FFmpeg coverage


Directory: ../../../ffmpeg/
File: src/libavcodec/ilbcdec.c
Date: 2024-11-20 23:03:26
Exec Total Coverage
Lines: 0 681 0.0%
Functions: 0 34 0.0%
Branches: 0 274 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 "codec_internal.h"
36 #include "decode.h"
37 #include "get_bits.h"
38 #include "ilbcdata.h"
39
40 #define LPC_N_20MS 1
41 #define LPC_N_30MS 2
42 #define LPC_N_MAX 2
43 #define LSF_NSPLIT 3
44 #define NASUB_MAX 4
45 #define LPC_FILTERORDER 10
46 #define NSUB_MAX 6
47 #define SUBL 40
48
49 #define ST_MEM_L_TBL 85
50 #define MEM_LF_TBL 147
51 #define STATE_SHORT_LEN_20MS 57
52 #define STATE_SHORT_LEN_30MS 58
53
54 #define BLOCKL_MAX 240
55 #define CB_MEML 147
56 #define CB_NSTAGES 3
57 #define CB_HALFFILTERLEN 4
58 #define CB_FILTERLEN 8
59
60 #define ENH_NBLOCKS_TOT 8
61 #define ENH_BLOCKL 80
62 #define ENH_BUFL (ENH_NBLOCKS_TOT)*ENH_BLOCKL
63 #define ENH_BUFL_FILTEROVERHEAD 3
64 #define BLOCKL_MAX 240
65 #define NSUB_20MS 4
66 #define NSUB_30MS 6
67 #define NSUB_MAX 6
68 #define NASUB_20MS 2
69 #define NASUB_30MS 4
70 #define NASUB_MAX 4
71 #define STATE_LEN 80
72 #define STATE_SHORT_LEN_30MS 58
73 #define STATE_SHORT_LEN_20MS 57
74
75 #define SPL_MUL_16_16(a, b) ((int32_t) (((int16_t)(a)) * ((int16_t)(b))))
76 #define SPL_MUL_16_16_RSFT(a, b, c) (SPL_MUL_16_16(a, b) >> (c))
77
78 typedef struct ILBCFrame {
79 int16_t lsf[LSF_NSPLIT*LPC_N_MAX];
80 int16_t cb_index[CB_NSTAGES*(NASUB_MAX + 1)];
81 int16_t gain_index[CB_NSTAGES*(NASUB_MAX + 1)];
82 int16_t ifm;
83 int16_t state_first;
84 int16_t idx[STATE_SHORT_LEN_30MS];
85 int16_t firstbits;
86 int16_t start;
87 } ILBCFrame;
88
89 typedef struct ILBCContext {
90 AVClass *class;
91 int enhancer;
92
93 int mode;
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, const uint8_t *buf, int size)
130 {
131 ILBCFrame *frame = &s->frame;
132 GetBitContext gb0, *const gb = &gb0;
133 int j, ret;
134
135 if ((ret = init_get_bits8(gb, buf, size)) < 0)
136 return ret;
137
138 frame->lsf[0] = get_bits(gb, 6);
139 frame->lsf[1] = get_bits(gb, 7);
140 frame->lsf[2] = get_bits(gb, 7);
141
142 if (s->mode == 20) {
143 frame->start = get_bits(gb, 2);
144 frame->state_first = get_bits1(gb);
145 frame->ifm = get_bits(gb, 6);
146 frame->cb_index[0] = get_bits(gb, 6) << 1;
147 frame->gain_index[0] = get_bits(gb, 2) << 3;
148 frame->gain_index[1] = get_bits1(gb) << 3;
149 frame->cb_index[3] = get_bits(gb, 7) << 1;
150 frame->gain_index[3] = get_bits1(gb) << 4;
151 frame->gain_index[4] = get_bits1(gb) << 3;
152 frame->gain_index[6] = get_bits1(gb) << 4;
153 } else {
154 frame->lsf[3] = get_bits(gb, 6);
155 frame->lsf[4] = get_bits(gb, 7);
156 frame->lsf[5] = get_bits(gb, 7);
157 frame->start = get_bits(gb, 3);
158 frame->state_first = get_bits1(gb);
159 frame->ifm = get_bits(gb, 6);
160 frame->cb_index[0] = get_bits(gb, 4) << 3;
161 frame->gain_index[0] = get_bits1(gb) << 4;
162 frame->gain_index[1] = get_bits1(gb) << 3;
163 frame->cb_index[3] = get_bits(gb, 6) << 2;
164 frame->gain_index[3] = get_bits1(gb) << 4;
165 frame->gain_index[4] = get_bits1(gb) << 3;
166 }
167
168 for (j = 0; j < 48; j++)
169 frame->idx[j] = get_bits1(gb) << 2;
170
171 if (s->mode == 20) {
172 for (; j < 57; j++)
173 frame->idx[j] = get_bits1(gb) << 2;
174
175 frame->gain_index[1] |= get_bits1(gb) << 2;
176 frame->gain_index[3] |= get_bits(gb, 2) << 2;
177 frame->gain_index[4] |= get_bits1(gb) << 2;
178 frame->gain_index[6] |= get_bits1(gb) << 3;
179 frame->gain_index[7] = get_bits(gb, 2) << 2;
180 } else {
181 for (; j < 58; j++)
182 frame->idx[j] = get_bits1(gb) << 2;
183
184 frame->cb_index[0] |= get_bits(gb, 2) << 1;
185 frame->gain_index[0] |= get_bits1(gb) << 3;
186 frame->gain_index[1] |= get_bits1(gb) << 2;
187 frame->cb_index[3] |= get_bits1(gb) << 1;
188 frame->cb_index[6] = get_bits1(gb) << 7;
189 frame->cb_index[6] |= get_bits(gb, 6) << 1;
190 frame->cb_index[9] = get_bits(gb, 7) << 1;
191 frame->cb_index[12] = get_bits(gb, 3) << 5;
192 frame->cb_index[12] |= get_bits(gb, 4) << 1;
193 frame->gain_index[3] |= get_bits(gb, 2) << 2;
194 frame->gain_index[4] |= get_bits(gb, 2) << 1;
195 frame->gain_index[6] = get_bits(gb, 2) << 3;
196 frame->gain_index[7] = get_bits(gb, 2) << 2;
197 frame->gain_index[9] = get_bits1(gb) << 4;
198 frame->gain_index[10] = get_bits1(gb) << 3;
199 frame->gain_index[12] = get_bits1(gb) << 4;
200 frame->gain_index[13] = get_bits1(gb) << 3;
201 }
202
203 for (j = 0; j < 56; j++)
204 frame->idx[j] |= get_bits(gb, 2);
205
206 if (s->mode == 20) {
207 frame->idx[56] |= get_bits(gb, 2);
208 frame->cb_index[0] |= get_bits1(gb);
209 frame->cb_index[1] = get_bits(gb, 7);
210 frame->cb_index[2] = get_bits(gb, 6) << 1;
211 frame->cb_index[2] |= get_bits1(gb);
212 frame->gain_index[0] |= get_bits(gb, 3);
213 frame->gain_index[1] |= get_bits(gb, 2);
214 frame->gain_index[2] = get_bits(gb, 3);
215 frame->cb_index[3] |= get_bits1(gb);
216 frame->cb_index[4] = get_bits(gb, 6) << 1;
217 frame->cb_index[4] |= get_bits1(gb);
218 frame->cb_index[5] = get_bits(gb, 7);
219 frame->cb_index[6] = get_bits(gb, 8);
220 frame->cb_index[7] = get_bits(gb, 8);
221 frame->cb_index[8] = get_bits(gb, 8);
222 frame->gain_index[3] |= get_bits(gb, 2);
223 frame->gain_index[4] |= get_bits(gb, 2);
224 frame->gain_index[5] = get_bits(gb, 3);
225 frame->gain_index[6] |= get_bits(gb, 3);
226 frame->gain_index[7] |= get_bits(gb, 2);
227 frame->gain_index[8] = get_bits(gb, 3);
228 } else {
229 frame->idx[56] |= get_bits(gb, 2);
230 frame->idx[57] |= get_bits(gb, 2);
