cf137e046aa82cc27e3ad312b3faeee1d5f98cb3
[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/channel_layout.h"
28 #include "libavutil/common.h"
29 #include "libavutil/float_dsp.h"
30 #include "libavutil/lfg.h"
31
32 #include "avcodec.h"
33 #include "lsp.h"
34 #include "celp_filters.h"
35 #include "acelp_filters.h"
36 #include "acelp_vectors.h"
37 #include "acelp_pitch_delay.h"
38 #include "internal.h"
39
40 #define AMR_USE_16BIT_TABLES
41 #include "amr.h"
42
43 #include "amrwbdata.h"
44
45 typedef struct {
46 AMRWBFrame frame; ///< AMRWB parameters decoded from bitstream
47 enum Mode fr_cur_mode; ///< mode index of current frame
48 uint8_t fr_quality; ///< frame quality index (FQI)
49 float isf_cur[LP_ORDER]; ///< working ISF vector from current frame
50 float isf_q_past[LP_ORDER]; ///< quantized ISF vector of the previous frame
51 float isf_past_final[LP_ORDER]; ///< final processed ISF vector of the previous frame
52 double isp[4][LP_ORDER]; ///< ISP vectors from current frame
53 double isp_sub4_past[LP_ORDER]; ///< ISP vector for the 4th subframe of the previous frame
54
55 float lp_coef[4][LP_ORDER]; ///< Linear Prediction Coefficients from ISP vector
56
57 uint8_t base_pitch_lag; ///< integer part of pitch lag for the next relative subframe
58 uint8_t pitch_lag_int; ///< integer part of pitch lag of the previous subframe
59
60 float excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 2 + AMRWB_SFR_SIZE]; ///< current excitation and all necessary excitation history
61 float *excitation; ///< points to current excitation in excitation_buf[]
62
63 float pitch_vector[AMRWB_SFR_SIZE]; ///< adaptive codebook (pitch) vector for current subframe
64 float fixed_vector[AMRWB_SFR_SIZE]; ///< algebraic codebook (fixed) vector for current subframe
65
66 float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes
67 float pitch_gain[6]; ///< quantified pitch gains for the current and previous five subframes
68 float fixed_gain[2]; ///< quantified fixed gains for the current and previous subframes
69
70 float tilt_coef; ///< {beta_1} related to the voicing of the previous subframe
71
72 float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness to determine "onset"
73 uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none
74 float prev_tr_gain; ///< previous initial gain used by noise enhancer for threshold
75
76 float samples_az[LP_ORDER + AMRWB_SFR_SIZE]; ///< low-band samples and memory from synthesis at 12.8kHz
77 float samples_up[UPS_MEM_SIZE + AMRWB_SFR_SIZE]; ///< low-band samples and memory processed for upsampling
78 float samples_hb[LP_ORDER_16k + AMRWB_SFR_SIZE_16k]; ///< high-band samples and memory from synthesis at 16kHz
79
80 float hpf_31_mem[2], hpf_400_mem[2]; ///< previous values in the high pass filters
81 float demph_mem[1]; ///< previous value in the de-emphasis filter
82 float bpf_6_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band band pass filter
83 float lpf_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band low pass filter
84
85 AVLFG prng; ///< random number generator for white noise excitation
86 uint8_t first_frame; ///< flag active during decoding of the first frame
87 } AMRWBContext;
88
89 static av_cold int amrwb_decode_init(AVCodecContext *avctx)
90 {
91 AMRWBContext *ctx = avctx->priv_data;
92 int i;
93
94 if (avctx->channels > 1) {
95 av_log_missing_feature(avctx, "multi-channel AMR", 0);
96 return AVERROR_PATCHWELCOME;
97 }
98
99 avctx->channels = 1;
100 avctx->channel_layout = AV_CH_LAYOUT_MONO;
101 avctx->sample_rate = 16000;
102 avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
103
104 av_lfg_init(&ctx->prng, 1);
105
106 ctx->excitation = &ctx->excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 1];
107 ctx->first_frame = 1;
108
109 for (i = 0; i < LP_ORDER; i++)
110 ctx->isf_past_final[i] = isf_init[i] * (1.0f / (1 << 15));
111
112 for (i = 0; i < 4; i++)
113 ctx->prediction_error[i] = MIN_ENERGY;
114
115 return 0;
116 }
117
118 /**
119 * Decode the frame header in the "MIME/storage" format. This format
120 * is simpler and does not carry the auxiliary frame information.
