
871 return -1;
872 }
873
874 for (i = 0; i < length; i++) {
875 absolute = FFABS(vector[i]);
876 if (absolute > maximum)
877 maximum = absolute;
878 }
879
880 // Guard the case for abs(-32768).
881 return FFMIN(maximum, INT16_MAX);
882 }
883
884 static int16_t get_size_in_bits(uint32_t n)
885 {
886 int16_t bits;
887
888 if (0xFFFF0000 & n) {
889 bits = 16;
890 } else {
891 bits = 0;
892 }
893
894 if (0x0000FF00 & (n >> bits)) bits += 8;
895 if (0x000000F0 & (n >> bits)) bits += 4;
896 if (0x0000000C & (n >> bits)) bits += 2;
897 if (0x00000002 & (n >> bits)) bits += 1;
898 if (0x00000001 & (n >> bits)) bits += 1;
899
900 return bits;
901 }
902
903 static int32_t scale_dot_product(const int16_t *v1, const int16_t *v2, int length, int scaling)
904 {
905 int64_t sum = 0;
906
907 for (int i = 0; i < length; i++)
908 sum += (v1[i] * v2[i]) >> scaling;
909
910 return av_clipl_int32(sum);
911 }
912
913 static void correlation(int32_t *corr, int32_t *ener, int16_t *buffer,
914 int16_t lag, int16_t blen, int16_t srange, int16_t scale)
915 {
916 int16_t *w16ptr;
917
918 w16ptr = &buffer[blen - srange - lag];
919
920 *corr = scale_dot_product(&buffer[blen - srange], w16ptr, srange, scale);
921 *ener = scale_dot_product(w16ptr, w16ptr, srange, scale);
922
923 if (*ener == 0) {
924 *corr = 0;
925 *ener = 1;
926 }
927 }
928
929 #define SPL_SHIFT_W32(x, c) (((c) >= 0) ? ((x) << (c)) : ((x) >> (-(c))))
930
931 static int16_t norm_w32(int32_t a)
932 {
933 if (a == 0) {
934 return 0;
935 } else if (a < 0) {
936 a = ~a;
937 }
938
939 return ff_clz(a);
940 }
941
942 static int32_t div_w32_w16(int32_t num, int16_t den)
943 {
944 if (den != 0)
945 return num / den;
946 else
947 return 0x7FFFFFFF;
948 }
949
950 static void do_plc(int16_t *plc_residual, /* (o) concealed residual */
951 int16_t *plc_lpc, /* (o) concealed LP parameters */
952 int16_t PLI, /* (i) packet loss indicator
953 0 - no PL, 1 = PL */
954 int16_t *decresidual, /* (i) decoded residual */
955 int16_t *lpc, /* (i) decoded LPC (only used for no PL) */
956 int16_t inlag, /* (i) pitch lag */
957 ILBCContext *s) /* (i/o) decoder instance */
958 {
959 int16_t i, pick;
960 int32_t cross, ener, cross_comp, ener_comp = 0;
961 int32_t measure, max_measure, energy;
962 int16_t max, cross_square_max, cross_square;
963 int16_t j, lag, tmp1, tmp2, randlag;
964 int16_t shift1, shift2, shift3, shift_max;
965 int16_t scale3;
966 int16_t corrLen;
967 int32_t tmpW32, tmp2W32;
968 int16_t use_gain;
969 int16_t tot_gain;
970 int16_t max_perSquare;
971 int16_t scale1, scale2;
972 int16_t totscale;
973 int32_t nom;
974 int16_t denom;
975 int16_t pitchfact;
976 int16_t use_lag;
977 int ind;
978 int16_t randvec[BLOCKL_MAX];
979
980 /* Packet Loss */
981 if (PLI == 1) {
982
983 s->consPLICount += 1;
984
985 /* if previous frame not lost,
986 determine pitch pred. gain */
987
988 if (s->prevPLI != 1) {
989
990 /* Maximum 60 samples are correlated, preserve as high accuracy
991 as possible without getting overflow */
992 max = max_abs_value_w16(s->prevResidual, s->block_samples);
993 scale3 = (get_size_in_bits(max) << 1) - 25;
994 if (scale3 < 0) {
995 scale3 = 0;
996 }
997
998 /* Store scale for use when interpolating between the
999 * concealment and the received packet */