231 frame->cb_index[0] |= get_bits1(gb);
232 frame->cb_index[1] = get_bits(gb, 7);
233 frame->cb_index[2] = get_bits(gb, 4) << 3;
234 frame->cb_index[2] |= get_bits(gb, 3);
235 frame->gain_index[0] |= get_bits(gb, 3);
236 frame->gain_index[1] |= get_bits(gb, 2);
237 frame->gain_index[2] = get_bits(gb, 3);
238 frame->cb_index[3] |= get_bits1(gb);
239 frame->cb_index[4] = get_bits(gb, 4) << 3;
240 frame->cb_index[4] |= get_bits(gb, 3);
241 frame->cb_index[5] = get_bits(gb, 7);
242 frame->cb_index[6] |= get_bits1(gb);
243 frame->cb_index[7] = get_bits(gb, 5) << 3;
244 frame->cb_index[7] |= get_bits(gb, 3);
245 frame->cb_index[8] = get_bits(gb, 8);
246 frame->cb_index[9] |= get_bits1(gb);
247 frame->cb_index[10] = get_bits(gb, 4) << 4;
248 frame->cb_index[10] |= get_bits(gb, 4);
249 frame->cb_index[11] = get_bits(gb, 8);
250 frame->cb_index[12] |= get_bits1(gb);
251 frame->cb_index[13] = get_bits(gb, 3) << 5;
252 frame->cb_index[13] |= get_bits(gb, 5);
253 frame->cb_index[14] = get_bits(gb, 8);
254 frame->gain_index[3] |= get_bits(gb, 2);
255 frame->gain_index[4] |= get_bits1(gb);
256 frame->gain_index[5] = get_bits(gb, 3);
257 frame->gain_index[6] |= get_bits(gb, 3);
258 frame->gain_index[7] |= get_bits(gb, 2);
259 frame->gain_index[8] = get_bits(gb, 3);
260 frame->gain_index[9] |= get_bits(gb, 4);
261 frame->gain_index[10] |= get_bits1(gb) << 2;
262 frame->gain_index[10] |= get_bits(gb, 2);
263 frame->gain_index[11] = get_bits(gb, 3);
264 frame->gain_index[12] |= get_bits(gb, 4);
265 frame->gain_index[13] |= get_bits(gb, 3);
266 frame->gain_index[14] = get_bits(gb, 3);
267 }
268
269 return get_bits1(gb);
270 }
271
272 static void index_conv(int16_t *index)
273 {
274 int k;
275
276 for (k = 4; k < 6; k++) {
277 if (index[k] >= 44 && index[k] < 108) {
278 index[k] += 64;
279 } else if (index[k] >= 108 && index[k] < 128) {
280 index[k] += 128;
281 }
282 }
283 }
284
285 static void lsf_dequantization(int16_t *lsfdeq, int16_t *index, int16_t lpc_n)
286 {
287 int i, j, pos = 0, cb_pos = 0;
288
289 for (i = 0; i < LSF_NSPLIT; i++) {
290 for (j = 0; j < lsf_dim_codebook[i]; j++) {
291 lsfdeq[pos + j] = lsf_codebook[cb_pos + index[i] * lsf_dim_codebook[i] + j];
292 }
293
294 pos += lsf_dim_codebook[i];
295 cb_pos += lsf_size_codebook[i] * lsf_dim_codebook[i];
296 }
297
298 if (lpc_n > 1) {
299 pos = 0;
300 cb_pos = 0;
301 for (i = 0; i < LSF_NSPLIT; i++) {
302 for (j = 0; j < lsf_dim_codebook[i]; j++) {
303 lsfdeq[LPC_FILTERORDER + pos + j] = lsf_codebook[cb_pos +
304 index[LSF_NSPLIT + i] * lsf_dim_codebook[i] + j];
305 }
306
307 pos += lsf_dim_codebook[i];
308 cb_pos += lsf_size_codebook[i] * lsf_dim_codebook[i];
309 }
310 }
311 }
312
313 static void lsf_check_stability(int16_t *lsf, int dim, int nb_vectors)
314 {
315 for (int n = 0; n < 2; n++) {
316 for (int m = 0; m < nb_vectors; m++) {
317 for (int k = 0; k < dim - 1; k++) {
318 int i = m * dim + k;
319
320 if ((lsf[i + 1] - lsf[i]) < 319) {
321 if (lsf[i + 1] < lsf[i]) {
322 lsf[i + 1] = lsf[i] + 160;
323 lsf[i] = lsf[i + 1] - 160;
324 } else {
325 lsf[i] -= 160;
326 lsf[i + 1] += 160;
327 }
328 }
329
330 lsf[i] = av_clip(lsf[i], 82, 25723);
331 }
332 }
333 }
334 }
335
336 static void lsf_interpolate(int16_t *out, const int16_t *in1,
337 const int16_t *in2, int16_t coef,
338 int size)
339 {
340 int invcoef = 16384 - coef, i;
341
342 for (i = 0; i < size; i++)
343 out[i] = (coef * in1[i] + invcoef * in2[i] + 8192) >> 14;
344 }
345
346 static void lsf2lsp(const int16_t *lsf, int16_t *lsp, int order)
347 {
348 int16_t diff, freq;
349 int32_t tmp;
350 int i, k;
351
352 for (i = 0; i < order; i++) {
353 freq = (lsf[i] * 20861) >> 15;
354 /* 20861: 1.0/(2.0*PI) in Q17 */
355 /*
356 Upper 8 bits give the index k and
357 Lower 8 bits give the difference, which needs
358 to be approximated linearly
359 */
360 k = FFMIN(freq >> 8, 63);
361 diff = freq & 0xFF;
362
363 /* Calculate linear approximation */
364 tmp = cos_derivative_tbl[k] * diff;
365 lsp[i] = cos_tbl[k] + (tmp >> 12);
366 }
367 }
368
369 static void get_lsp_poly(const int16_t *lsp, int32_t *f)
370 {
371 int16_t high, low;
372 int i, j, k, l;
373 int32_t tmp;
374
375 f[0] = 16777216;
376 f[1] = lsp[0] * -1024;
377
378 for (i = 2, k = 2, l = 2; i <= 5; i++, k += 2) {
379 f[l] = f[l - 2];
380
381 for (j = i; j > 1; j--, l--) {
382 high = f[l - 1] >> 16;
383 low = (f[l - 1] - (high * (1 << 16))) >> 1;
384
385 tmp = ((high * lsp[k]) * 4) + (((low * lsp[k]) >> 15) * 4);
386
387 f[l] += f[l - 2];
388 f[l] -= (unsigned)tmp;
389 }
390
391 f[l] -= lsp[k] * (1 << 10);
392 l += i;
393 }
394 }
395
396 static void lsf2poly(int16_t *a, const int16_t *lsf)
397 {
398 int32_t f[2][6];
399 int16_t lsp[10];
400 int32_t tmp;
401 int i;
402
403 lsf2lsp(lsf, lsp, LPC_FILTERORDER);
404
405 get_lsp_poly(&lsp[0], f[0]);
406 get_lsp_poly(&lsp[1], f[1]);
407
408 for (i = 5; i > 0; i--) {
409 f[0][i] += (unsigned)f[0][i - 1];
410 f[1][i] -= (unsigned)f[1][i - 1];
411 }
412
413 a[0] = 4096;
414 for (i = 5; i > 0; i--) {
415 tmp = f[0][6 - i] + (unsigned)f[1][6 - i] + 4096;
416 a[6 - i] = tmp >> 13;
417
418 tmp = f[0][6 - i] - (unsigned)f[1][6 - i] + 4096;
419 a[5 + i] = tmp >> 13;
420 }
421 }
422
423 static void lsp_interpolate2polydec(int16_t *a, const int16_t *lsf1,
424 const int16_t *lsf2, int coef, int length)
425 {
426 int16_t lsftmp[LPC_FILTERORDER];
427
428 lsf_interpolate(lsftmp, lsf1, lsf2, coef, length);
429 lsf2poly(a, lsftmp);
430 }
431
432 static void bw_expand(int16_t *out, const int16_t *in, const int16_t *coef, int length)