121 *
122 * @param[in] ctx The Context
123 * @param[in] buf Pointer to the input buffer
124 *
125 * @return The decoded header length in bytes
126 */
127 static int decode_mime_header(AMRWBContext *ctx, const uint8_t *buf)
128 {
129 /* Decode frame header (1st octet) */
130 ctx->fr_cur_mode = buf[0] >> 3 & 0x0F;
131 ctx->fr_quality = (buf[0] & 0x4) == 0x4;
132
133 return 1;
134 }
135
136 /**
137 * Decode quantized ISF vectors using 36-bit indexes (6K60 mode only).
138 *
139 * @param[in] ind Array of 5 indexes
140 * @param[out] isf_q Buffer for isf_q[LP_ORDER]
141 *
142 */
143 static void decode_isf_indices_36b(uint16_t *ind, float *isf_q)
144 {
145 int i;
146
147 for (i = 0; i < 9; i++)
148 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
149
150 for (i = 0; i < 7; i++)
151 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
152
153 for (i = 0; i < 5; i++)
154 isf_q[i] += dico21_isf_36b[ind[2]][i] * (1.0f / (1 << 15));
155
156 for (i = 0; i < 4; i++)
157 isf_q[i + 5] += dico22_isf_36b[ind[3]][i] * (1.0f / (1 << 15));
158
159 for (i = 0; i < 7; i++)
160 isf_q[i + 9] += dico23_isf_36b[ind[4]][i] * (1.0f / (1 << 15));
161 }
162
163 /**
164 * Decode quantized ISF vectors using 46-bit indexes (except 6K60 mode).
165 *
166 * @param[in] ind Array of 7 indexes
167 * @param[out] isf_q Buffer for isf_q[LP_ORDER]
168 *
169 */
170 static void decode_isf_indices_46b(uint16_t *ind, float *isf_q)
171 {
172 int i;
173
174 for (i = 0; i < 9; i++)
175 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
176
177 for (i = 0; i < 7; i++)
178 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
179
180 for (i = 0; i < 3; i++)
181 isf_q[i] += dico21_isf[ind[2]][i] * (1.0f / (1 << 15));
182
183 for (i = 0; i < 3; i++)
184 isf_q[i + 3] += dico22_isf[ind[3]][i] * (1.0f / (1 << 15));
185
186 for (i = 0; i < 3; i++)
187 isf_q[i + 6] += dico23_isf[ind[4]][i] * (1.0f / (1 << 15));
188
189 for (i = 0; i < 3; i++)
190 isf_q[i + 9] += dico24_isf[ind[5]][i] * (1.0f / (1 << 15));
191
192 for (i = 0; i < 4; i++)
193 isf_q[i + 12] += dico25_isf[ind[6]][i] * (1.0f / (1 << 15));
194 }
195
196 /**
197 * Apply mean and past ISF values using the prediction factor.
198 * Updates past ISF vector.
199 *
200 * @param[in,out] isf_q Current quantized ISF
201 * @param[in,out] isf_past Past quantized ISF
202 *
203 */
204 static void isf_add_mean_and_past(float *isf_q, float *isf_past)
205 {
206 int i;
207 float tmp;
208
209 for (i = 0; i < LP_ORDER; i++) {
210 tmp = isf_q[i];
211 isf_q[i] += isf_mean[i] * (1.0f / (1 << 15));
212 isf_q[i] += PRED_FACTOR * isf_past[i];
213 isf_past[i] = tmp;
214 }
215 }
216
217 /**
218 * Interpolate the fourth ISP vector from current and past frames
219 * to obtain an ISP vector for each subframe.
220 *
221 * @param[in,out] isp_q ISPs for each subframe
222 * @param[in] isp4_past Past ISP for subframe 4
223 */
224 static void interpolate_isp(double isp_q[4][LP_ORDER], const double *isp4_past)
225 {
226 int i, k;
227
228 for (k = 0; k < 3; k++) {
229 float c = isfp_inter[k];
230 for (i = 0; i < LP_ORDER; i++)
231 isp_q[k][i] = (1.0 - c) * isp4_past[i] + c * isp_q[3][i];
232 }
233 }
234
235 /**
236 * Decode an adaptive codebook index into pitch lag (except 6k60, 8k85 modes).
