doxygen: misc consistency, spelling and wording fixes
[libav.git] / libavcodec / amrwbdec.c
1 /*
2 * AMR wideband decoder
3 * Copyright (c) 2010 Marcelo Galvao Povoa
4 *
5 * This file is part of Libav.
6 *
7 * Libav is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
11 *
12 * Libav is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A particular PURPOSE. See the GNU
15 * Lesser General Public License for more details.
16 *
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with Libav; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 */
21
22 /**
23 * @file
24 * AMR wideband decoder
25 */
26
27 #include "libavutil/lfg.h"
28
29 #include "avcodec.h"
30 #include "get_bits.h"
31 #include "lsp.h"
32 #include "celp_math.h"
33 #include "celp_filters.h"
34 #include "acelp_filters.h"
35 #include "acelp_vectors.h"
36 #include "acelp_pitch_delay.h"
37
38 #define AMR_USE_16BIT_TABLES
39 #include "amr.h"
40
41 #include "amrwbdata.h"
42
43 typedef struct {
44 AVFrame avframe; ///< AVFrame for decoded samples
45 AMRWBFrame frame; ///< AMRWB parameters decoded from bitstream
46 enum Mode fr_cur_mode; ///< mode index of current frame
47 uint8_t fr_quality; ///< frame quality index (FQI)
48 float isf_cur[LP_ORDER]; ///< working ISF vector from current frame
49 float isf_q_past[LP_ORDER]; ///< quantized ISF vector of the previous frame
50 float isf_past_final[LP_ORDER]; ///< final processed ISF vector of the previous frame
51 double isp[4][LP_ORDER]; ///< ISP vectors from current frame
52 double isp_sub4_past[LP_ORDER]; ///< ISP vector for the 4th subframe of the previous frame
53
54 float lp_coef[4][LP_ORDER]; ///< Linear Prediction Coefficients from ISP vector
55
56 uint8_t base_pitch_lag; ///< integer part of pitch lag for the next relative subframe
57 uint8_t pitch_lag_int; ///< integer part of pitch lag of the previous subframe
58
59 float excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 2 + AMRWB_SFR_SIZE]; ///< current excitation and all necessary excitation history
60 float *excitation; ///< points to current excitation in excitation_buf[]
61
62 float pitch_vector[AMRWB_SFR_SIZE]; ///< adaptive codebook (pitch) vector for current subframe
63 float fixed_vector[AMRWB_SFR_SIZE]; ///< algebraic codebook (fixed) vector for current subframe
64
65 float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes
66 float pitch_gain[6]; ///< quantified pitch gains for the current and previous five subframes
67 float fixed_gain[2]; ///< quantified fixed gains for the current and previous subframes
68
69 float tilt_coef; ///< {beta_1} related to the voicing of the previous subframe
70
71 float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness to determine "onset"
72 uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none
73 float prev_tr_gain; ///< previous initial gain used by noise enhancer for threshold
74
75 float samples_az[LP_ORDER + AMRWB_SFR_SIZE]; ///< low-band samples and memory from synthesis at 12.8kHz
76 float samples_up[UPS_MEM_SIZE + AMRWB_SFR_SIZE]; ///< low-band samples and memory processed for upsampling
77 float samples_hb[LP_ORDER_16k + AMRWB_SFR_SIZE_16k]; ///< high-band samples and memory from synthesis at 16kHz
78
79 float hpf_31_mem[2], hpf_400_mem[2]; ///< previous values in the high pass filters
80 float demph_mem[1]; ///< previous value in the de-emphasis filter
81 float bpf_6_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band band pass filter
82 float lpf_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band low pass filter
83
84 AVLFG prng; ///< random number generator for white noise excitation
85 uint8_t first_frame; ///< flag active during decoding of the first frame
86 } AMRWBContext;
87
88 static av_cold int amrwb_decode_init(AVCodecContext *avctx)
89 {
90 AMRWBContext *ctx = avctx->priv_data;
91 int i;
92
93 avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
94
95 av_lfg_init(&ctx->prng, 1);
96
97 ctx->excitation = &ctx->excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 1];
98 ctx->first_frame = 1;
99
100 for (i = 0; i < LP_ORDER; i++)
101 ctx->isf_past_final[i] = isf_init[i] * (1.0f / (1 << 15));
102
103 for (i = 0; i < 4; i++)
104 ctx->prediction_error[i] = MIN_ENERGY;
105
106 avcodec_get_frame_defaults(&ctx->avframe);
107 avctx->coded_frame = &ctx->avframe;
108
109 return 0;
110 }
111
112 /**
113 * Decode the frame header in the "MIME/storage" format. This format
114 * is simpler and does not carry the auxiliary frame information.
115 *
116 * @param[in] ctx The Context
117 * @param[in] buf Pointer to the input buffer
118 *
119 * @return The decoded header length in bytes
120 */
121 static int decode_mime_header(AMRWBContext *ctx, const uint8_t *buf)
122 {
123 GetBitContext gb;
124 init_get_bits(&gb, buf, 8);
125
126 /* Decode frame header (1st octet) */
127 skip_bits(&gb, 1); // padding bit
128 ctx->fr_cur_mode = get_bits(&gb, 4);
129 ctx->fr_quality = get_bits1(&gb);
130 skip_bits(&gb, 2); // padding bits
131
132 return 1;
133 }
134
135 /**
136 * Decode quantized ISF vectors using 36-bit indexes (6K60 mode only).
137 *
138 * @param[in] ind Array of 5 indexes
139 * @param[out] isf_q Buffer for isf_q[LP_ORDER]
140 *
141 */
142 static void decode_isf_indices_36b(uint16_t *ind, float *isf_q)
143 {
144 int i;
145
146 for (i = 0; i < 9; i++)
147 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
148
149 for (i = 0; i < 7; i++)
150 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
151
152 for (i = 0; i < 5; i++)
153 isf_q[i] += dico21_isf_36b[ind[2]][i] * (1.0f / (1 << 15));
154
155 for (i = 0; i < 4; i++)
156 isf_q[i + 5] += dico22_isf_36b[ind[3]][i] * (1.0f / (1 << 15));
157
158 for (i = 0; i < 7; i++)
159 isf_q[i + 9] += dico23_isf_36b[ind[4]][i] * (1.0f / (1 << 15));
160 }
161
162 /**
163 * Decode quantized ISF vectors using 46-bit indexes (except 6K60 mode).