433 {
434 int i;
435
436 out[0] = in[0];
437 for (i = 1; i < length; i++)
438 out[i] = (coef[i] * in[i] + 16384) >> 15;
439 }
440
441 static void lsp_interpolate(int16_t *syntdenum, int16_t *weightdenum,
442 const int16_t *lsfdeq, int16_t length,
443 ILBCContext *s)
444 {
445 int16_t lp[LPC_FILTERORDER + 1];
446 const int16_t *const lsfdeq2 = lsfdeq + length;
447 int i, pos, lp_length;
448
449 lp_length = length + 1;
450
451 if (s->mode == 30) {
452 lsp_interpolate2polydec(lp, (*s).lsfdeqold, lsfdeq, lsf_weight_30ms[0], length);
453 memcpy(syntdenum, lp, lp_length * 2);
454 bw_expand(weightdenum, lp, kLpcChirpSyntDenum, lp_length);
455
456 pos = lp_length;
457 for (i = 1; i < 6; i++) {
458 lsp_interpolate2polydec(lp, lsfdeq, lsfdeq2,
459 lsf_weight_30ms[i],
460 length);
461 memcpy(syntdenum + pos, lp, lp_length * 2);
462 bw_expand(weightdenum + pos, lp, kLpcChirpSyntDenum, lp_length);
463 pos += lp_length;
464 }
465 } else {
466 pos = 0;
467 for (i = 0; i < s->nsub; i++) {
468 lsp_interpolate2polydec(lp, s->lsfdeqold, lsfdeq,
469 lsf_weight_20ms[i], length);
470 memcpy(syntdenum + pos, lp, lp_length * 2);
471 bw_expand(weightdenum + pos, lp, kLpcChirpSyntDenum, lp_length);
472 pos += lp_length;
473 }
474 }
475
476 if (s->mode == 30) {
477 memcpy(s->lsfdeqold, lsfdeq2, length * 2);
478 } else {
479 memcpy(s->lsfdeqold, lsfdeq, length * 2);
480 }
481 }
482
483 static void filter_mafq12(const int16_t *in_ptr, int16_t *out_ptr,
484 const int16_t *B, int16_t B_length,
485 int16_t length)
486 {
487 int o, i, j;
488
489 for (i = 0; i < length; i++) {
490 const int16_t *b_ptr = &B[0];
491 const int16_t *x_ptr = &in_ptr[i];
492
493 o = 0;
494 for (j = 0; j < B_length; j++)
495 o += b_ptr[j] * *x_ptr--;
496
497 o = av_clip(o, -134217728, 134215679);
498
499 out_ptr[i] = ((o + 2048) >> 12);
500 }
501 }
502
503 static void filter_arfq12(const int16_t *data_in,
504 int16_t *data_out,
505 const int16_t *coefficients,
506 int coefficients_length,
507 int data_length)
508 {
509 int i, j;
510
511 for (i = 0; i < data_length; i++) {
512 int output = 0, sum = 0;
513
514 for (j = coefficients_length - 1; j > 0; j--) {
515 sum += (unsigned)(coefficients[j] * data_out[i - j]);
516 }
517
518 output = coefficients[0] * data_in[i] - (unsigned)sum;
519 output = av_clip(output, -134217728, 134215679);
520
521 data_out[i] = (output + 2048) >> 12;
522 }
523 }
524
525 static void state_construct(int16_t ifm, const int16_t *idx,
526 const int16_t *synt_denum, int16_t *Out_fix,
527 int16_t len)
528 {
529 int k;
530 int16_t maxVal;
531 int16_t *tmp1, *tmp3;
532 const int16_t *tmp2;
533 /* Stack based */
534 int16_t numerator[1 + LPC_FILTERORDER];
535 int16_t sampleValVec[2 * STATE_SHORT_LEN_30MS + LPC_FILTERORDER];
536 int16_t sampleMaVec[2 * STATE_SHORT_LEN_30MS + LPC_FILTERORDER];
537 int16_t *sampleVal = &sampleValVec[LPC_FILTERORDER];
538 int16_t *sampleMa = &sampleMaVec[LPC_FILTERORDER];
539 int16_t *sampleAr = &sampleValVec[LPC_FILTERORDER];
540
541 /* initialization of coefficients */
542
543 for (k = 0; k < LPC_FILTERORDER + 1; k++) {
544 numerator[k] = synt_denum[LPC_FILTERORDER - k];
545 }
546
547 /* decoding of the maximum value */
548
549 maxVal = frg_quant_mod[ifm];
550
551 /* decoding of the sample values */
552 tmp1 = sampleVal;
553 tmp2 = &idx[len - 1];
554
555 if (ifm < 37) {
556 for (k = 0; k < len; k++) {
557 /*the shifting is due to the Q13 in sq4_fixQ13[i], also the adding of 2097152 (= 0.5 << 22)
558 maxVal is in Q8 and result is in Q(-1) */
559 (*tmp1) = (int16_t) ((SPL_MUL_16_16(maxVal, ilbc_state[(*tmp2)]) + 2097152) >> 22);
560 tmp1++;
561 tmp2--;
562 }
563 } else if (ifm < 59) {
564 for (k = 0; k < len; k++) {
565 /*the shifting is due to the Q13 in sq4_fixQ13[i], also the adding of 262144 (= 0.5 << 19)
566 maxVal is in Q5 and result is in Q(-1) */
567 (*tmp1) = (int16_t) ((SPL_MUL_16_16(maxVal, ilbc_state[(*tmp2)]) + 262144) >> 19);
568 tmp1++;
569 tmp2--;
570 }
571 } else {
572 for (k = 0; k < len; k++) {
573 /*the shifting is due to the Q13 in sq4_fixQ13[i], also the adding of 65536 (= 0.5 << 17)
574 maxVal is in Q3 and result is in Q(-1) */
575 (*tmp1) = (int16_t) ((SPL_MUL_16_16(maxVal, ilbc_state[(*tmp2)]) + 65536) >> 17);
576 tmp1++;
577 tmp2--;
578 }
579 }
580
581 /* Set the rest of the data to zero */
582 memset(&sampleVal[len], 0, len * 2);
583
584 /* circular convolution with all-pass filter */
585
586 /* Set the state to zero */
587 memset(sampleValVec, 0, LPC_FILTERORDER * 2);
588
589 /* Run MA filter + AR filter */
590 filter_mafq12(sampleVal, sampleMa, numerator, LPC_FILTERORDER + 1, len + LPC_FILTERORDER);
591 memset(&sampleMa[len + LPC_FILTERORDER], 0, (len - LPC_FILTERORDER) * 2);
592 filter_arfq12(sampleMa, sampleAr, synt_denum, LPC_FILTERORDER + 1, 2 * len);
593
594 tmp1 = &sampleAr[len - 1];
595 tmp2 = &sampleAr[2 * len - 1];
596 tmp3 = Out_fix;
597 for (k = 0; k < len; k++) {
598 (*tmp3) = (*tmp1) + (*tmp2);
599 tmp1--;
600 tmp2--;
601 tmp3++;
602 }
603 }
604
605 static int16_t gain_dequantization(int index, int max_in, int stage)
606 {
607 int16_t scale = FFMAX(1638, FFABS(max_in));
608
609 return ((scale * ilbc_gain[stage][index]) + 8192) >> 14;
610 }
611
612 static void vector_rmultiplication(int16_t *out, const int16_t *in,
613 const int16_t *win,
614 int length, int shift)
615 {
616 for (int i = 0; i < length; i++)
617 out[i] = (in[i] * win[-i]) >> shift;
618 }
619
620 static void vector_multiplication(int16_t *out, const int16_t *in,
621 const int16_t *win, int length,