237 * Calculate integer lag and fractional lag always using 1/4 resolution.
238 * In 1st and 3rd subframes the index is relative to last subframe integer lag.
239 *
240 * @param[out] lag_int Decoded integer pitch lag
241 * @param[out] lag_frac Decoded fractional pitch lag
242 * @param[in] pitch_index Adaptive codebook pitch index
243 * @param[in,out] base_lag_int Base integer lag used in relative subframes
244 * @param[in] subframe Current subframe index (0 to 3)
245 */
246 static void decode_pitch_lag_high(int *lag_int, int *lag_frac, int pitch_index,
247 uint8_t *base_lag_int, int subframe)
248 {
249 if (subframe == 0 || subframe == 2) {
250 if (pitch_index < 376) {
251 *lag_int = (pitch_index + 137) >> 2;
252 *lag_frac = pitch_index - (*lag_int << 2) + 136;
253 } else if (pitch_index < 440) {
254 *lag_int = (pitch_index + 257 - 376) >> 1;
255 *lag_frac = (pitch_index - (*lag_int << 1) + 256 - 376) << 1;
256 /* the actual resolution is 1/2 but expressed as 1/4 */
257 } else {
258 *lag_int = pitch_index - 280;
259 *lag_frac = 0;
260 }
261 /* minimum lag for next subframe */
262 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
263 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);
264 // XXX: the spec states clearly that *base_lag_int should be
265 // the nearest integer to *lag_int (minus 8), but the ref code
266 // actually always uses its floor, I'm following the latter
267 } else {
268 *lag_int = (pitch_index + 1) >> 2;
269 *lag_frac = pitch_index - (*lag_int << 2);
270 *lag_int += *base_lag_int;
271 }
272 }
273
274 /**
275 * Decode an adaptive codebook index into pitch lag for 8k85 and 6k60 modes.
276 * The description is analogous to decode_pitch_lag_high, but in 6k60 the
277 * relative index is used for all subframes except the first.
278 */
279 static void decode_pitch_lag_low(int *lag_int, int *lag_frac, int pitch_index,
280 uint8_t *base_lag_int, int subframe, enum Mode mode)
281 {
282 if (subframe == 0 || (subframe == 2 && mode != MODE_6k60)) {
283 if (pitch_index < 116) {
284 *lag_int = (pitch_index + 69) >> 1;
285 *lag_frac = (pitch_index - (*lag_int << 1) + 68) << 1;
286 } else {
287 *lag_int = pitch_index - 24;
288 *lag_frac = 0;
289 }
290 // XXX: same problem as before
291 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
292 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);
293 } else {
294 *lag_int = (pitch_index + 1) >> 1;
295 *lag_frac = (pitch_index - (*lag_int << 1)) << 1;
296 *lag_int += *base_lag_int;
297 }
298 }
299
300 /**
301 * Find the pitch vector by interpolating the past excitation at the
302 * pitch delay, which is obtained in this function.