164 *
165 * @param[in] ind Array of 7 indexes
166 * @param[out] isf_q Buffer for isf_q[LP_ORDER]
167 *
168 */
169 static void decode_isf_indices_46b(uint16_t *ind, float *isf_q)
170 {
171 int i;
172
173 for (i = 0; i < 9; i++)
174 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
175
176 for (i = 0; i < 7; i++)
177 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
178
179 for (i = 0; i < 3; i++)
180 isf_q[i] += dico21_isf[ind[2]][i] * (1.0f / (1 << 15));
181
182 for (i = 0; i < 3; i++)
183 isf_q[i + 3] += dico22_isf[ind[3]][i] * (1.0f / (1 << 15));
184
185 for (i = 0; i < 3; i++)
186 isf_q[i + 6] += dico23_isf[ind[4]][i] * (1.0f / (1 << 15));
187
188 for (i = 0; i < 3; i++)
189 isf_q[i + 9] += dico24_isf[ind[5]][i] * (1.0f / (1 << 15));
190
191 for (i = 0; i < 4; i++)
192 isf_q[i + 12] += dico25_isf[ind[6]][i] * (1.0f / (1 << 15));
193 }
194
195 /**
196 * Apply mean and past ISF values using the prediction factor.
197 * Updates past ISF vector.
198 *
199 * @param[in,out] isf_q Current quantized ISF
200 * @param[in,out] isf_past Past quantized ISF
201 *
202 */
203 static void isf_add_mean_and_past(float *isf_q, float *isf_past)
204 {
205 int i;
206 float tmp;
207
208 for (i = 0; i < LP_ORDER; i++) {
209 tmp = isf_q[i];
210 isf_q[i] += isf_mean[i] * (1.0f / (1 << 15));
211 isf_q[i] += PRED_FACTOR * isf_past[i];
212 isf_past[i] = tmp;
213 }
214 }
215
216 /**
217 * Interpolate the fourth ISP vector from current and past frames
218 * to obtain an ISP vector for each subframe.
219 *
220 * @param[in,out] isp_q ISPs for each subframe
221 * @param[in] isp4_past Past ISP for subframe 4
222 */
223 static void interpolate_isp(double isp_q[4][LP_ORDER], const double *isp4_past)
224 {
225 int i, k;
226
227 for (k = 0; k < 3; k++) {
228 float c = isfp_inter[k];
229 for (i = 0; i < LP_ORDER; i++)
230 isp_q[k][i] = (1.0 - c) * isp4_past[i] + c * isp_q[3][i];
231 }
232 }
233
234 /**
235 * Decode an adaptive codebook index into pitch lag (except 6k60, 8k85 modes).
236 * Calculate integer lag and fractional lag always using 1/4 resolution.
237 * In 1st and 3rd subframes the index is relative to last subframe integer lag.
238 *
239 * @param[out] lag_int Decoded integer pitch lag
240 * @param[out] lag_frac Decoded fractional pitch lag
241 * @param[in] pitch_index Adaptive codebook pitch index
242 * @param[in,out] base_lag_int Base integer lag used in relative subframes
243 * @param[in] subframe Current subframe index (0 to 3)
244 */
245 static void decode_pitch_lag_high(int *lag_int, int *lag_frac, int pitch_index,
246 uint8_t *base_lag_int, int subframe)
247 {
248 if (subframe == 0 || subframe == 2) {
249 if (pitch_index < 376) {
250 *lag_int = (pitch_index + 137) >> 2;
251 *lag_frac = pitch_index - (*lag_int << 2) + 136;
252 } else if (pitch_index < 440) {
253 *lag_int = (pitch_index + 257 - 376) >> 1;
254 *lag_frac = (pitch_index - (*lag_int << 1) + 256 - 376) << 1;
255 /* the actual resolution is 1/2 but expressed as 1/4 */
256 } else {
257 *lag_int = pitch_index - 280;
258 *lag_frac = 0;
259 }
260 /* minimum lag for next subframe */
261 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
262 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);
263 // XXX: the spec states clearly that *base_lag_int should be
264 // the nearest integer to *lag_int (minus 8), but the ref code
265 // actually always uses its floor, I'm following the latter
266 } else {
267 *lag_int = (pitch_index + 1) >> 2;
268 *lag_frac = pitch_index - (*lag_int << 2);
269 *lag_int += *base_lag_int;
270 }
271 }
272
273 /**
274 * Decode an adaptive codebook index into pitch lag for 8k85 and 6k60 modes.
275 * The description is analogous to decode_pitch_lag_high, but in 6k60 the
276 * relative index is used for all subframes except the first.
277 */
278 static void decode_pitch_lag_low(int *lag_int, int *lag_frac, int pitch_index,
279 uint8_t *base_lag_int, int subframe, enum Mode mode)
280 {
281 if (subframe == 0 || (subframe == 2 && mode != MODE_6k60)) {
282 if (pitch_index < 116) {
283 *lag_int = (pitch_index + 69) >> 1;
284 *lag_frac = (pitch_index - (*lag_int << 1) + 68) << 1;
285 } else {
286 *lag_int = pitch_index - 24;
287 *lag_frac = 0;
288 }
289 // XXX: same problem as before
290 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
291 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);
292 } else {
293 *lag_int = (pitch_index + 1) >> 1;
294 *lag_frac = (pitch_index - (*lag_int << 1)) << 1;
295 *lag_int += *base_lag_int;
296 }
297 }
298
299 /**
300 * Find the pitch vector by interpolating the past excitation at the
301 * pitch delay, which is obtained in this function.
302 *
303 * @param[in,out] ctx The context
304 * @param[in] amr_subframe Current subframe data
305 * @param[in] subframe Current subframe index (0 to 3)
306 */
307 static void decode_pitch_vector(AMRWBContext *ctx,
308 const AMRWBSubFrame *amr_subframe,
309 const int subframe)
310 {
311 int pitch_lag_int, pitch_lag_frac;
312 int i;
313 float *exc = ctx->excitation;
314 enum Mode mode = ctx->fr_cur_mode;
315
316 if (mode <= MODE_8k85) {
317 decode_pitch_lag_low(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
318 &ctx->base_pitch_lag, subframe, mode);
319 } else
320 decode_pitch_lag_high(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
321 &ctx->base_pitch_lag, subframe);
322
323 ctx->pitch_lag_int = pitch_lag_int;
324 pitch_lag_int += pitch_lag_frac > 0;
325
326 /* Calculate the pitch vector by interpolating the past excitation at the
327 pitch lag using a hamming windowed sinc function */
328 ff_acelp_interpolatef(exc, exc + 1 - pitch_lag_int,
329 ac_inter, 4,
330 pitch_lag_frac + (pitch_lag_frac > 0 ? 0 : 4),
331 LP_ORDER, AMRWB_SFR_SIZE + 1);
332
333 /* Check which pitch signal path should be used
334 * 6k60 and 8k85 modes have the ltp flag set to 0 */
335 if (amr_subframe->ltp) {
336 memcpy(ctx->pitch_vector, exc, AMRWB_SFR_SIZE * sizeof(float));
337 } else {
338 for (i = 0; i < AMRWB_SFR_SIZE; i++)
339 ctx->pitch_vector[i] = 0.18 * exc[i - 1] + 0.64 * exc[i] +
340 0.18 * exc[i + 1];
341 memcpy(exc, ctx->pitch_vector, AMRWB_SFR_SIZE * sizeof(float));
342 }
343 }
344
345 /** Get x bits in the index interval [lsb,lsb+len-1] inclusive */
346 #define BIT_STR(x,lsb,len) (((x) >> (lsb)) & ((1 << (len)) - 1))
347
348 /** Get the bit at specified position */
349 #define BIT_POS(x, p) (((x) >> (p)) & 1)
350
351 /**
352 * The next six functions decode_[i]p_track decode exactly i pulses
353 * positions and amplitudes (-1 or 1) in a subframe track using
354 * an encoded pulse indexing (TS 26.190 section 5.8.2).