622 int shift)
623 {
624 for (int i = 0; i < length; i++)
625 out[i] = (in[i] * win[i]) >> shift;
626 }
627
628 static void add_vector_and_shift(int16_t *out, const int16_t *in1,
629 const int16_t *in2, int length,
630 int shift)
631 {
632 for (int i = 0; i < length; i++)
633 out[i] = (in1[i] + in2[i]) >> shift;
634 }
635
636 static void create_augmented_vector(int index, const int16_t *buffer, int16_t *cbVec)
637 {
638 int16_t cbVecTmp[4];
639 int interpolation_length = FFMIN(4, index);
640 int16_t ilow = index - interpolation_length;
641
642 memcpy(cbVec, buffer - index, index * 2);
643
644 vector_multiplication(&cbVec[ilow], buffer - index - interpolation_length, alpha, interpolation_length, 15);
645 vector_rmultiplication(cbVecTmp, buffer - interpolation_length, &alpha[interpolation_length - 1], interpolation_length, 15);
646 add_vector_and_shift(&cbVec[ilow], &cbVec[ilow], cbVecTmp, interpolation_length, 0);
647
648 memcpy(cbVec + index, buffer - index, FFMIN(SUBL - index, index) * sizeof(*cbVec));
649 }
650
651 static void get_codebook(int16_t * cbvec, /* (o) Constructed codebook vector */
652 int16_t * mem, /* (i) Codebook buffer */
653 int16_t index, /* (i) Codebook index */
654 int16_t lMem, /* (i) Length of codebook buffer */
655 int16_t cbveclen /* (i) Codebook vector length */
656 )
657 {
658 int16_t k, base_size;
659 int16_t lag;
660 /* Stack based */
661 int16_t tempbuff2[SUBL + 5];
662
663 /* Determine size of codebook sections */
664 base_size = lMem - cbveclen + 1;
665
666 if (cbveclen == SUBL) {
667 base_size += cbveclen / 2;
668 }
669
670 /* No filter -> First codebook section */
671 if (index < lMem - cbveclen + 1) {
672 /* first non-interpolated vectors */
673
674 k = index + cbveclen;
675 /* get vector */
676 memcpy(cbvec, mem + lMem - k, cbveclen * 2);
677 } else if (index < base_size) {
678
679 /* Calculate lag */
680
681 k = (int16_t) SPL_MUL_16_16(2, (index - (lMem - cbveclen + 1))) + cbveclen;
682
683 lag = k / 2;
684
685 create_augmented_vector(lag, mem + lMem, cbvec);
686 } else {
687 int16_t memIndTest;
688
689 /* first non-interpolated vectors */
690
691 if (index - base_size < lMem - cbveclen + 1) {
692
693 /* Set up filter memory, stuff zeros outside memory buffer */
694
695 memIndTest = lMem - (index - base_size + cbveclen);
696
697 memset(mem - CB_HALFFILTERLEN, 0, CB_HALFFILTERLEN * 2);
698 memset(mem + lMem, 0, CB_HALFFILTERLEN * 2);
699
700 /* do filtering to get the codebook vector */
701
702 filter_mafq12(&mem[memIndTest + 4], cbvec, kCbFiltersRev, CB_FILTERLEN, cbveclen);
703 } else {
704 /* interpolated vectors */
705 /* Stuff zeros outside memory buffer */
706 memIndTest = lMem - cbveclen - CB_FILTERLEN;
707 memset(mem + lMem, 0, CB_HALFFILTERLEN * 2);
708
709 /* do filtering */
710 filter_mafq12(&mem[memIndTest + 7], tempbuff2, kCbFiltersRev, CB_FILTERLEN, (int16_t) (cbveclen + 5));
711
712 /* Calculate lag index */
713 lag = (cbveclen << 1) - 20 + index - base_size - lMem - 1;
714
715 create_augmented_vector(lag, tempbuff2 + SUBL + 5, cbvec);
716 }
717 }
718 }
719
720 static void construct_vector (
721 int16_t *decvector, /* (o) Decoded vector */
722 const int16_t *index, /* (i) Codebook indices */
723 const int16_t *gain_index, /* (i) Gain quantization indices */
724 int16_t *mem, /* (i) Buffer for codevector construction */
725 int16_t lMem, /* (i) Length of buffer */
726 int16_t veclen)
727 {
728 int16_t gain[CB_NSTAGES];
729 int16_t cbvec0[SUBL];
730 int16_t cbvec1[SUBL];
731 int16_t cbvec2[SUBL];
732 unsigned a32;
733 int16_t *gainPtr;
734 int j;
735
736 /* gain de-quantization */
737
738 gain[0] = gain_dequantization(gain_index[0], 16384, 0);
739 gain[1] = gain_dequantization(gain_index[1], gain[0], 1);
740 gain[2] = gain_dequantization(gain_index[2], gain[1], 2);
741
742 /* codebook vector construction and construction of total vector */
743
744 /* Stack based */
745 get_codebook(cbvec0, mem, index[0], lMem, veclen);
746 get_codebook(cbvec1, mem, index[1], lMem, veclen);
747 get_codebook(cbvec2, mem, index[2], lMem, veclen);
748
749 gainPtr = &gain[0];
750 for (j = 0; j < veclen; j++) {
751 a32 = SPL_MUL_16_16(*gainPtr++, cbvec0[j]);
752 a32 += SPL_MUL_16_16(*gainPtr++, cbvec1[j]);
753 a32 += SPL_MUL_16_16(*gainPtr, cbvec2[j]);
754 gainPtr -= 2;
755 decvector[j] = (int)(a32 + 8192) >> 14;
756 }
757 }
758
759 static void reverse_memcpy(int16_t *dest, const int16_t *source, int length)
760 {
761 int16_t* destPtr = dest;
762 const int16_t *sourcePtr = source;
763 int j;
764
765 for (j = 0; j < length; j++)
766 *destPtr-- = *sourcePtr++;
767 }
768
769 static void decode_residual(ILBCContext *s,
770 ILBCFrame *encbits,
771 int16_t *decresidual,
772 const int16_t *syntdenum)
773 {
774 int16_t meml_gotten, Nfor, Nback, diff, start_pos;
775 int16_t subcount, subframe;
776 int16_t *reverseDecresidual = s->enh_buf; /* Reversed decoded data, used for decoding backwards in time (reuse memory in state) */
777 int16_t *memVec = s->prevResidual;
778 int16_t *mem = &memVec[CB_HALFFILTERLEN]; /* Memory for codebook */
779
780 diff = STATE_LEN - s->state_short_len;
781
782 if (encbits->state_first == 1) {
783 start_pos = (encbits->start - 1) * SUBL;
784 } else {
785 start_pos = (encbits->start - 1) * SUBL + diff;
786 }
787
788 /* decode scalar part of start state */
789
790 state_construct(encbits->ifm, encbits->idx, &syntdenum[(encbits->start - 1) * (LPC_FILTERORDER + 1)], &decresidual[start_pos], s->state_short_len);
791
792 if (encbits->state_first) { /* put adaptive part in the end */