1000 s->prevScale = scale3;
1001
1002 /* Search around the previous lag +/-3 to find the
1003 best pitch period */
1004 lag = inlag - 3;
1005
1006 /* Guard against getting outside the frame */
1007 corrLen = FFMIN(60, s->block_samples - (inlag + 3));
1008
1009 correlation(&cross, &ener, s->prevResidual, lag, s->block_samples, corrLen, scale3);
1010
1011 /* Normalize and store cross^2 and the number of shifts */
1012 shift_max = get_size_in_bits(FFABS(cross)) - 15;
1013 cross_square_max = (int16_t) SPL_MUL_16_16_RSFT(SPL_SHIFT_W32(cross, -shift_max), SPL_SHIFT_W32(cross, -shift_max), 15);
1014
1015 for (j = inlag - 2; j <= inlag + 3; j++) {
1016 correlation(&cross_comp, &ener_comp, s->prevResidual, j, s->block_samples, corrLen, scale3);
1017
1018 /* Use the criteria (corr*corr)/energy to compare if
1019 this lag is better or not. To avoid the division,
1020 do a cross multiplication */
1021 shift1 = get_size_in_bits(FFABS(cross_comp)) - 15;
1022 cross_square = (int16_t) SPL_MUL_16_16_RSFT(SPL_SHIFT_W32(cross_comp, -shift1), SPL_SHIFT_W32(cross_comp, -shift1), 15);
1023
1024 shift2 = get_size_in_bits(ener) - 15;
1025 measure = SPL_MUL_16_16(SPL_SHIFT_W32(ener, -shift2), cross_square);
1026
1027 shift3 = get_size_in_bits(ener_comp) - 15;
1028 max_measure = SPL_MUL_16_16(SPL_SHIFT_W32(ener_comp, -shift3), cross_square_max);
1029
1030 /* Calculate shift value, so that the two measures can
1031 be put in the same Q domain */
1032 if (((shift_max << 1) + shift3) > ((shift1 << 1) + shift2)) {
1033 tmp1 = FFMIN(31, (shift_max << 1) + shift3 - (shift1 << 1) - shift2);
1034 tmp2 = 0;
1035 } else {
1036 tmp1 = 0;
1037 tmp2 = FFMIN(31, (shift1 << 1) + shift2 - (shift_max << 1) - shift3);
1038 }
1039
1040 if ((measure >> tmp1) > (max_measure >> tmp2)) {
1041 /* New lag is better => record lag, measure and domain */
1042 lag = j;
1043 cross_square_max = cross_square;
1044 cross = cross_comp;
1045 shift_max = shift1;
1046 ener = ener_comp;
1047 }
1048 }
1049
1050 /* Calculate the periodicity for the lag with the maximum correlation.
1051
1052 Definition of the periodicity:
1053 abs(corr(vec1, vec2))/(sqrt(energy(vec1))*sqrt(energy(vec2)))
1054
1055 Work in the Square domain to simplify the calculations
1056 max_perSquare is less than 1 (in Q15)
1057 */
1058 tmp2W32 = scale_dot_product(&s->prevResidual[s->block_samples - corrLen], &s->prevResidual[s->block_samples - corrLen], corrLen, scale3);
1059
1060 if ((tmp2W32 > 0) && (ener_comp > 0)) {
1061 /* norm energies to int16_t, compute the product of the energies and
1062 use the upper int16_t as the denominator */
1063
1064 scale1 = norm_w32(tmp2W32) - 16;
1065 tmp1 = SPL_SHIFT_W32(tmp2W32, scale1);
1066
1067 scale2 = norm_w32(ener) - 16;
1068 tmp2 = SPL_SHIFT_W32(ener, scale2);
1069 denom = SPL_MUL_16_16_RSFT(tmp1, tmp2, 16); /* denom in Q(scale1+scale2-16) */
1070
1071 /* Square the cross correlation and norm it such that max_perSquare
1072 will be in Q15 after the division */
1073
1074 totscale = scale1 + scale2 - 1;
1075 tmp1 = SPL_SHIFT_W32(cross, (totscale >> 1));
1076 tmp2 = SPL_SHIFT_W32(cross, totscale - (totscale >> 1));
1077
1078 nom = SPL_MUL_16_16(tmp1, tmp2);
1079 max_perSquare = div_w32_w16(nom, denom);
1080 } else {
1081 max_perSquare = 0;