303 *
304 * @param[in,out] ctx The context
305 * @param[in] amr_subframe Current subframe data
306 * @param[in] subframe Current subframe index (0 to 3)
307 */
308 static void decode_pitch_vector(AMRWBContext *ctx,
309 const AMRWBSubFrame *amr_subframe,
310 const int subframe)
311 {
312 int pitch_lag_int, pitch_lag_frac;
313 int i;
314 float *exc = ctx->excitation;
315 enum Mode mode = ctx->fr_cur_mode;
316
317 if (mode <= MODE_8k85) {
318 decode_pitch_lag_low(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
319 &ctx->base_pitch_lag, subframe, mode);
320 } else
321 decode_pitch_lag_high(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
322 &ctx->base_pitch_lag, subframe);
323
324 ctx->pitch_lag_int = pitch_lag_int;
325 pitch_lag_int += pitch_lag_frac > 0;
326
327 /* Calculate the pitch vector by interpolating the past excitation at the
328 pitch lag using a hamming windowed sinc function */
329 ff_acelp_interpolatef(exc, exc + 1 - pitch_lag_int,
330 ac_inter, 4,
331 pitch_lag_frac + (pitch_lag_frac > 0 ? 0 : 4),
332 LP_ORDER, AMRWB_SFR_SIZE + 1);
333
334 /* Check which pitch signal path should be used
335 * 6k60 and 8k85 modes have the ltp flag set to 0 */
336 if (amr_subframe->ltp) {
337 memcpy(ctx->pitch_vector, exc, AMRWB_SFR_SIZE * sizeof(float));
338 } else {
339 for (i = 0; i < AMRWB_SFR_SIZE; i++)
340 ctx->pitch_vector[i] = 0.18 * exc[i - 1] + 0.64 * exc[i] +
341 0.18 * exc[i + 1];
342 memcpy(exc, ctx->pitch_vector, AMRWB_SFR_SIZE * sizeof(float));
343 }
344 }
345
346 /** Get x bits in the index interval [lsb,lsb+len-1] inclusive */
347 #define BIT_STR(x,lsb,len) (((x) >> (lsb)) & ((1 << (len)) - 1))
348
349 /** Get the bit at specified position */
350 #define BIT_POS(x, p) (((x) >> (p)) & 1)
351
352 /**
353 * The next six functions decode_[i]p_track decode exactly i pulses
354 * positions and amplitudes (-1 or 1) in a subframe track using
355 * an encoded pulse indexing (TS 26.190 section 5.8.2).
356 *
357 * The results are given in out[], in which a negative number means
358 * amplitude -1 and vice versa (i.e., ampl(x) = x / abs(x) ).
359 *
360 * @param[out] out Output buffer (writes i elements)
361 * @param[in] code Pulse index (no. of bits varies, see below)
362 * @param[in] m (log2) Number of potential positions
363 * @param[in] off Offset for decoded positions
364 */
365 static inline void decode_1p_track(int *out, int code, int m, int off)
366 {
367 int pos = BIT_STR(code, 0, m) + off; ///code: m+1 bits
368
369 out[0] = BIT_POS(code, m) ? -pos : pos;
370 }
371
372 static inline void decode_2p_track(int *out, int code, int m, int off) ///code: 2m+1 bits
373 {
374 int pos0 = BIT_STR(code, m, m) + off;
375 int pos1 = BIT_STR(code, 0, m) + off;
376
377 out[0] = BIT_POS(code, 2*m) ? -pos0 : pos0;
378 out[1] = BIT_POS(code, 2*m) ? -pos1 : pos1;
379 out[1] = pos0 > pos1 ? -out[1] : out[1];
380 }
381
382 static void decode_3p_track(int *out, int code, int m, int off) ///code: 3m+1 bits
383 {
384 int half_2p = BIT_POS(code, 2*m - 1) << (m - 1);
385
386 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
387 m - 1, off + half_2p);
388 decode_1p_track(out + 2, BIT_STR(code, 2*m, m + 1), m, off);
389 }
390
391 static void decode_4p_track(int *out, int code, int m, int off) ///code: 4m bits
392 {
393 int half_4p, subhalf_2p;
394 int b_offset = 1 << (m - 1);
395
396 switch (BIT_STR(code, 4*m - 2, 2)) { /* case ID (2 bits) */
397 case 0: /* 0 pulses in A, 4 pulses in B or vice versa */
398 half_4p = BIT_POS(code, 4*m - 3) << (m - 1); // which has 4 pulses
399 subhalf_2p = BIT_POS(code, 2*m - 3) << (m - 2);
400
401 decode_2p_track(out, BIT_STR(code, 0, 2*m - 3),
402 m - 2, off + half_4p + subhalf_2p);
403 decode_2p_track(out + 2, BIT_STR(code, 2*m - 2, 2*m - 1),
404 m - 1, off + half_4p);
405 break;
406 case 1: /* 1 pulse in A, 3 pulses in B */
407 decode_1p_track(out, BIT_STR(code, 3*m - 2, m),
408 m - 1, off);
409 decode_3p_track(out + 1, BIT_STR(code, 0, 3*m - 2),
410 m - 1, off + b_offset);
411 break;
412 case 2: /* 2 pulses in each half */
413 decode_2p_track(out, BIT_STR(code, 2*m - 1, 2*m - 1),
414 m - 1, off);
415 decode_2p_track(out + 2, BIT_STR(code, 0, 2*m - 1),
416 m - 1, off + b_offset);
417 break;
418 case 3: /* 3 pulses in A, 1 pulse in B */