355 *
356 * The results are given in out[], in which a negative number means
357 * amplitude -1 and vice versa (i.e., ampl(x) = x / abs(x) ).
358 *
359 * @param[out] out Output buffer (writes i elements)
360 * @param[in] code Pulse index (no. of bits varies, see below)
361 * @param[in] m (log2) Number of potential positions
362 * @param[in] off Offset for decoded positions
363 */
364 static inline void decode_1p_track(int *out, int code, int m, int off)
365 {
366 int pos = BIT_STR(code, 0, m) + off; ///code: m+1 bits
367
368 out[0] = BIT_POS(code, m) ? -pos : pos;
369 }
370
371 static inline void decode_2p_track(int *out, int code, int m, int off) ///code: 2m+1 bits
372 {
373 int pos0 = BIT_STR(code, m, m) + off;
374 int pos1 = BIT_STR(code, 0, m) + off;
375
376 out[0] = BIT_POS(code, 2*m) ? -pos0 : pos0;
377 out[1] = BIT_POS(code, 2*m) ? -pos1 : pos1;
378 out[1] = pos0 > pos1 ? -out[1] : out[1];
379 }
380
381 static void decode_3p_track(int *out, int code, int m, int off) ///code: 3m+1 bits
382 {
383 int half_2p = BIT_POS(code, 2*m - 1) << (m - 1);
384
385 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
386 m - 1, off + half_2p);
387 decode_1p_track(out + 2, BIT_STR(code, 2*m, m + 1), m, off);
388 }
389
390 static void decode_4p_track(int *out, int code, int m, int off) ///code: 4m bits
391 {
392 int half_4p, subhalf_2p;
393 int b_offset = 1 << (m - 1);
394
395 switch (BIT_STR(code, 4*m - 2, 2)) { /* case ID (2 bits) */
396 case 0: /* 0 pulses in A, 4 pulses in B or vice versa */
397 half_4p = BIT_POS(code, 4*m - 3) << (m - 1); // which has 4 pulses
398 subhalf_2p = BIT_POS(code, 2*m - 3) << (m - 2);
399
400 decode_2p_track(out, BIT_STR(code, 0, 2*m - 3),
401 m - 2, off + half_4p + subhalf_2p);
402 decode_2p_track(out + 2, BIT_STR(code, 2*m - 2, 2*m - 1),
403 m - 1, off + half_4p);
404 break;
405 case 1: /* 1 pulse in A, 3 pulses in B */
406 decode_1p_track(out, BIT_STR(code, 3*m - 2, m),
407 m - 1, off);
408 decode_3p_track(out + 1, BIT_STR(code, 0, 3*m - 2),
409 m - 1, off + b_offset);
410 break;
411 case 2: /* 2 pulses in each half */
412 decode_2p_track(out, BIT_STR(code, 2*m - 1, 2*m - 1),
413 m - 1, off);
414 decode_2p_track(out + 2, BIT_STR(code, 0, 2*m - 1),
415 m - 1, off + b_offset);
416 break;
417 case 3: /* 3 pulses in A, 1 pulse in B */
418 decode_3p_track(out, BIT_STR(code, m, 3*m - 2),
419 m - 1, off);
420 decode_1p_track(out + 3, BIT_STR(code, 0, m),
421 m - 1, off + b_offset);
422 break;
423 }
424 }
425
426 static void decode_5p_track(int *out, int code, int m, int off) ///code: 5m bits
427 {
428 int half_3p = BIT_POS(code, 5*m - 1) << (m - 1);
429
430 decode_3p_track(out, BIT_STR(code, 2*m + 1, 3*m - 2),
431 m - 1, off + half_3p);
432
433 decode_2p_track(out + 3, BIT_STR(code, 0, 2*m + 1), m, off);
434 }
435
436 static void decode_6p_track(int *out, int code, int m, int off) ///code: 6m-2 bits
437 {
438 int b_offset = 1 << (m - 1);
439 /* which half has more pulses in cases 0 to 2 */
440 int half_more = BIT_POS(code, 6*m - 5) << (m - 1);
441 int half_other = b_offset - half_more;
442
443 switch (BIT_STR(code, 6*m - 4, 2)) { /* case ID (2 bits) */
444 case 0: /* 0 pulses in A, 6 pulses in B or vice versa */
445 decode_1p_track(out, BIT_STR(code, 0, m),
446 m - 1, off + half_more);
447 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
448 m - 1, off + half_more);
449 break;
450 case 1: /* 1 pulse in A, 5 pulses in B or vice versa */
451 decode_1p_track(out, BIT_STR(code, 0, m),
452 m - 1, off + half_other);
453 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
454 m - 1, off + half_more);
455 break;
456 case 2: /* 2 pulses in A, 4 pulses in B or vice versa */
457 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
458 m - 1, off + half_other);
459 decode_4p_track(out + 2, BIT_STR(code, 2*m - 1, 4*m - 4),
460 m - 1, off + half_more);
461 break;
462 case 3: /* 3 pulses in A, 3 pulses in B */
463 decode_3p_track(out, BIT_STR(code, 3*m - 2, 3*m - 2),
464 m - 1, off);
465 decode_3p_track(out + 3, BIT_STR(code, 0, 3*m - 2),
466 m - 1, off + b_offset);
467 break;
468 }
469 }
470
471 /**
472 * Decode the algebraic codebook index to pulse positions and signs,
473 * then construct the algebraic codebook vector.