1082 }
1083 } else {
1084 /* previous frame lost, use recorded lag and gain */
1085 lag = s->prevLag;
1086 max_perSquare = s->per_square;
1087 }
1088
1089 /* Attenuate signal and scale down pitch pred gain if
1090 several frames lost consecutively */
1091
1092 use_gain = 32767; /* 1.0 in Q15 */
1093
1094 if (s->consPLICount * s->block_samples > 320) {
1095 use_gain = 29491; /* 0.9 in Q15 */
1096 } else if (s->consPLICount * s->block_samples > 640) {
1097 use_gain = 22938; /* 0.7 in Q15 */
1098 } else if (s->consPLICount * s->block_samples > 960) {
1099 use_gain = 16384; /* 0.5 in Q15 */
1100 } else if (s->consPLICount * s->block_samples > 1280) {
1101 use_gain = 0; /* 0.0 in Q15 */
1102 }
1103
1104 /* Compute mixing factor of picth repeatition and noise:
1105 for max_per>0.7 set periodicity to 1.0
1106 0.4<max_per<0.7 set periodicity to (maxper-0.4)/0.7-0.4)
1107 max_per<0.4 set periodicity to 0.0
1108 */
1109
1110 if (max_perSquare > 7868) { /* periodicity > 0.7 (0.7^4=0.2401 in Q15) */
1111 pitchfact = 32767;
1112 } else if (max_perSquare > 839) { /* 0.4 < periodicity < 0.7 (0.4^4=0.0256 in Q15) */
1113 /* find best index and interpolate from that */
1114 ind = 5;
1115 while ((max_perSquare < kPlcPerSqr[ind]) && (ind > 0)) {
1116 ind--;
1117 }
1118 /* pitch fact is approximated by first order */
1119 tmpW32 = kPlcPitchFact[ind] + SPL_MUL_16_16_RSFT(kPlcPfSlope[ind], (max_perSquare - kPlcPerSqr[ind]), 11);
1120
1121 pitchfact = FFMIN(tmpW32, 32767); /* guard against overflow */
1122
1123 } else { /* periodicity < 0.4 */
1124 pitchfact = 0;
1125 }
1126
1127 /* avoid repetition of same pitch cycle (buzzyness) */
1128 use_lag = lag;
1129 if (lag < 80) {
1130 use_lag = 2 * lag;
1131 }
1132
1133 /* compute concealed residual */
1134 energy = 0;
1135
1136 for (i = 0; i < s->block_samples; i++) {
1137 /* noise component - 52 < randlagFIX < 117 */
1138 s->seed = SPL_MUL_16_16(s->seed, 31821) + 13849;
1139 randlag = 53 + (s->seed & 63);
1140
1141 pick = i - randlag;
1142
1143 if (pick < 0) {
1144 randvec[i] = s->prevResidual[s->block_samples + pick];
1145 } else {
1146 randvec[i] = s->prevResidual[pick];
1147 }
1148
1149 /* pitch repeatition component */
1150 pick = i - use_lag;
1151
1152 if (pick < 0) {
1153 plc_residual[i] = s->prevResidual[s->block_samples + pick];
1154 } else {
1155 plc_residual[i] = plc_residual[pick];
1156 }
1157
1158 /* Attinuate total gain for each 10 ms */
1159 if (i < 80) {
1160 tot_gain = use_gain;
1161 } else if (i < 160) {
1162 tot_gain = SPL_MUL_16_16_RSFT(31130, use_gain, 15); /* 0.95*use_gain */
1163 } else {
1164 tot_gain = SPL_MUL_16_16_RSFT(29491, use_gain, 15); /* 0.9*use_gain */
1165 }
1166
1167 /* mix noise and pitch repeatition */
1168 plc_residual[i] = SPL_MUL_16_16_RSFT(tot_gain, (pitchfact * plc_residual[i] + (32767 - pitchfact) * randvec[i] + 16384) >> 15, 15);
1169
1170 /* Shifting down the result one step extra to ensure that no overflow
1171 will occur */
1172 energy += SPL_MUL_16_16_RSFT(plc_residual[i], plc_residual[i], (s->prevScale + 1));
1173
1174 }
1175
1176 /* less than 30 dB, use only noise */
1177 if (energy < SPL_SHIFT_W32(s->block_samples * 900, -s->prevScale - 1)) {
1178 energy = 0;
1179 for (i = 0; i < s->block_samples; i++) {