419 decode_3p_track(out, BIT_STR(code, m, 3*m - 2),
420 m - 1, off);
421 decode_1p_track(out + 3, BIT_STR(code, 0, m),
422 m - 1, off + b_offset);
423 break;
424 }
425 }
426
427 static void decode_5p_track(int *out, int code, int m, int off) ///code: 5m bits
428 {
429 int half_3p = BIT_POS(code, 5*m - 1) << (m - 1);
430
431 decode_3p_track(out, BIT_STR(code, 2*m + 1, 3*m - 2),
432 m - 1, off + half_3p);
433
434 decode_2p_track(out + 3, BIT_STR(code, 0, 2*m + 1), m, off);
435 }
436
437 static void decode_6p_track(int *out, int code, int m, int off) ///code: 6m-2 bits
438 {
439 int b_offset = 1 << (m - 1);
440 /* which half has more pulses in cases 0 to 2 */
441 int half_more = BIT_POS(code, 6*m - 5) << (m - 1);
442 int half_other = b_offset - half_more;
443
444 switch (BIT_STR(code, 6*m - 4, 2)) { /* case ID (2 bits) */
445 case 0: /* 0 pulses in A, 6 pulses in B or vice versa */
446 decode_1p_track(out, BIT_STR(code, 0, m),
447 m - 1, off + half_more);
448 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
449 m - 1, off + half_more);
450 break;
451 case 1: /* 1 pulse in A, 5 pulses in B or vice versa */
452 decode_1p_track(out, BIT_STR(code, 0, m),
453 m - 1, off + half_other);
454 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
455 m - 1, off + half_more);
456 break;
457 case 2: /* 2 pulses in A, 4 pulses in B or vice versa */
458 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
459 m - 1, off + half_other);
460 decode_4p_track(out + 2, BIT_STR(code, 2*m - 1, 4*m - 4),
461 m - 1, off + half_more);
462 break;
463 case 3: /* 3 pulses in A, 3 pulses in B */
464 decode_3p_track(out, BIT_STR(code, 3*m - 2, 3*m - 2),
465 m - 1, off);
466 decode_3p_track(out + 3, BIT_STR(code, 0, 3*m - 2),
467 m - 1, off + b_offset);
468 break;
469 }
470 }
471
472 /**
473 * Decode the algebraic codebook index to pulse positions and signs,
474 * then construct the algebraic codebook vector.
475 *
476 * @param[out] fixed_vector Buffer for the fixed codebook excitation
477 * @param[in] pulse_hi MSBs part of the pulse index array (higher modes only)
478 * @param[in] pulse_lo LSBs part of the pulse index array
479 * @param[in] mode Mode of the current frame
480 */
481 static void decode_fixed_vector(float *fixed_vector, const uint16_t *pulse_hi,
482 const uint16_t *pulse_lo, const enum Mode mode)
483 {
484 /* sig_pos stores for each track the decoded pulse position indexes
485 * (1-based) multiplied by its corresponding amplitude (+1 or -1) */
486 int sig_pos[4][6];
487 int spacing = (mode == MODE_6k60) ? 2 : 4;
488 int i, j;
489
490 switch (mode) {
491 case MODE_6k60:
492 for (i = 0; i < 2; i++)
493 decode_1p_track(sig_pos[i], pulse_lo[i], 5, 1);
494 break;
495 case MODE_8k85:
496 for (i = 0; i < 4; i++)
497 decode_1p_track(sig_pos[i], pulse_lo[i], 4, 1);
498 break;
499 case MODE_12k65:
500 for (i = 0; i < 4; i++)
501 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
502 break;
503 case MODE_14k25:
504 for (i = 0; i < 2; i++)
505 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
506 for (i = 2; i < 4; i++)
507 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
508 break;
509 case MODE_15k85:
510 for (i = 0; i < 4; i++)
511 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
512 break;
513 case MODE_18k25:
514 for (i = 0; i < 4; i++)
515 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
516 ((int) pulse_hi[i] << 14), 4, 1);
517 break;
518 case MODE_19k85:
519 for (i = 0; i < 2; i++)
520 decode_5p_track(sig_pos[i], (int) pulse_lo[i] +
521 ((int) pulse_hi[i] << 10), 4, 1);
522 for (i = 2; i < 4; i++)
523 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
524 ((int) pulse_hi[i] << 14), 4, 1);
525 break;
526 case MODE_23k05:
527 case MODE_23k85:
528 for (i = 0; i < 4; i++)
529 decode_6p_track(sig_pos[i], (int) pulse_lo[i] +
530 ((int) pulse_hi[i] << 11), 4, 1);
531 break;
532 }
533
534 memset(fixed_vector, 0, sizeof(float) * AMRWB_SFR_SIZE);
535
536 for (i = 0; i < 4; i++)
537 for (j = 0; j < pulses_nb_per_mode_tr[mode][i]; j++) {
538 int pos = (FFABS(sig_pos[i][j]) - 1) * spacing + i;
539
540 fixed_vector[pos] += sig_pos[i][j] < 0 ? -1.0 : 1.0;
541 }
542 }
543
544 /**
545 * Decode pitch gain and fixed gain correction factor.