474 *
475 * @param[out] fixed_vector Buffer for the fixed codebook excitation
476 * @param[in] pulse_hi MSBs part of the pulse index array (higher modes only)
477 * @param[in] pulse_lo LSBs part of the pulse index array
478 * @param[in] mode Mode of the current frame
479 */
480 static void decode_fixed_vector(float *fixed_vector, const uint16_t *pulse_hi,
481 const uint16_t *pulse_lo, const enum Mode mode)
482 {
483 /* sig_pos stores for each track the decoded pulse position indexes
484 * (1-based) multiplied by its corresponding amplitude (+1 or -1) */
485 int sig_pos[4][6];
486 int spacing = (mode == MODE_6k60) ? 2 : 4;
487 int i, j;
488
489 switch (mode) {
490 case MODE_6k60:
491 for (i = 0; i < 2; i++)
492 decode_1p_track(sig_pos[i], pulse_lo[i], 5, 1);
493 break;
494 case MODE_8k85:
495 for (i = 0; i < 4; i++)
496 decode_1p_track(sig_pos[i], pulse_lo[i], 4, 1);
497 break;
498 case MODE_12k65:
499 for (i = 0; i < 4; i++)
500 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
501 break;
502 case MODE_14k25:
503 for (i = 0; i < 2; i++)
504 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
505 for (i = 2; i < 4; i++)
506 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
507 break;
508 case MODE_15k85:
509 for (i = 0; i < 4; i++)
510 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
511 break;
512 case MODE_18k25:
513 for (i = 0; i < 4; i++)
514 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
515 ((int) pulse_hi[i] << 14), 4, 1);
516 break;
517 case MODE_19k85:
518 for (i = 0; i < 2; i++)
519 decode_5p_track(sig_pos[i], (int) pulse_lo[i] +
520 ((int) pulse_hi[i] << 10), 4, 1);
521 for (i = 2; i < 4; i++)
522 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
523 ((int) pulse_hi[i] << 14), 4, 1);
524 break;
525 case MODE_23k05:
526 case MODE_23k85:
527 for (i = 0; i < 4; i++)
528 decode_6p_track(sig_pos[i], (int) pulse_lo[i] +
529 ((int) pulse_hi[i] << 11), 4, 1);
530 break;
531 }
532
533 memset(fixed_vector, 0, sizeof(float) * AMRWB_SFR_SIZE);
534
535 for (i = 0; i < 4; i++)
536 for (j = 0; j < pulses_nb_per_mode_tr[mode][i]; j++) {
537 int pos = (FFABS(sig_pos[i][j]) - 1) * spacing + i;
538
539 fixed_vector[pos] += sig_pos[i][j] < 0 ? -1.0 : 1.0;
540 }
541 }
542
543 /**
544 * Decode pitch gain and fixed gain correction factor.
545 *
546 * @param[in] vq_gain Vector-quantized index for gains
547 * @param[in] mode Mode of the current frame
548 * @param[out] fixed_gain_factor Decoded fixed gain correction factor
549 * @param[out] pitch_gain Decoded pitch gain
550 */
551 static void decode_gains(const uint8_t vq_gain, const enum Mode mode,
552 float *fixed_gain_factor, float *pitch_gain)
553 {
554 const int16_t *gains = (mode <= MODE_8k85 ? qua_gain_6b[vq_gain] :
555 qua_gain_7b[vq_gain]);
556
557 *pitch_gain = gains[0] * (1.0f / (1 << 14));
558 *fixed_gain_factor = gains[1] * (1.0f / (1 << 11));
559 }
560
561 /**
562 * Apply pitch sharpening filters to the fixed codebook vector.
563 *
564 * @param[in] ctx The context
565 * @param[in,out] fixed_vector Fixed codebook excitation
566 */
567 // XXX: Spec states this procedure should be applied when the pitch
568 // lag is less than 64, but this checking seems absent in reference and AMR-NB
569 static void pitch_sharpening(AMRWBContext *ctx, float *fixed_vector)
570 {
571 int i;
572
573 /* Tilt part */
574 for (i = AMRWB_SFR_SIZE - 1; i != 0; i--)
575 fixed_vector[i] -= fixed_vector[i - 1] * ctx->tilt_coef;
576
577 /* Periodicity enhancement part */
578 for (i = ctx->pitch_lag_int; i < AMRWB_SFR_SIZE; i++)
579 fixed_vector[i] += fixed_vector[i - ctx->pitch_lag_int] * 0.85;
580 }
581
582 /**
583 * Calculate the voicing factor (-1.0 = unvoiced to 1.0 = voiced).
584 *
585 * @param[in] p_vector, f_vector Pitch and fixed excitation vectors
586 * @param[in] p_gain, f_gain Pitch and fixed gains
587 */
588 // XXX: There is something wrong with the precision here! The magnitudes
589 // of the energies are not correct. Please check the reference code carefully
590 static float voice_factor(float *p_vector, float p_gain,
591 float *f_vector, float f_gain)
592 {
593 double p_ener = (double) ff_dot_productf(p_vector, p_vector,
594 AMRWB_SFR_SIZE) * p_gain * p_gain;
595 double f_ener = (double) ff_dot_productf(f_vector, f_vector,
596 AMRWB_SFR_SIZE) * f_gain * f_gain;
597
598 return (p_ener - f_ener) / (p_ener + f_ener);
599 }
600
601 /**
602 * Reduce fixed vector sparseness by smoothing with one of three IR filters,
603 * also known as "adaptive phase dispersion".
604 *
605 * @param[in] ctx The context
606 * @param[in,out] fixed_vector Unfiltered fixed vector
607 * @param[out] buf Space for modified vector if necessary
608 *
609 * @return The potentially overwritten filtered fixed vector address
610 */
611 static float *anti_sparseness(AMRWBContext *ctx,
612 float *fixed_vector, float *buf)
613 {
614 int ir_filter_nr;
615
616 if (ctx->fr_cur_mode > MODE_8k85) // no filtering in higher modes
617 return fixed_vector;
618
619 if (ctx->pitch_gain[0] < 0.6) {
620 ir_filter_nr = 0; // strong filtering
621 } else if (ctx->pitch_gain[0] < 0.9) {
622 ir_filter_nr = 1; // medium filtering
623 } else
624 ir_filter_nr = 2; // no filtering
625
626 /* detect 'onset' */
627 if (ctx->fixed_gain[0] > 3.0 * ctx->fixed_gain[1]) {
628 if (ir_filter_nr < 2)
629 ir_filter_nr++;
630 } else {
631 int i, count = 0;
632
633 for (i = 0; i < 6; i++)
634 if (ctx->pitch_gain[i] < 0.6)
635 count++;
636
637 if (count > 2)
638 ir_filter_nr = 0;
639
640 if (ir_filter_nr > ctx->prev_ir_filter_nr + 1)
641 ir_filter_nr--;
642 }
643
644 /* update ir filter strength history */
645 ctx->prev_ir_filter_nr = ir_filter_nr;
646
647 ir_filter_nr += (ctx->fr_cur_mode == MODE_8k85);
648
649 if (ir_filter_nr < 2) {
650 int i;
651 const float *coef = ir_filters_lookup[ir_filter_nr];
652
653 /* Circular convolution code in the reference
654 * decoder was modified to avoid using one
655 * extra array. The filtered vector is given by:
656 *
657 * c2(n) = sum(i,0,len-1){ c(i) * coef( (n - i + len) % len ) }
658 */
659
660 memset(buf, 0, sizeof(float) * AMRWB_SFR_SIZE);
661 for (i = 0; i < AMRWB_SFR_SIZE; i++)
662 if (fixed_vector[i])
663 ff_celp_circ_addf(buf, buf, coef, i, fixed_vector[i],
664 AMRWB_SFR_SIZE);
665 fixed_vector = buf;
666 }
667
668 return fixed_vector;
669 }
670
671 /**
672 * Calculate a stability factor {teta} based on distance between
673 * current and past isf. A value of 1 shows maximum signal stability.