1180 plc_residual[i] = randvec[i];
1181 }
1182 }
1183
1184 /* use the old LPC */
1185 memcpy(plc_lpc, (*s).prev_lpc, (LPC_FILTERORDER + 1) * 2);
1186
1187 /* Update state in case there are multiple frame losses */
1188 s->prevLag = lag;
1189 s->per_square = max_perSquare;
1190 } else { /* no packet loss, copy input */
1191 memcpy(plc_residual, decresidual, s->block_samples * 2);
1192 memcpy(plc_lpc, lpc, (LPC_FILTERORDER + 1) * 2);
1193 s->consPLICount = 0;
1194 }
1195
1196 /* update state */
1197 s->prevPLI = PLI;
1198 memcpy(s->prev_lpc, plc_lpc, (LPC_FILTERORDER + 1) * 2);
1199 memcpy(s->prevResidual, plc_residual, s->block_samples * 2);
1200
1201 return;
1202 }
1203
1204 static int xcorr_coeff(int16_t *target, int16_t *regressor,
1205 int16_t subl, int16_t searchLen,
1206 int16_t offset, int16_t step)
1207 {
1208 int16_t maxlag;
1209 int16_t pos;
1210 int16_t max;
1211 int16_t cross_corr_scale, energy_scale;
1212 int16_t cross_corr_sg_mod, cross_corr_sg_mod_max;
1213 int32_t cross_corr, energy;
1214 int16_t cross_corr_mod, energy_mod, enery_mod_max;
1215 int16_t *tp, *rp;
1216 int16_t *rp_beg, *rp_end;
1217 int16_t totscale, totscale_max;
1218 int16_t scalediff;
1219 int32_t new_crit, max_crit;
1220 int shifts;
1221 int k;
1222
1223 /* Initializations, to make sure that the first one is selected */
1224 cross_corr_sg_mod_max = 0;
1225 enery_mod_max = INT16_MAX;
1226 totscale_max = -500;
1227 maxlag = 0;
1228 pos = 0;
1229
1230 /* Find scale value and start position */
1231 if (step == 1) {
1232 max = max_abs_value_w16(regressor, (int16_t) (subl + searchLen - 1));
1233 rp_beg = regressor;
1234 rp_end = &regressor[subl];
1235 } else { /* step== -1 */
1236 max = max_abs_value_w16(&regressor[-searchLen], (int16_t) (subl + searchLen - 1));
1237 rp_beg = &regressor[-1];
1238 rp_end = &regressor[subl - 1];
1239 }
1240
1241 /* Introduce a scale factor on the energy in int32_t in
1242 order to make sure that the calculation does not
1243 overflow */
1244
1245 if (max > 5000) {
1246 shifts = 2;
1247 } else {
1248 shifts = 0;
1249 }
1250
1251 /* Calculate the first energy, then do a +/- to get the other energies */
1252 energy = scale_dot_product(regressor, regressor, subl, shifts);
1253
1254 for (k = 0; k < searchLen; k++) {
1255 tp = target;
1256 rp = &regressor[pos];
1257
1258 cross_corr = scale_dot_product(tp, rp, subl, shifts);
1259
1260 if ((energy > 0) && (cross_corr > 0)) {
1261 /* Put cross correlation and energy on 16 bit word */
1262 cross_corr_scale = norm_w32(cross_corr) - 16;
1263 cross_corr_mod = (int16_t) SPL_SHIFT_W32(cross_corr, cross_corr_scale);
1264 energy_scale = norm_w32(energy) - 16;
1265 energy_mod = (int16_t) SPL_SHIFT_W32(energy, energy_scale);
1266
1267 /* Square cross correlation and store upper int16_t */
1268 cross_corr_sg_mod = (int16_t) SPL_MUL_16_16_RSFT(cross_corr_mod, cross_corr_mod, 16);
1269
1270 /* Calculate the total number of (dynamic) right shifts that have
1271 been performed on (cross_corr*cross_corr)/energy
1272 */
1273 totscale = energy_scale - (cross_corr_scale * 2);
1274
1275 /* Calculate the shift difference in order to be able to compare the two
1276 (cross_corr*cross_corr)/energy in the same domain
1277 */
1278 scalediff = totscale - totscale_max;