546 *
547 * @param[in] vq_gain Vector-quantized index for gains
548 * @param[in] mode Mode of the current frame
549 * @param[out] fixed_gain_factor Decoded fixed gain correction factor
550 * @param[out] pitch_gain Decoded pitch gain
551 */
552 static void decode_gains(const uint8_t vq_gain, const enum Mode mode,
553 float *fixed_gain_factor, float *pitch_gain)
554 {
555 const int16_t *gains = (mode <= MODE_8k85 ? qua_gain_6b[vq_gain] :
556 qua_gain_7b[vq_gain]);
557
558 *pitch_gain = gains[0] * (1.0f / (1 << 14));
559 *fixed_gain_factor = gains[1] * (1.0f / (1 << 11));
560 }
561
562 /**
563 * Apply pitch sharpening filters to the fixed codebook vector.
564 *
565 * @param[in] ctx The context
566 * @param[in,out] fixed_vector Fixed codebook excitation
567 */
568 // XXX: Spec states this procedure should be applied when the pitch
569 // lag is less than 64, but this checking seems absent in reference and AMR-NB
570 static void pitch_sharpening(AMRWBContext *ctx, float *fixed_vector)
571 {
572 int i;
573
574 /* Tilt part */
575 for (i = AMRWB_SFR_SIZE - 1; i != 0; i--)
576 fixed_vector[i] -= fixed_vector[i - 1] * ctx->tilt_coef;
577
578 /* Periodicity enhancement part */
579 for (i = ctx->pitch_lag_int; i < AMRWB_SFR_SIZE; i++)
580 fixed_vector[i] += fixed_vector[i - ctx->pitch_lag_int] * 0.85;
581 }
582
583 /**
584 * Calculate the voicing factor (-1.0 = unvoiced to 1.0 = voiced).
585 *
586 * @param[in] p_vector, f_vector Pitch and fixed excitation vectors
587 * @param[in] p_gain, f_gain Pitch and fixed gains
588 */
589 // XXX: There is something wrong with the precision here! The magnitudes
590 // of the energies are not correct. Please check the reference code carefully
591 static float voice_factor(float *p_vector, float p_gain,
592 float *f_vector, float f_gain)
593 {
594 double p_ener = (double) avpriv_scalarproduct_float_c(p_vector, p_vector,
595 AMRWB_SFR_SIZE) *
596 p_gain * p_gain;
597 double f_ener = (double) avpriv_scalarproduct_float_c(f_vector, f_vector,
598 AMRWB_SFR_SIZE) *
599 f_gain * f_gain;
600
601 return (p_ener - f_ener) / (p_ener + f_ener);
602 }
603
604 /**
605 * Reduce fixed vector sparseness by smoothing with one of three IR filters,
606 * also known as "adaptive phase dispersion".