674 */
675 static float stability_factor(const float *isf, const float *isf_past)
676 {
677 int i;
678 float acc = 0.0;
679
680 for (i = 0; i < LP_ORDER - 1; i++)
681 acc += (isf[i] - isf_past[i]) * (isf[i] - isf_past[i]);
682
683 // XXX: This part is not so clear from the reference code
684 // the result is more accurate changing the "/ 256" to "* 512"
685 return FFMAX(0.0, 1.25 - acc * 0.8 * 512);
686 }
687
688 /**
689 * Apply a non-linear fixed gain smoothing in order to reduce
690 * fluctuation in the energy of excitation.
691 *
692 * @param[in] fixed_gain Unsmoothed fixed gain
693 * @param[in,out] prev_tr_gain Previous threshold gain (updated)
694 * @param[in] voice_fac Frame voicing factor
695 * @param[in] stab_fac Frame stability factor
696 *
697 * @return The smoothed gain
698 */
699 static float noise_enhancer(float fixed_gain, float *prev_tr_gain,
700 float voice_fac, float stab_fac)
701 {
702 float sm_fac = 0.5 * (1 - voice_fac) * stab_fac;
703 float g0;
704
705 // XXX: the following fixed-point constants used to in(de)crement
706 // gain by 1.5dB were taken from the reference code, maybe it could
707 // be simpler
708 if (fixed_gain < *prev_tr_gain) {
709 g0 = FFMIN(*prev_tr_gain, fixed_gain + fixed_gain *
710 (6226 * (1.0f / (1 << 15)))); // +1.5 dB
711 } else
712 g0 = FFMAX(*prev_tr_gain, fixed_gain *
713 (27536 * (1.0f / (1 << 15)))); // -1.5 dB
714
715 *prev_tr_gain = g0; // update next frame threshold
716
717 return sm_fac * g0 + (1 - sm_fac) * fixed_gain;
718 }
719
720 /**
721 * Filter the fixed_vector to emphasize the higher frequencies.
722 *
723 * @param[in,out] fixed_vector Fixed codebook vector
724 * @param[in] voice_fac Frame voicing factor
725 */
726 static void pitch_enhancer(float *fixed_vector, float voice_fac)
727 {
728 int i;
729 float cpe = 0.125 * (1 + voice_fac);
730 float last = fixed_vector[0]; // holds c(i - 1)
731
732 fixed_vector[0] -= cpe * fixed_vector[1];
733
734 for (i = 1; i < AMRWB_SFR_SIZE - 1; i++) {
735 float cur = fixed_vector[i];
736
737 fixed_vector[i] -= cpe * (last + fixed_vector[i + 1]);
738 last = cur;
739 }
740
741 fixed_vector[AMRWB_SFR_SIZE - 1] -= cpe * last;
742 }
743
744 /**
745 * Conduct 16th order linear predictive coding synthesis from excitation.
746 *
747 * @param[in] ctx Pointer to the AMRWBContext
748 * @param[in] lpc Pointer to the LPC coefficients
749 * @param[out] excitation Buffer for synthesis final excitation
750 * @param[in] fixed_gain Fixed codebook gain for synthesis
751 * @param[in] fixed_vector Algebraic codebook vector
752 * @param[in,out] samples Pointer to the output samples and memory
753 */
754 static void synthesis(AMRWBContext *ctx, float *lpc, float *excitation,
755 float fixed_gain, const float *fixed_vector,
756 float *samples)
757 {
758 ff_weighted_vector_sumf(excitation, ctx->pitch_vector, fixed_vector,
759 ctx->pitch_gain[0], fixed_gain, AMRWB_SFR_SIZE);
760
761 /* emphasize pitch vector contribution in low bitrate modes */
762 if (ctx->pitch_gain[0] > 0.5 && ctx->fr_cur_mode <= MODE_8k85) {
763 int i;
764 float energy = ff_dot_productf(excitation, excitation,
765 AMRWB_SFR_SIZE);
766
767 // XXX: Weird part in both ref code and spec. A unknown parameter
768 // {beta} seems to be identical to the current pitch gain
769 float pitch_factor = 0.25 * ctx->pitch_gain[0] * ctx->pitch_gain[0];
770
771 for (i = 0; i < AMRWB_SFR_SIZE; i++)
772 excitation[i] += pitch_factor * ctx->pitch_vector[i];
773
774 ff_scale_vector_to_given_sum_of_squares(excitation, excitation,
775 energy, AMRWB_SFR_SIZE);
776 }
777
778 ff_celp_lp_synthesis_filterf(samples, lpc, excitation,
779 AMRWB_SFR_SIZE, LP_ORDER);
780 }
781
782 /**
783 * Apply to synthesis a de-emphasis filter of the form:
784 * H(z) = 1 / (1 - m * z^-1)
785 *
786 * @param[out] out Output buffer
787 * @param[in] in Input samples array with in[-1]
788 * @param[in] m Filter coefficient
789 * @param[in,out] mem State from last filtering
790 */
791 static void de_emphasis(float *out, float *in, float m, float mem[1])
792 {
793 int i;
794
795 out[0] = in[0] + m * mem[0];
796
797 for (i = 1; i < AMRWB_SFR_SIZE; i++)
798 out[i] = in[i] + out[i - 1] * m;
799
800 mem[0] = out[AMRWB_SFR_SIZE - 1];
801 }
802
803 /**
804 * Upsample a signal by 5/4 ratio (from 12.8kHz to 16kHz) using
805 * a FIR interpolation filter. Uses past data from before *in address.
806 *
807 * @param[out] out Buffer for interpolated signal
808 * @param[in] in Current signal data (length 0.8*o_size)
809 * @param[in] o_size Output signal length
810 */
811 static void upsample_5_4(float *out, const float *in, int o_size)
812 {
813 const float *in0 = in - UPS_FIR_SIZE + 1;
814 int i, j, k;
815 int int_part = 0, frac_part;
816
817 i = 0;
818 for (j = 0; j < o_size / 5; j++) {
819 out[i] = in[int_part];
820 frac_part = 4;
821 i++;
822
823 for (k = 1; k < 5; k++) {
824 out[i] = ff_dot_productf(in0 + int_part, upsample_fir[4 - frac_part],
825 UPS_MEM_SIZE);
826 int_part++;
827 frac_part--;
828 i++;
829 }
830 }
831 }
832
833 /**
834 * Calculate the high-band gain based on encoded index (23k85 mode) or
835 * on the low-band speech signal and the Voice Activity Detection flag.