793 /* setup memory */
794 memset(mem, 0, (int16_t) (CB_MEML - s->state_short_len) * 2);
795 memcpy(mem + CB_MEML - s->state_short_len, decresidual + start_pos, s->state_short_len * 2);
796
797 /* construct decoded vector */
798
799 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);
800
801 } else { /* put adaptive part in the beginning */
802 /* setup memory */
803 meml_gotten = s->state_short_len;
804 reverse_memcpy(mem + CB_MEML - 1, decresidual + start_pos, meml_gotten);
805 memset(mem, 0, (int16_t) (CB_MEML - meml_gotten) * 2);
806
807 /* construct decoded vector */
808 construct_vector(reverseDecresidual, encbits->cb_index, encbits->gain_index, mem + CB_MEML - ST_MEM_L_TBL, ST_MEM_L_TBL, diff);
809
810 /* get decoded residual from reversed vector */
811 reverse_memcpy(&decresidual[start_pos - 1], reverseDecresidual, diff);
812 }
813
814 /* counter for predicted subframes */
815 subcount = 1;
816
817 /* forward prediction of subframes */
818 Nfor = s->nsub - encbits->start - 1;
819
820 if (Nfor > 0) {
821 /* setup memory */
822 memset(mem, 0, (CB_MEML - STATE_LEN) * 2);
823 memcpy(mem + CB_MEML - STATE_LEN, decresidual + (encbits->start - 1) * SUBL, STATE_LEN * 2);
824
825 /* loop over subframes to encode */
826 for (subframe = 0; subframe < Nfor; subframe++) {
827 /* construct decoded vector */
828 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);
829
830 /* update memory */
831 memmove(mem, mem + SUBL, (CB_MEML - SUBL) * sizeof(*mem));
832 memcpy(mem + CB_MEML - SUBL, &decresidual[(encbits->start + 1 + subframe) * SUBL], SUBL * 2);
833
834 subcount++;
835 }
836
837 }
838
839 /* backward prediction of subframes */
840 Nback = encbits->start - 1;
841
842 if (Nback > 0) {
843 /* setup memory */
844 meml_gotten = SUBL * (s->nsub + 1 - encbits->start);
845 if (meml_gotten > CB_MEML) {
846 meml_gotten = CB_MEML;
847 }
848
849 reverse_memcpy(mem + CB_MEML - 1, decresidual + (encbits->start - 1) * SUBL, meml_gotten);
850 memset(mem, 0, (int16_t) (CB_MEML - meml_gotten) * 2);
851
852 /* loop over subframes to decode */
853 for (subframe = 0; subframe < Nback; subframe++) {
854 /* construct decoded vector */
855 construct_vector(&reverseDecresidual[subframe * SUBL], encbits->cb_index + subcount * CB_NSTAGES,
856 encbits->gain_index + subcount * CB_NSTAGES, mem, MEM_LF_TBL, SUBL);
857
858 /* update memory */
859 memmove(mem, mem + SUBL, (CB_MEML - SUBL) * sizeof(*mem));
860 memcpy(mem + CB_MEML - SUBL, &reverseDecresidual[subframe * SUBL], SUBL * 2);
861
862 subcount++;
863 }
864
865 /* get decoded residual from reversed vector */
866 reverse_memcpy(decresidual + SUBL * Nback - 1, reverseDecresidual, SUBL * Nback);
867 }
868 }
869
870 static int16_t max_abs_value_w16(const int16_t* vector, int length)
871 {
872 int i = 0, absolute = 0, maximum = 0;
873
874 if (vector == NULL || length <= 0) {
875 return -1;
876 }
877
878 for (i = 0; i < length; i++) {
879 absolute = FFABS(vector[i]);
880 if (absolute > maximum)
881 maximum = absolute;
882 }
883
884 // Guard the case for abs(-32768).
885 return FFMIN(maximum, INT16_MAX);
886 }
887
888 static int16_t get_size_in_bits(uint32_t n)
889 {
890 int16_t bits;
891
892 if (0xFFFF0000 & n) {
893 bits = 16;
894 } else {
895 bits = 0;
896 }
897
898 if (0x0000FF00 & (n >> bits)) bits += 8;
899 if (0x000000F0 & (n >> bits)) bits += 4;
900 if (0x0000000C & (n >> bits)) bits += 2;
901 if (0x00000002 & (n >> bits)) bits += 1;
902 if (0x00000001 & (n >> bits)) bits += 1;
903
904 return bits;
905 }
906
907 static int32_t scale_dot_product(const int16_t *v1, const int16_t *v2, int length, int scaling)
908 {
909 int64_t sum = 0;
910
911 for (int i = 0; i < length; i++)
912 sum += (v1[i] * v2[i]) >> scaling;
913
914 return av_clipl_int32(sum);
915 }
916
917 static void correlation(int32_t *corr, int32_t *ener, const int16_t *buffer,
918 int16_t lag, int16_t blen, int16_t srange, int16_t scale)
919 {
920 const int16_t *w16ptr = &buffer[blen - srange - lag];
921
922 *corr = scale_dot_product(&buffer[blen - srange], w16ptr, srange, scale);
923 *ener = scale_dot_product(w16ptr, w16ptr, srange, scale);
924
925 if (*ener == 0) {
926 *corr = 0;
927 *ener = 1;
928 }
929 }
930
931 #define SPL_SHIFT_W32(x, c) (((c) >= 0) ? ((x) << (c)) : ((x) >> (-(c))))
932
933 static int16_t norm_w32(int32_t a)
934 {
935 if (a == 0) {
936 return 0;
937 } else if (a < 0) {
938 a = ~a;
939 }
940
941 return ff_clz(a);
942 }
943
944 static int32_t div_w32_w16(int32_t num, int16_t den)
945 {
946 if (den != 0)
947 return num / den;
948 else
949 return 0x7FFFFFFF;
950 }
951
952 static void do_plc(int16_t *plc_residual, /* (o) concealed residual */
953 int16_t *plc_lpc, /* (o) concealed LP parameters */
954 int16_t PLI, /* (i) packet loss indicator
955 0 - no PL, 1 = PL */
956 const int16_t *decresidual, /* (i) decoded residual */
957 const int16_t *lpc, /* (i) decoded LPC (only used for no PL) */
958 int16_t inlag, /* (i) pitch lag */
959 ILBCContext *s) /* (i/o) decoder instance */
960 {
961 int16_t i, pick;
962 int32_t cross, ener, cross_comp, ener_comp = 0;
963 int32_t measure, max_measure, energy;
964 int16_t max, cross_square_max, cross_square;
965 int16_t j, lag, tmp1, tmp2, randlag;
966 int16_t shift1, shift2, shift3, shift_max;
967 int16_t scale3;
968 int16_t corrLen;
969 int32_t tmpW32, tmp2W32;
970 int16_t use_gain;
971 int16_t tot_gain;
972 int16_t max_perSquare;
973 int16_t scale1, scale2;
974 int16_t totscale;
975 int32_t nom;
976 int16_t denom;