1279 scalediff = FFMIN(scalediff, 31);
1280 scalediff = FFMAX(scalediff, -31);
1281
1282 /* Compute the cross multiplication between the old best criteria
1283 and the new one to be able to compare them without using a
1284 division */
1285
1286 if (scalediff < 0) {
1287 new_crit = ((int32_t) cross_corr_sg_mod * enery_mod_max) >> (-scalediff);
1288 max_crit = ((int32_t) cross_corr_sg_mod_max * energy_mod);
1289 } else {
1290 new_crit = ((int32_t) cross_corr_sg_mod * enery_mod_max);
1291 max_crit = ((int32_t) cross_corr_sg_mod_max * energy_mod) >> scalediff;
1292 }
1293
1294 /* Store the new lag value if the new criteria is larger
1295 than previous largest criteria */
1296
1297 if (new_crit > max_crit) {
1298 cross_corr_sg_mod_max = cross_corr_sg_mod;
1299 enery_mod_max = energy_mod;
1300 totscale_max = totscale;
1301 maxlag = k;
1302 }
1303 }
1304 pos += step;
1305
1306 /* Do a +/- to get the next energy */
1307 energy += (unsigned)step * ((*rp_end * *rp_end - *rp_beg * *rp_beg) >> shifts);
1308
1309 rp_beg += step;
1310 rp_end += step;
1311 }
1312
1313 return maxlag + offset;
1314 }
1315
1316 static void hp_output(int16_t *signal, const int16_t *ba, int16_t *y,
1317 int16_t *x, int16_t len)
1318 {
1319 int32_t tmp;
1320
1321 for (int i = 0; i < len; i++) {
1322 tmp = SPL_MUL_16_16(y[1], ba[3]); /* (-a[1])*y[i-1] (low part) */
1323 tmp += SPL_MUL_16_16(y[3], ba[4]); /* (-a[2])*y[i-2] (low part) */
1324 tmp = (tmp >> 15);
1325 tmp += SPL_MUL_16_16(y[0], ba[3]); /* (-a[1])*y[i-1] (high part) */
1326 tmp += SPL_MUL_16_16(y[2], ba[4]); /* (-a[2])*y[i-2] (high part) */
1327 tmp = (tmp * 2);
1328
1329 tmp += SPL_MUL_16_16(signal[i], ba[0]); /* b[0]*x[0] */
1330 tmp += SPL_MUL_16_16(x[0], ba[1]); /* b[1]*x[i-1] */
1331 tmp += SPL_MUL_16_16(x[1], ba[2]); /* b[2]*x[i-2] */
1332
1333 /* Update state (input part) */
1334 x[1] = x[0];
1335 x[0] = signal[i];
1336
1337 /* Convert back to Q0 and multiply with 2 */
1338 signal[i] = av_clip_intp2(tmp + 1024, 26) >> 11;
1339
1340 /* Update state (filtered part) */
1341 y[2] = y[0];
1342 y[3] = y[1];
1343
1344 /* upshift tmp by 3 with saturation */
1345 if (tmp > 268435455) {
1346 tmp = INT32_MAX;
1347 } else if (tmp < -268435456) {
1348 tmp = INT32_MIN;
1349 } else {
1350 tmp = tmp * 8;
1351 }
1352
1353 y[0] = tmp >> 16;
1354 y[1] = (tmp - (y[0] * (1 << 16))) >> 1;
1355 }
1356 }
1357
1358 static int ilbc_decode_frame(AVCodecContext *avctx, void *data,
1359 int *got_frame_ptr, AVPacket *avpkt)
1360 {
1361 const uint8_t *buf = avpkt->data;
1362 AVFrame *frame = data;
1363 ILBCContext *s = avctx->priv_data;
1364 int mode = s->mode, ret;
1365 int16_t *plc_data = &s->plc_residual[LPC_FILTERORDER];
1366
1367 if ((ret = init_get_bits8(&s->gb, buf, avpkt->size)) < 0)
1368 return ret;
1369 memset(&s->frame, 0, sizeof(ILBCFrame));
1370
1371 frame->nb_samples = s->block_samples;
1372 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
1373 return ret;
1374
1375 if (unpack_frame(s))
1376 mode = 0;
1377 if (s->frame.start < 1 || s->frame.start > 5)
1378 mode = 0;
1379
1380 if (mode) {
1381 index_conv(s->frame.cb_index);
1382
1383 lsf_dequantization(s->lsfdeq, s->frame.lsf, s->lpc_n);
1384 lsf_check_stability(s->lsfdeq, LPC_FILTERORDER, s->lpc_n);
1385 lsp_interpolate(s->syntdenum, s->weightdenum,