836 *
837 * @param[in] ctx The context
838 * @param[in] synth LB speech synthesis at 12.8k
839 * @param[in] hb_idx Gain index for mode 23k85 only
840 * @param[in] vad VAD flag for the frame
841 */
842 static float find_hb_gain(AMRWBContext *ctx, const float *synth,
843 uint16_t hb_idx, uint8_t vad)
844 {
845 int wsp = (vad > 0);
846 float tilt;
847
848 if (ctx->fr_cur_mode == MODE_23k85)
849 return qua_hb_gain[hb_idx] * (1.0f / (1 << 14));
850
851 tilt = ff_dot_productf(synth, synth + 1, AMRWB_SFR_SIZE - 1) /
852 ff_dot_productf(synth, synth, AMRWB_SFR_SIZE);
853
854 /* return gain bounded by [0.1, 1.0] */
855 return av_clipf((1.0 - FFMAX(0.0, tilt)) * (1.25 - 0.25 * wsp), 0.1, 1.0);
856 }
857
858 /**
859 * Generate the high-band excitation with the same energy from the lower
860 * one and scaled by the given gain.
861 *
862 * @param[in] ctx The context
863 * @param[out] hb_exc Buffer for the excitation
864 * @param[in] synth_exc Low-band excitation used for synthesis
865 * @param[in] hb_gain Wanted excitation gain
866 */
867 static void scaled_hb_excitation(AMRWBContext *ctx, float *hb_exc,
868 const float *synth_exc, float hb_gain)
869 {
870 int i;
871 float energy = ff_dot_productf(synth_exc, synth_exc, AMRWB_SFR_SIZE);
872
873 /* Generate a white-noise excitation */
874 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
875 hb_exc[i] = 32768.0 - (uint16_t) av_lfg_get(&ctx->prng);
876
877 ff_scale_vector_to_given_sum_of_squares(hb_exc, hb_exc,
878 energy * hb_gain * hb_gain,
879 AMRWB_SFR_SIZE_16k);
880 }
881
882 /**
883 * Calculate the auto-correlation for the ISF difference vector.
884 */
885 static float auto_correlation(float *diff_isf, float mean, int lag)
886 {
887 int i;
888 float sum = 0.0;
889
890 for (i = 7; i < LP_ORDER - 2; i++) {
891 float prod = (diff_isf[i] - mean) * (diff_isf[i - lag] - mean);
892 sum += prod * prod;
893 }
894 return sum;
895 }
896
897 /**
898 * Extrapolate a ISF vector to the 16kHz range (20th order LP)
899 * used at mode 6k60 LP filter for the high frequency band.
900 *
901 * @param[out] out Buffer for extrapolated isf
902 * @param[in] isf Input isf vector
903 */
904 static void extrapolate_isf(float out[LP_ORDER_16k], float isf[LP_ORDER])
905 {
906 float diff_isf[LP_ORDER - 2], diff_mean;
907 float *diff_hi = diff_isf - LP_ORDER + 1; // diff array for extrapolated indexes
908 float corr_lag[3];
909 float est, scale;
910 int i, i_max_corr;
911
912 memcpy(out, isf, (LP_ORDER - 1) * sizeof(float));
913 out[LP_ORDER_16k - 1] = isf[LP_ORDER - 1];
914
915 /* Calculate the difference vector */
916 for (i = 0; i < LP_ORDER - 2; i++)
917 diff_isf[i] = isf[i + 1] - isf[i];
918
919 diff_mean = 0.0;
920 for (i = 2; i < LP_ORDER - 2; i++)
921 diff_mean += diff_isf[i] * (1.0f / (LP_ORDER - 4));
922
923 /* Find which is the maximum autocorrelation */
924 i_max_corr = 0;
925 for (i = 0; i < 3; i++) {
926 corr_lag[i] = auto_correlation(diff_isf, diff_mean, i + 2);
927
928 if (corr_lag[i] > corr_lag[i_max_corr])
929 i_max_corr = i;
930 }
931 i_max_corr++;
932
933 for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
934 out[i] = isf[i - 1] + isf[i - 1 - i_max_corr]
935 - isf[i - 2 - i_max_corr];
936
937 /* Calculate an estimate for ISF(18) and scale ISF based on the error */
938 est = 7965 + (out[2] - out[3] - out[4]) / 6.0;
939 scale = 0.5 * (FFMIN(est, 7600) - out[LP_ORDER - 2]) /
940 (out[LP_ORDER_16k - 2] - out[LP_ORDER - 2]);
941
942 for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
943 diff_hi[i] = scale * (out[i] - out[i - 1]);
944
945 /* Stability insurance */
946 for (i = LP_ORDER; i < LP_ORDER_16k - 1; i++)
947 if (diff_hi[i] + diff_hi[i - 1] < 5.0) {
948 if (diff_hi[i] > diff_hi[i - 1]) {
949 diff_hi[i - 1] = 5.0 - diff_hi[i];
950 } else
951 diff_hi[i] = 5.0 - diff_hi[i - 1];
952 }
953
954 for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
955 out[i] = out[i - 1] + diff_hi[i] * (1.0f / (1 << 15));
956
957 /* Scale the ISF vector for 16000 Hz */
958 for (i = 0; i < LP_ORDER_16k - 1; i++)
959 out[i] *= 0.8;
960 }
961
962 /**
963 * Spectral expand the LP coefficients using the equation:
964 * y[i] = x[i] * (gamma ** i)
965 *
966 * @param[out] out Output buffer (may use input array)
967 * @param[in] lpc LP coefficients array
968 * @param[in] gamma Weighting factor
969 * @param[in] size LP array size
970 */
971 static void lpc_weighting(float *out, const float *lpc, float gamma, int size)
972 {
973 int i;
974 float fac = gamma;
975
976 for (i = 0; i < size; i++) {
977 out[i] = lpc[i] * fac;
978 fac *= gamma;
979 }
980 }
981
982 /**
983 * Conduct 20th order linear predictive coding synthesis for the high
984 * frequency band excitation at 16kHz.