977 int16_t pitchfact;
978 int16_t use_lag;
979 int ind;
980 int16_t randvec[BLOCKL_MAX];
981
982 /* Packet Loss */
983 if (PLI == 1) {
984
985 s->consPLICount += 1;
986
987 /* if previous frame not lost,
988 determine pitch pred. gain */
989
990 if (s->prevPLI != 1) {
991
992 /* Maximum 60 samples are correlated, preserve as high accuracy
993 as possible without getting overflow */
994 max = max_abs_value_w16(s->prevResidual, s->block_samples);
995 scale3 = (get_size_in_bits(max) << 1) - 25;
996 if (scale3 < 0) {
997 scale3 = 0;
998 }
999
1000 /* Store scale for use when interpolating between the
1001 * concealment and the received packet */
1002 s->prevScale = scale3;
1003
1004 /* Search around the previous lag +/-3 to find the
1005 best pitch period */
1006 lag = inlag - 3;
1007
1008 /* Guard against getting outside the frame */
1009 corrLen = FFMIN(60, s->block_samples - (inlag + 3));
1010
1011 correlation(&cross, &ener, s->prevResidual, lag, s->block_samples, corrLen, scale3);
1012
1013 /* Normalize and store cross^2 and the number of shifts */
1014 shift_max = get_size_in_bits(FFABS(cross)) - 15;
1015 cross_square_max = (int16_t) SPL_MUL_16_16_RSFT(SPL_SHIFT_W32(cross, -shift_max), SPL_SHIFT_W32(cross, -shift_max), 15);
1016
1017 for (j = inlag - 2; j <= inlag + 3; j++) {
1018 correlation(&cross_comp, &ener_comp, s->prevResidual, j, s->block_samples, corrLen, scale3);
1019
1020 /* Use the criteria (corr*corr)/energy to compare if
1021 this lag is better or not. To avoid the division,
1022 do a cross multiplication */
1023 shift1 = get_size_in_bits(FFABS(cross_comp)) - 15;
1024 cross_square = (int16_t) SPL_MUL_16_16_RSFT(SPL_SHIFT_W32(cross_comp, -shift1), SPL_SHIFT_W32(cross_comp, -shift1), 15);
1025
1026 shift2 = get_size_in_bits(ener) - 15;
1027 measure = SPL_MUL_16_16(SPL_SHIFT_W32(ener, -shift2), cross_square);
1028
1029 shift3 = get_size_in_bits(ener_comp) - 15;
1030 max_measure = SPL_MUL_16_16(SPL_SHIFT_W32(ener_comp, -shift3), cross_square_max);
1031
1032 /* Calculate shift value, so that the two measures can
1033 be put in the same Q domain */
1034 if (((shift_max << 1) + shift3) > ((shift1 << 1) + shift2)) {
1035 tmp1 = FFMIN(31, (shift_max << 1) + shift3 - (shift1 << 1) - shift2);
1036 tmp2 = 0;
1037 } else {
1038 tmp1 = 0;
1039 tmp2 = FFMIN(31, (shift1 << 1) + shift2 - (shift_max << 1) - shift3);
1040 }
1041
1042 if ((measure >> tmp1) > (max_measure >> tmp2)) {
1043 /* New lag is better => record lag, measure and domain */
1044 lag = j;
1045 cross_square_max = cross_square;
1046 cross = cross_comp;
1047 shift_max = shift1;
1048 ener = ener_comp;
1049 }
1050 }
1051
1052 /* Calculate the periodicity for the lag with the maximum correlation.
1053
1054 Definition of the periodicity:
1055 abs(corr(vec1, vec2))/(sqrt(energy(vec1))*sqrt(energy(vec2)))
1056
1057 Work in the Square domain to simplify the calculations
1058 max_perSquare is less than 1 (in Q15)
1059 */
1060 tmp2W32 = scale_dot_product(&s->prevResidual[s->block_samples - corrLen], &s->prevResidual[s->block_samples - corrLen], corrLen, scale3);
1061
1062 if ((tmp2W32 > 0) && (ener_comp > 0)) {
1063 /* norm energies to int16_t, compute the product of the energies and
1064 use the upper int16_t as the denominator */
1065
1066 scale1 = norm_w32(tmp2W32) - 16;
1067 tmp1 = SPL_SHIFT_W32(tmp2W32, scale1);
1068
1069 scale2 = norm_w32(ener) - 16;
1070 tmp2 = SPL_SHIFT_W32(ener, scale2);
1071 denom = SPL_MUL_16_16_RSFT(tmp1, tmp2, 16); /* denom in Q(scale1+scale2-16) */
1072
1073 /* Square the cross correlation and norm it such that max_perSquare
1074 will be in Q15 after the division */
1075
1076 totscale = scale1 + scale2 - 1;
1077 tmp1 = SPL_SHIFT_W32(cross, (totscale >> 1));
1078 tmp2 = SPL_SHIFT_W32(cross, totscale - (totscale >> 1));
1079
1080 nom = SPL_MUL_16_16(tmp1, tmp2);
1081 max_perSquare = div_w32_w16(nom, denom);
1082 } else {
1083 max_perSquare = 0;
1084 }
1085 } else {
1086 /* previous frame lost, use recorded lag and gain */
1087 lag = s->prevLag;
1088 max_perSquare = s->per_square;
1089 }
1090
1091 /* Attenuate signal and scale down pitch pred gain if
1092 several frames lost consecutively */
1093
1094 use_gain = 32767; /* 1.0 in Q15 */
1095
1096 if (s->consPLICount * s->block_samples > 320) {
1097 use_gain = 29491; /* 0.9 in Q15 */
1098 }
1099
1100 /* Compute mixing factor of picth repeatition and noise:
1101 for max_per>0.7 set periodicity to 1.0
1102 0.4<max_per<0.7 set periodicity to (maxper-0.4)/0.7-0.4)
1103 max_per<0.4 set periodicity to 0.0
1104 */
1105
1106 if (max_perSquare > 7868) { /* periodicity > 0.7 (0.7^4=0.2401 in Q15) */
1107 pitchfact = 32767;
1108 } else if (max_perSquare > 839) { /* 0.4 < periodicity < 0.7 (0.4^4=0.0256 in Q15) */
1109 /* find best index and interpolate from that */
1110 ind = 5;
1111 while ((max_perSquare < kPlcPerSqr[ind]) && (ind > 0)) {
1112 ind--;
1113 }
1114 /* pitch fact is approximated by first order */
1115 tmpW32 = kPlcPitchFact[ind] + SPL_MUL_16_16_RSFT(kPlcPfSlope[ind], (max_perSquare - kPlcPerSqr[ind]), 11);
1116
1117 pitchfact = FFMIN(tmpW32, 32767); /* guard against overflow */
1118
1119 } else { /* periodicity < 0.4 */
1120 pitchfact = 0;
1121 }
1122
1123 /* avoid repetition of same pitch cycle (buzzyness) */
1124 use_lag = lag;
1125 if (lag < 80) {
1126 use_lag = 2 * lag;
1127 }
1128
1129 /* compute concealed residual */
1130 energy = 0;
1131
1132 for (i = 0; i < s->block_samples; i++) {
1133 /* noise component - 52 < randlagFIX < 117 */
1134 s->seed = SPL_MUL_16_16(s->seed, 31821) + 13849;