1386 s->lsfdeq, LPC_FILTERORDER, s);
1387 decode_residual(s, &s->frame, s->decresidual, s->syntdenum);
1388
1389 do_plc(s->plc_residual, s->plc_lpc, 0,
1390 s->decresidual, s->syntdenum + (LPC_FILTERORDER + 1) * (s->nsub - 1),
1391 s->last_lag, s);
1392
1393 memcpy(s->decresidual, s->plc_residual, s->block_samples * 2);
1394 }
1395
1396 if (s->enhancer) {
1397 /* TODO */
1398 } else {
1399 int16_t lag, i;
1400
1401 /* Find last lag (since the enhancer is not called to give this info) */
1402 if (s->mode == 20) {
1403 lag = xcorr_coeff(&s->decresidual[s->block_samples-60], &s->decresidual[s->block_samples-80],
1404 60, 80, 20, -1);
1405 } else {
1406 lag = xcorr_coeff(&s->decresidual[s->block_samples-ENH_BLOCKL],
1407 &s->decresidual[s->block_samples-ENH_BLOCKL-20],
1408 ENH_BLOCKL, 100, 20, -1);
1409 }
1410
1411 /* Store lag (it is needed if next packet is lost) */
1412 s->last_lag = lag;
1413
1414 /* copy data and run synthesis filter */
1415 memcpy(plc_data, s->decresidual, s->block_samples * 2);
1416
1417 /* Set up the filter state */
1418 memcpy(&plc_data[-LPC_FILTERORDER], s->syntMem, LPC_FILTERORDER * 2);
1419
1420 for (i = 0; i < s->nsub; i++) {
1421 filter_arfq12(plc_data+i*SUBL, plc_data+i*SUBL,
1422 s->syntdenum + i*(LPC_FILTERORDER + 1),
1423 LPC_FILTERORDER + 1, SUBL);
1424 }
1425
1426 /* Save the filter state */
1427 memcpy(s->syntMem, &plc_data[s->block_samples-LPC_FILTERORDER], LPC_FILTERORDER * 2);
1428 }
1429
1430 memcpy(frame->data[0], plc_data, s->block_samples * 2);
1431
1432 hp_output((int16_t *)frame->data[0], hp_out_coeffs,
1433 s->hpimemy, s->hpimemx, s->block_samples);
1434
1435 memcpy(s->old_syntdenum, s->syntdenum, s->nsub*(LPC_FILTERORDER + 1) * 2);
1436
1437 s->prev_enh_pl = 0;
1438 if (mode == 0)
1439 s->prev_enh_pl = 1;
1440
1441 *got_frame_ptr = 1;
1442
1443 return avpkt->size;
1444 }
1445
1446 static av_cold int ilbc_decode_init(AVCodecContext *avctx)
1447 {
1448 ILBCContext *s = avctx->priv_data;
1449
1450 if (avctx->block_align == 38)
1451 s->mode = 20;
1452 else if (avctx->block_align == 50)
1453 s->mode = 30;
1454 else if (avctx->bit_rate > 0)
1455 s->mode = avctx->bit_rate <= 14000 ? 30 : 20;
1456 else
1457 return AVERROR_INVALIDDATA;
1458
1459 avctx->channels = 1;
1460 avctx->channel_layout = AV_CH_LAYOUT_MONO;
1461 avctx->sample_rate = 8000;
1462 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
1463
1464 if (s->mode == 30) {
1465 s->block_samples = 240;
1466 s->nsub = NSUB_30MS;
1467 s->nasub = NASUB_30MS;
1468 s->lpc_n = LPC_N_30MS;
1469 s->state_short_len = STATE_SHORT_LEN_30MS;
1470 } else {
1471 s->block_samples = 160;
1472 s->nsub = NSUB_20MS;
1473 s->nasub = NASUB_20MS;
1474 s->lpc_n = LPC_N_20MS;
1475 s->state_short_len = STATE_SHORT_LEN_20MS;
1476 }
1477
1478 return 0;
1479 }
1480
1481 const AVCodec ff_ilbc_decoder = {
1482 .name = "ilbc",
1483 .long_name = NULL_IF_CONFIG_SMALL("iLBC (Internet Low Bitrate Codec)"),
1484 .type = AVMEDIA_TYPE_AUDIO,
1485 .id = AV_CODEC_ID_ILBC,
1486 .init = ilbc_decode_init,
1487 .decode = ilbc_decode_frame,
1488 .capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF,
1489 .priv_data_size = sizeof(ILBCContext),
1490 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE,
1491 };
1492