985 *
986 * @param[in] ctx The context
987 * @param[in] subframe Current subframe index (0 to 3)
988 * @param[in,out] samples Pointer to the output speech samples
989 * @param[in] exc Generated white-noise scaled excitation
990 * @param[in] isf Current frame isf vector
991 * @param[in] isf_past Past frame final isf vector
992 */
993 static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples,
994 const float *exc, const float *isf, const float *isf_past)
995 {
996 float hb_lpc[LP_ORDER_16k];
997 enum Mode mode = ctx->fr_cur_mode;
998
999 if (mode == MODE_6k60) {
1000 float e_isf[LP_ORDER_16k]; // ISF vector for extrapolation
1001 double e_isp[LP_ORDER_16k];
1002
1003 ff_weighted_vector_sumf(e_isf, isf_past, isf, isfp_inter[subframe],
1004 1.0 - isfp_inter[subframe], LP_ORDER);
1005
1006 extrapolate_isf(e_isf, e_isf);
1007
1008 e_isf[LP_ORDER_16k - 1] *= 2.0;
1009 ff_acelp_lsf2lspd(e_isp, e_isf, LP_ORDER_16k);
1010 ff_amrwb_lsp2lpc(e_isp, hb_lpc, LP_ORDER_16k);
1011
1012 lpc_weighting(hb_lpc, hb_lpc, 0.9, LP_ORDER_16k);
1013 } else {
1014 lpc_weighting(hb_lpc, ctx->lp_coef[subframe], 0.6, LP_ORDER);
1015 }
1016
1017 ff_celp_lp_synthesis_filterf(samples, hb_lpc, exc, AMRWB_SFR_SIZE_16k,
1018 (mode == MODE_6k60) ? LP_ORDER_16k : LP_ORDER);
1019 }
1020
1021 /**
1022 * Apply a 15th order filter to high-band samples.
1023 * The filter characteristic depends on the given coefficients.
1024 *
1025 * @param[out] out Buffer for filtered output
1026 * @param[in] fir_coef Filter coefficients
1027 * @param[in,out] mem State from last filtering (updated)
1028 * @param[in] in Input speech data (high-band)
1029 *
1030 * @remark It is safe to pass the same array in in and out parameters
1031 */
1032 static void hb_fir_filter(float *out, const float fir_coef[HB_FIR_SIZE + 1],
1033 float mem[HB_FIR_SIZE], const float *in)
1034 {
1035 int i, j;
1036 float data[AMRWB_SFR_SIZE_16k + HB_FIR_SIZE]; // past and current samples
1037
1038 memcpy(data, mem, HB_FIR_SIZE * sizeof(float));
1039 memcpy(data + HB_FIR_SIZE, in, AMRWB_SFR_SIZE_16k * sizeof(float));
1040
1041 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) {
1042 out[i] = 0.0;
1043 for (j = 0; j <= HB_FIR_SIZE; j++)
1044 out[i] += data[i + j] * fir_coef[j];
1045 }
1046
1047 memcpy(mem, data + AMRWB_SFR_SIZE_16k, HB_FIR_SIZE * sizeof(float));
1048 }
1049
1050 /**
1051 * Update context state before the next subframe.
1052 */
1053 static void update_sub_state(AMRWBContext *ctx)
1054 {
1055 memmove(&ctx->excitation_buf[0], &ctx->excitation_buf[AMRWB_SFR_SIZE],
1056 (AMRWB_P_DELAY_MAX + LP_ORDER + 1) * sizeof(float));
1057
1058 memmove(&ctx->pitch_gain[1], &ctx->pitch_gain[0], 5 * sizeof(float));
1059 memmove(&ctx->fixed_gain[1], &ctx->fixed_gain[0], 1 * sizeof(float));
1060
1061 memmove(&ctx->samples_az[0], &ctx->samples_az[AMRWB_SFR_SIZE],
1062 LP_ORDER * sizeof(float));
1063 memmove(&ctx->samples_up[0], &ctx->samples_up[AMRWB_SFR_SIZE],
1064 UPS_MEM_SIZE * sizeof(float));
1065 memmove(&ctx->samples_hb[0], &ctx->samples_hb[AMRWB_SFR_SIZE_16k],
1066 LP_ORDER_16k * sizeof(float));
1067 }
1068
1069 static int amrwb_decode_frame(AVCodecContext *avctx, void *data,
1070 int *got_frame_ptr, AVPacket *avpkt)
1071 {
1072 AMRWBContext *ctx = avctx->priv_data;
1073 AMRWBFrame *cf = &ctx->frame;
1074 const uint8_t *buf = avpkt->data;
1075 int buf_size = avpkt->size;
1076 int expected_fr_size, header_size;
1077 float *buf_out;
1078 float spare_vector[AMRWB_SFR_SIZE]; // extra stack space to hold result from anti-sparseness processing
1079 float fixed_gain_factor; // fixed gain correction factor (gamma)
1080 float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
1081 float synth_fixed_gain; // the fixed gain that synthesis should use
1082 float voice_fac, stab_fac; // parameters used for gain smoothing
1083 float synth_exc[AMRWB_SFR_SIZE]; // post-processed excitation for synthesis
1084 float hb_exc[AMRWB_SFR_SIZE_16k]; // excitation for the high frequency band
1085 float hb_samples[AMRWB_SFR_SIZE_16k]; // filtered high-band samples from synthesis
1086 float hb_gain;
1087 int sub, i, ret;
1088
1089 /* get output buffer */
1090 ctx->avframe.nb_samples = 4 * AMRWB_SFR_SIZE_16k;
1091 if ((ret = avctx->get_buffer(avctx, &ctx->avframe)) < 0) {
1092 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1093 return ret;
1094 }
1095 buf_out = (float *)ctx->avframe.data[0];
1096
1097 header_size = decode_mime_header(ctx, buf);
1098 expected_fr_size = ((cf_sizes_wb[ctx->fr_cur_mode] + 7) >> 3) + 1;
1099
1100 if (buf_size < expected_fr_size) {
1101 av_log(avctx, AV_LOG_ERROR,
1102 "Frame too small (%d bytes). Truncated file?\n", buf_size);
1103 *got_frame_ptr = 0;
1104 return buf_size;
1105 }
1106
1107 if (!ctx->fr_quality || ctx->fr_cur_mode > MODE_SID)
1108 av_log(avctx, AV_LOG_ERROR, "Encountered a bad or corrupted frame\n");
1109
1110 if (ctx->fr_cur_mode == MODE_SID) /* Comfort noise frame */
1111 av_log_missing_feature(avctx, "SID mode", 1);
1112
1113 if (ctx->fr_cur_mode >= MODE_SID)
1114 return -1;
1115