1135 randlag = 53 + (s->seed & 63);
1136
1137 pick = i - randlag;
1138
1139 if (pick < 0) {
1140 randvec[i] = s->prevResidual[s->block_samples + pick];
1141 } else {
1142 randvec[i] = s->prevResidual[pick];
1143 }
1144
1145 /* pitch repeatition component */
1146 pick = i - use_lag;
1147
1148 if (pick < 0) {
1149 plc_residual[i] = s->prevResidual[s->block_samples + pick];
1150 } else {
1151 plc_residual[i] = plc_residual[pick];
1152 }
1153
1154 /* Attinuate total gain for each 10 ms */
1155 if (i < 80) {
1156 tot_gain = use_gain;
1157 } else if (i < 160) {
1158 tot_gain = SPL_MUL_16_16_RSFT(31130, use_gain, 15); /* 0.95*use_gain */
1159 } else {
1160 tot_gain = SPL_MUL_16_16_RSFT(29491, use_gain, 15); /* 0.9*use_gain */
1161 }
1162
1163 /* mix noise and pitch repeatition */
1164 plc_residual[i] = SPL_MUL_16_16_RSFT(tot_gain, (pitchfact * plc_residual[i] + (32767 - pitchfact) * randvec[i] + 16384) >> 15, 15);
1165
1166 /* Shifting down the result one step extra to ensure that no overflow
1167 will occur */
1168 energy += SPL_MUL_16_16_RSFT(plc_residual[i], plc_residual[i], (s->prevScale + 1));
1169
1170 }
1171
1172 /* less than 30 dB, use only noise */
1173 if (energy < SPL_SHIFT_W32(s->block_samples * 900, -s->prevScale - 1)) {
1174 energy = 0;
1175 for (i = 0; i < s->block_samples; i++) {
1176 plc_residual[i] = randvec[i];
1177 }
1178 }
1179
1180 /* use the old LPC */
1181 memcpy(plc_lpc, (*s).prev_lpc, (LPC_FILTERORDER + 1) * 2);
1182
1183 /* Update state in case there are multiple frame losses */
1184 s->prevLag = lag;
1185 s->per_square = max_perSquare;
1186 } else { /* no packet loss, copy input */
1187 memcpy(plc_residual, decresidual, s->block_samples * 2);
1188 memcpy(plc_lpc, lpc, (LPC_FILTERORDER + 1) * 2);
1189 s->consPLICount = 0;
1190 }
1191
1192 /* update state */
1193 s->prevPLI = PLI;
1194 memcpy(s->prev_lpc, plc_lpc, (LPC_FILTERORDER + 1) * 2);
1195 memcpy(s->prevResidual, plc_residual, s->block_samples * 2);
1196
1197 return;
1198 }
1199
1200 static int xcorr_coeff(const int16_t *target, const int16_t *regressor,
1201 int16_t subl, int16_t searchLen,
1202 int16_t offset, int16_t step)
1203 {
1204 int16_t maxlag;
1205 int16_t pos;
1206 int16_t max;
1207 int16_t cross_corr_scale, energy_scale;
1208 int16_t cross_corr_sg_mod, cross_corr_sg_mod_max;
1209 int32_t cross_corr, energy;
1210 int16_t cross_corr_mod, energy_mod, enery_mod_max;
1211 const int16_t *rp;
1212 const int16_t *rp_beg, *rp_end;
1213 int16_t totscale, totscale_max;
1214 int16_t scalediff;
1215 int32_t new_crit, max_crit;
1216 int shifts;
1217 int k;
1218
1219 /* Initializations, to make sure that the first one is selected */
1220 cross_corr_sg_mod_max = 0;
1221 enery_mod_max = INT16_MAX;
1222 totscale_max = -500;
1223 maxlag = 0;
1224 pos = 0;
1225
1226 /* Find scale value and start position */
1227 if (step == 1) {
1228 max = max_abs_value_w16(regressor, (int16_t) (subl + searchLen - 1));
1229 rp_beg = regressor;
1230 rp_end = &regressor[subl];
1231 } else { /* step== -1 */
1232 max = max_abs_value_w16(&regressor[-searchLen], (int16_t) (subl + searchLen - 1));
1233 rp_beg = &regressor[-1];
1234 rp_end = &regressor[subl - 1];
1235 }
1236
1237 /* Introduce a scale factor on the energy in int32_t in
1238 order to make sure that the calculation does not
1239 overflow */
1240
1241 if (max > 5000) {
1242 shifts = 2;
1243 } else {
1244 shifts = 0;
1245 }
1246
1247 /* Calculate the first energy, then do a +/- to get the other energies */
1248 energy = scale_dot_product(regressor, regressor, subl, shifts);
1249
1250 for (k = 0; k < searchLen; k++) {
1251 rp = &regressor[pos];
1252
1253 cross_corr = scale_dot_product(target, rp, subl, shifts);
1254
1255 if ((energy > 0) && (cross_corr > 0)) {
1256 /* Put cross correlation and energy on 16 bit word */
1257 cross_corr_scale = norm_w32(cross_corr) - 16;
1258 cross_corr_mod = (int16_t) SPL_SHIFT_W32(cross_corr, cross_corr_scale);
1259 energy_scale = norm_w32(energy) - 16;
1260 energy_mod = (int16_t) SPL_SHIFT_W32(energy, energy_scale);
1261
1262 /* Square cross correlation and store upper int16_t */
1263 cross_corr_sg_mod = (int16_t) SPL_MUL_16_16_RSFT(cross_corr_mod, cross_corr_mod, 16);
1264
1265 /* Calculate the total number of (dynamic) right shifts that have
1266 been performed on (cross_corr*cross_corr)/energy
1267 */
1268 totscale = energy_scale - (cross_corr_scale * 2);
1269
1270 /* Calculate the shift difference in order to be able to compare the two
1271 (cross_corr*cross_corr)/energy in the same domain
1272 */
1273 scalediff = totscale - totscale_max;
1274 scalediff = FFMIN(scalediff, 31);
1275 scalediff = FFMAX(scalediff, -31);
1276
1277 /* Compute the cross multiplication between the old best criteria
1278 and the new one to be able to compare them without using a
1279 division */
1280
1281 if (scalediff < 0) {
1282 new_crit = ((int32_t) cross_corr_sg_mod * enery_mod_max) >> (-scalediff);
1283 max_crit = ((int32_t) cross_corr_sg_mod_max * energy_mod);
1284 } else {
1285 new_crit = ((int32_t) cross_corr_sg_mod * enery_mod_max);
1286 max_crit = ((int32_t) cross_corr_sg_mod_max * energy_mod) >> scalediff;
1287 }
1288
1289 /* Store the new lag value if the new criteria is larger
1290 than previous largest criteria */
1291
1292 if (new_crit > max_crit) {
1293 cross_corr_sg_mod_max = cross_corr_sg_mod;
1294 enery_mod_max = energy_mod;
1295 totscale_max = totscale;
1296 maxlag = k;
1297 }
1298 }
1299 pos += step;
1300
1301 /* Do a +/- to get the next energy */