607 *
608 * @param[in] ctx The context
609 * @param[in,out] fixed_vector Unfiltered fixed vector
610 * @param[out] buf Space for modified vector if necessary
611 *
612 * @return The potentially overwritten filtered fixed vector address
613 */
614 static float *anti_sparseness(AMRWBContext *ctx,
615 float *fixed_vector, float *buf)
616 {
617 int ir_filter_nr;
618
619 if (ctx->fr_cur_mode > MODE_8k85) // no filtering in higher modes
620 return fixed_vector;
621
622 if (ctx->pitch_gain[0] < 0.6) {
623 ir_filter_nr = 0; // strong filtering
624 } else if (ctx->pitch_gain[0] < 0.9) {
625 ir_filter_nr = 1; // medium filtering
626 } else
627 ir_filter_nr = 2; // no filtering
628
629 /* detect 'onset' */
630 if (ctx->fixed_gain[0] > 3.0 * ctx->fixed_gain[1]) {
631 if (ir_filter_nr < 2)
632 ir_filter_nr++;
633 } else {
634 int i, count = 0;
635
636 for (i = 0; i < 6; i++)
637 if (ctx->pitch_gain[i] < 0.6)
638 count++;
639
640 if (count > 2)
641 ir_filter_nr = 0;
642
643 if (ir_filter_nr > ctx->prev_ir_filter_nr + 1)
644 ir_filter_nr--;
645 }
646
647 /* update ir filter strength history */
648 ctx->prev_ir_filter_nr = ir_filter_nr;
649
650 ir_filter_nr += (ctx->fr_cur_mode == MODE_8k85);
651
652 if (ir_filter_nr < 2) {
653 int i;
654 const float *coef = ir_filters_lookup[ir_filter_nr];
655
656 /* Circular convolution code in the reference
657 * decoder was modified to avoid using one
658 * extra array. The filtered vector is given by:
659 *
660 * c2(n) = sum(i,0,len-1){ c(i) * coef( (n - i + len) % len ) }
661 */
662
663 memset(buf, 0, sizeof(float) * AMRWB_SFR_SIZE);
664 for (i = 0; i < AMRWB_SFR_SIZE; i++)
665 if (fixed_vector[i])
666 ff_celp_circ_addf(buf, buf, coef, i, fixed_vector[i],
667 AMRWB_SFR_SIZE);
668 fixed_vector = buf;
669 }
670
671 return fixed_vector;
672 }
673
674 /**
675 * Calculate a stability factor {teta} based on distance between
676 * current and past isf. A value of 1 shows maximum signal stability.
677 */
678 static float stability_factor(const float *isf, const float *isf_past)
679 {
680 int i;
681 float acc = 0.0;
682
683 for (i = 0; i < LP_ORDER - 1; i++)
684 acc += (isf[i] - isf_past[i]) * (isf[i] - isf_past[i]);
685
686 // XXX: This part is not so clear from the reference code
687 // the result is more accurate changing the "/ 256" to "* 512"
688 return FFMAX(0.0, 1.25 - acc * 0.8 * 512);
689 }
690
691 /**
692 * Apply a non-linear fixed gain smoothing in order to reduce
693 * fluctuation in the energy of excitation.
694 *
695 * @param[in] fixed_gain Unsmoothed fixed gain
696 * @param[in,out] prev_tr_gain Previous threshold gain (updated)
697 * @param[in] voice_fac Frame voicing factor
698 * @param[in] stab_fac Frame stability factor
699 *
700 * @return The smoothed gain
701 */
702 static float noise_enhancer(float fixed_gain, float *prev_tr_gain,
703 float voice_fac, float stab_fac)
704 {
705 float sm_fac = 0.5 * (1 - voice_fac) * stab_fac;
706 float g0;
707
708 // XXX: the following fixed-point constants used to in(de)crement
709 // gain by 1.5dB were taken from the reference code, maybe it could
710 // be simpler
711 if (fixed_gain < *prev_tr_gain) {
712 g0 = FFMIN(*prev_tr_gain, fixed_gain + fixed_gain *
713 (6226 * (1.0f / (1 << 15)))); // +1.5 dB
714 } else
715 g0 = FFMAX(*prev_tr_gain, fixed_gain *
716 (27536 * (1.0f / (1 << 15)))); // -1.5 dB
717
718 *prev_tr_gain = g0; // update next frame threshold
719
720 return sm_fac * g0 + (1 - sm_fac) * fixed_gain;
721 }
722
723 /**
724 * Filter the fixed_vector to emphasize the higher frequencies.