1116 ff_amr_bit_reorder((uint16_t *) &ctx->frame, sizeof(AMRWBFrame),
1117 buf + header_size, amr_bit_orderings_by_mode[ctx->fr_cur_mode]);
1118
1119 /* Decode the quantized ISF vector */
1120 if (ctx->fr_cur_mode == MODE_6k60) {
1121 decode_isf_indices_36b(cf->isp_id, ctx->isf_cur);
1122 } else {
1123 decode_isf_indices_46b(cf->isp_id, ctx->isf_cur);
1124 }
1125
1126 isf_add_mean_and_past(ctx->isf_cur, ctx->isf_q_past);
1127 ff_set_min_dist_lsf(ctx->isf_cur, MIN_ISF_SPACING, LP_ORDER - 1);
1128
1129 stab_fac = stability_factor(ctx->isf_cur, ctx->isf_past_final);
1130
1131 ctx->isf_cur[LP_ORDER - 1] *= 2.0;
1132 ff_acelp_lsf2lspd(ctx->isp[3], ctx->isf_cur, LP_ORDER);
1133
1134 /* Generate a ISP vector for each subframe */
1135 if (ctx->first_frame) {
1136 ctx->first_frame = 0;
1137 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(double));
1138 }
1139 interpolate_isp(ctx->isp, ctx->isp_sub4_past);
1140
1141 for (sub = 0; sub < 4; sub++)
1142 ff_amrwb_lsp2lpc(ctx->isp[sub], ctx->lp_coef[sub], LP_ORDER);
1143
1144 for (sub = 0; sub < 4; sub++) {
1145 const AMRWBSubFrame *cur_subframe = &cf->subframe[sub];
1146 float *sub_buf = buf_out + sub * AMRWB_SFR_SIZE_16k;
1147
1148 /* Decode adaptive codebook (pitch vector) */
1149 decode_pitch_vector(ctx, cur_subframe, sub);
1150 /* Decode innovative codebook (fixed vector) */
1151 decode_fixed_vector(ctx->fixed_vector, cur_subframe->pul_ih,
1152 cur_subframe->pul_il, ctx->fr_cur_mode);
1153
1154 pitch_sharpening(ctx, ctx->fixed_vector);
1155
1156 decode_gains(cur_subframe->vq_gain, ctx->fr_cur_mode,
1157 &fixed_gain_factor, &ctx->pitch_gain[0]);
1158
1159 ctx->fixed_gain[0] =
1160 ff_amr_set_fixed_gain(fixed_gain_factor,
1161 ff_dot_productf(ctx->fixed_vector, ctx->fixed_vector,
1162 AMRWB_SFR_SIZE) / AMRWB_SFR_SIZE,
1163 ctx->prediction_error,
1164 ENERGY_MEAN, energy_pred_fac);
1165
1166 /* Calculate voice factor and store tilt for next subframe */
1167 voice_fac = voice_factor(ctx->pitch_vector, ctx->pitch_gain[0],
1168 ctx->fixed_vector, ctx->fixed_gain[0]);
1169 ctx->tilt_coef = voice_fac * 0.25 + 0.25;
1170
1171 /* Construct current excitation */
1172 for (i = 0; i < AMRWB_SFR_SIZE; i++) {
1173 ctx->excitation[i] *= ctx->pitch_gain[0];
1174 ctx->excitation[i] += ctx->fixed_gain[0] * ctx->fixed_vector[i];
1175 ctx->excitation[i] = truncf(ctx->excitation[i]);
1176 }
1177
1178 /* Post-processing of excitation elements */
1179 synth_fixed_gain = noise_enhancer(ctx->fixed_gain[0], &ctx->prev_tr_gain,
1180 voice_fac, stab_fac);
1181
1182 synth_fixed_vector = anti_sparseness(ctx, ctx->fixed_vector,
1183 spare_vector);
1184
1185 pitch_enhancer(synth_fixed_vector, voice_fac);
1186
1187 synthesis(ctx, ctx->lp_coef[sub], synth_exc, synth_fixed_gain,
1188 synth_fixed_vector, &ctx->samples_az[LP_ORDER]);
1189
1190 /* Synthesis speech post-processing */
1191 de_emphasis(&ctx->samples_up[UPS_MEM_SIZE],
1192 &ctx->samples_az[LP_ORDER], PREEMPH_FAC, ctx->demph_mem);
1193
1194 ff_acelp_apply_order_2_transfer_function(&ctx->samples_up[UPS_MEM_SIZE],
1195 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_31_poles,
1196 hpf_31_gain, ctx->hpf_31_mem, AMRWB_SFR_SIZE);
1197
1198 upsample_5_4(sub_buf, &ctx->samples_up[UPS_FIR_SIZE],
1199 AMRWB_SFR_SIZE_16k);
1200
1201 /* High frequency band (6.4 - 7.0 kHz) generation part */
1202 ff_acelp_apply_order_2_transfer_function(hb_samples,
1203 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_400_poles,
1204 hpf_400_gain, ctx->hpf_400_mem, AMRWB_SFR_SIZE);
1205
1206 hb_gain = find_hb_gain(ctx, hb_samples,
1207 cur_subframe->hb_gain, cf->vad);
1208
1209 scaled_hb_excitation(ctx, hb_exc, synth_exc, hb_gain);
1210
1211 hb_synthesis(ctx, sub, &ctx->samples_hb[LP_ORDER_16k],
1212 hb_exc, ctx->isf_cur, ctx->isf_past_final);
1213
1214 /* High-band post-processing filters */
1215 hb_fir_filter(hb_samples, bpf_6_7_coef, ctx->bpf_6_7_mem,
1216 &ctx->samples_hb[LP_ORDER_16k]);
1217
1218 if (ctx->fr_cur_mode == MODE_23k85)
1219 hb_fir_filter(hb_samples, lpf_7_coef, ctx->lpf_7_mem,
1220 hb_samples);
1221
1222 /* Add the low and high frequency bands */
1223 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
1224 sub_buf[i] = (sub_buf[i] + hb_samples[i]) * (1.0f / (1 << 15));
1225
1226 /* Update buffers and history */
1227 update_sub_state(ctx);
1228 }
1229
1230 /* update state for next frame */
1231 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(ctx->isp[3][0]));
1232 memcpy(ctx->isf_past_final, ctx->isf_cur, LP_ORDER * sizeof(float));
1233
1234 *got_frame_ptr = 1;
1235 *(AVFrame *)data = ctx->avframe;
1236
1237 return expected_fr_size;
1238 }
1239
1240 AVCodec ff_amrwb_decoder = {
1241 .name = "amrwb",
1242 .type = AVMEDIA_TYPE_AUDIO,
1243 .id = CODEC_ID_AMR_WB,
1244 .priv_data_size = sizeof(AMRWBContext),
1245 .init = amrwb_decode_init,
1246 .decode = amrwb_decode_frame,
1247 .capabilities = CODEC_CAP_DR1,
1248 .long_name = NULL_IF_CONFIG_SMALL("Adaptive Multi-Rate WideBand"),
1249 .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_FLT,AV_SAMPLE_FMT_NONE},
1250 };