1302 energy += (unsigned)step * ((*rp_end * *rp_end - *rp_beg * *rp_beg) >> shifts);
1303
1304 rp_beg += step;
1305 rp_end += step;
1306 }
1307
1308 return maxlag + offset;
1309 }
1310
1311 static void hp_output(int16_t *signal, const int16_t *ba, int16_t *y,
1312 int16_t *x, int16_t len)
1313 {
1314 int32_t tmp;
1315
1316 for (int i = 0; i < len; i++) {
1317 tmp = SPL_MUL_16_16(y[1], ba[3]); /* (-a[1])*y[i-1] (low part) */
1318 tmp += SPL_MUL_16_16(y[3], ba[4]); /* (-a[2])*y[i-2] (low part) */
1319 tmp = (tmp >> 15);
1320 tmp += SPL_MUL_16_16(y[0], ba[3]); /* (-a[1])*y[i-1] (high part) */
1321 tmp += SPL_MUL_16_16(y[2], ba[4]); /* (-a[2])*y[i-2] (high part) */
1322 tmp = (tmp * 2);
1323
1324 tmp += SPL_MUL_16_16(signal[i], ba[0]); /* b[0]*x[0] */
1325 tmp += SPL_MUL_16_16(x[0], ba[1]); /* b[1]*x[i-1] */
1326 tmp += SPL_MUL_16_16(x[1], ba[2]); /* b[2]*x[i-2] */
1327
1328 /* Update state (input part) */
1329 x[1] = x[0];
1330 x[0] = signal[i];
1331
1332 /* Convert back to Q0 and multiply with 2 */
1333 signal[i] = av_clip_intp2(tmp + 1024, 26) >> 11;
1334
1335 /* Update state (filtered part) */
1336 y[2] = y[0];
1337 y[3] = y[1];
1338
1339 /* upshift tmp by 3 with saturation */
1340 if (tmp > 268435455) {
1341 tmp = INT32_MAX;
1342 } else if (tmp < -268435456) {
1343 tmp = INT32_MIN;
1344 } else {
1345 tmp = tmp * 8;
1346 }
1347
1348 y[0] = tmp >> 16;
1349 y[1] = (tmp - (y[0] * (1 << 16))) >> 1;
1350 }
1351 }
1352
1353 static int ilbc_decode_frame(AVCodecContext *avctx, AVFrame *frame,
1354 int *got_frame_ptr, AVPacket *avpkt)
1355 {
1356 ILBCContext *s = avctx->priv_data;
1357 int mode = s->mode, ret;
1358 int16_t *plc_data = &s->plc_residual[LPC_FILTERORDER];
1359
1360 memset(&s->frame, 0, sizeof(ILBCFrame));
1361 ret = unpack_frame(s, avpkt->data, avpkt->size);
1362 if (ret < 0)
1363 return ret;
1364 if (ret)
1365 mode = 0;
1366
1367 frame->nb_samples = s->block_samples;
1368 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
1369 return ret;
1370
1371 if (s->frame.start < 1 || s->frame.start > 5)
1372 mode = 0;
1373
1374 if (mode) {
1375 index_conv(s->frame.cb_index);
1376
1377 lsf_dequantization(s->lsfdeq, s->frame.lsf, s->lpc_n);
1378 lsf_check_stability(s->lsfdeq, LPC_FILTERORDER, s->lpc_n);
1379 lsp_interpolate(s->syntdenum, s->weightdenum,
1380 s->lsfdeq, LPC_FILTERORDER, s);
1381 decode_residual(s, &s->frame, s->decresidual, s->syntdenum);
1382
1383 do_plc(s->plc_residual, s->plc_lpc, 0,
1384 s->decresidual, s->syntdenum + (LPC_FILTERORDER + 1) * (s->nsub - 1),
1385 s->last_lag, s);
1386
1387 memcpy(s->decresidual, s->plc_residual, s->block_samples * 2);
1388 }
1389
1390 if (s->enhancer) {
1391 /* TODO */
1392 } else {
1393 int16_t lag, i;
1394
1395 /* Find last lag (since the enhancer is not called to give this info) */
1396 if (s->mode == 20) {
1397 lag = xcorr_coeff(&s->decresidual[s->block_samples-60], &s->decresidual[s->block_samples-80],
1398 60, 80, 20, -1);
1399 } else {
1400 lag = xcorr_coeff(&s->decresidual[s->block_samples-ENH_BLOCKL],
1401 &s->decresidual[s->block_samples-ENH_BLOCKL-20],
1402 ENH_BLOCKL, 100, 20, -1);
1403 }
1404
1405 /* Store lag (it is needed if next packet is lost) */
1406 s->last_lag = lag;
1407
1408 /* copy data and run synthesis filter */
1409 memcpy(plc_data, s->decresidual, s->block_samples * 2);
1410
1411 /* Set up the filter state */
1412 memcpy(&plc_data[-LPC_FILTERORDER], s->syntMem, LPC_FILTERORDER * 2);
1413
1414 for (i = 0; i < s->nsub; i++) {
1415 filter_arfq12(plc_data+i*SUBL, plc_data+i*SUBL,
1416 s->syntdenum + i*(LPC_FILTERORDER + 1),
1417 LPC_FILTERORDER + 1, SUBL);
1418 }
1419
1420 /* Save the filter state */
1421 memcpy(s->syntMem, &plc_data[s->block_samples-LPC_FILTERORDER], LPC_FILTERORDER * 2);
1422 }
1423
1424 memcpy(frame->data[0], plc_data, s->block_samples * 2);
1425
1426 hp_output((int16_t *)frame->data[0], hp_out_coeffs,
1427 s->hpimemy, s->hpimemx, s->block_samples);
1428
1429 memcpy(s->old_syntdenum, s->syntdenum, s->nsub*(LPC_FILTERORDER + 1) * 2);
1430
1431 s->prev_enh_pl = 0;
1432 if (mode == 0)
1433 s->prev_enh_pl = 1;
1434
1435 *got_frame_ptr = 1;
1436
1437 return avpkt->size;
1438 }
1439
1440 static av_cold int ilbc_decode_init(AVCodecContext *avctx)
1441 {
1442 ILBCContext *s = avctx->priv_data;
1443
1444 if (avctx->block_align == 38)
1445 s->mode = 20;
1446 else if (avctx->block_align == 50)
1447 s->mode = 30;
1448 else if (avctx->bit_rate > 0)
1449 s->mode = avctx->bit_rate <= 14000 ? 30 : 20;
1450 else
1451 return AVERROR_INVALIDDATA;
1452
1453 av_channel_layout_uninit(&avctx->ch_layout);
1454 avctx->ch_layout = (AVChannelLayout)AV_CHANNEL_LAYOUT_MONO;
1455 avctx->sample_rate = 8000;
1456 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
1457
1458 if (s->mode == 30) {
1459 s->block_samples = 240;
1460 s->nsub = NSUB_30MS;
1461 s->nasub = NASUB_30MS;
1462 s->lpc_n = LPC_N_30MS;
1463 s->state_short_len = STATE_SHORT_LEN_30MS;
1464 } else {
1465 s->block_samples = 160;
1466 s->nsub = NSUB_20MS;
1467 s->nasub = NASUB_20MS;
1468 s->lpc_n = LPC_N_20MS;
1469 s->state_short_len = STATE_SHORT_LEN_20MS;
1470 }
1471
1472 return 0;
1473 }
1474
1475 const FFCodec ff_ilbc_decoder = {
1476 .p.name = "ilbc",
1477 CODEC_LONG_NAME("iLBC (Internet Low Bitrate Codec)"),
1478 .p.type = AVMEDIA_TYPE_AUDIO,
1479 .p.id = AV_CODEC_ID_ILBC,
1480 .init = ilbc_decode_init,
1481 FF_CODEC_DECODE_CB(ilbc_decode_frame),
1482 .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF,
1483 .priv_data_size = sizeof(ILBCContext),
1484 };
1485