725 *
726 * @param[in,out] fixed_vector Fixed codebook vector
727 * @param[in] voice_fac Frame voicing factor
728 */
729 static void pitch_enhancer(float *fixed_vector, float voice_fac)
730 {
731 int i;
732 float cpe = 0.125 * (1 + voice_fac);
733 float last = fixed_vector[0]; // holds c(i - 1)
734
735 fixed_vector[0] -= cpe * fixed_vector[1];
736
737 for (i = 1; i < AMRWB_SFR_SIZE - 1; i++) {
738 float cur = fixed_vector[i];
739
740 fixed_vector[i] -= cpe * (last + fixed_vector[i + 1]);
741 last = cur;
742 }
743
744 fixed_vector[AMRWB_SFR_SIZE - 1] -= cpe * last;
745 }
746
747 /**
748 * Conduct 16th order linear predictive coding synthesis from excitation.
749 *
750 * @param[in] ctx Pointer to the AMRWBContext
751 * @param[in] lpc Pointer to the LPC coefficients
752 * @param[out] excitation Buffer for synthesis final excitation
753 * @param[in] fixed_gain Fixed codebook gain for synthesis
754 * @param[in] fixed_vector Algebraic codebook vector
755 * @param[in,out] samples Pointer to the output samples and memory
756 */
757 static void synthesis(AMRWBContext *ctx, float *lpc, float *excitation,
758 float fixed_gain, const float *fixed_vector,
759 float *samples)
760 {
761 ff_weighted_vector_sumf(excitation, ctx->pitch_vector, fixed_vector,
762 ctx->pitch_gain[0], fixed_gain, AMRWB_SFR_SIZE);
763
764 /* emphasize pitch vector contribution in low bitrate modes */
765 if (ctx->pitch_gain[0] > 0.5 && ctx->fr_cur_mode <= MODE_8k85) {
766 int i;
767 float energy = avpriv_scalarproduct_float_c(excitation, excitation,
768 AMRWB_SFR_SIZE);
769
770 // XXX: Weird part in both ref code and spec. A unknown parameter
771 // {beta} seems to be identical to the current pitch gain
772 float pitch_factor = 0.25 * ctx->pitch_gain[0] * ctx->pitch_gain[0];
773
774 for (i = 0; i < AMRWB_SFR_SIZE; i++)
775 excitation[i] += pitch_factor * ctx->pitch_vector[i];
776
777 ff_scale_vector_to_given_sum_of_squares(excitation, excitation,
778 energy, AMRWB_SFR_SIZE);
779 }
780
781 ff_celp_lp_synthesis_filterf(samples, lpc, excitation,
782 AMRWB_SFR_SIZE, LP_ORDER);
783 }
784
785 /**
786 * Apply to synthesis a de-emphasis filter of the form:
787 * H(z) = 1 / (1 - m * z^-1)
788 *
789 * @param[out] out Output buffer
790 * @param[in] in Input samples array with in[-1]
791 * @param[in] m Filter coefficient
792 * @param[in,out] mem State from last filtering
793 */
794 static void de_emphasis(float *out, float *in, float m, float mem[1])
795 {
796 int i;
797
798 out[0] = in[0] + m * mem[0];
799
800 for (i = 1; i < AMRWB_SFR_SIZE; i++)
801 out[i] = in[i] + out[i - 1] * m;
802
803 mem[0] = out[AMRWB_SFR_SIZE - 1];
804 }
805
806 /**
807 * Upsample a signal by 5/4 ratio (from 12.8kHz to 16kHz) using
808 * a FIR interpolation filter. Uses past data from before *in address.
809 *
810 * @param[out] out Buffer for interpolated signal
811 * @param[in] in Current signal data (length 0.8*o_size)
812 * @param[in] o_size Output signal length
813 */
814 static void upsample_5_4(float *out, const float *in, int o_size)
815 {
816 const float *in0 = in - UPS_FIR_SIZE + 1;
817 int i, j, k;
818 int int_part = 0, frac_part;
819
820 i = 0;
821 for (j = 0; j < o_size / 5; j++) {
822 out[i] = in[int_part];
823 frac_part = 4;
824 i++;
825
826 for (k = 1; k < 5; k++) {
827 out[i] = avpriv_scalarproduct_float_c(in0 + int_part,
828 upsample_fir[4 - frac_part],
829 UPS_MEM_SIZE);
830 int_part++;
831 frac_part--;
832 i++;
833 }
834 }
835 }
836
837 /**
838 * Calculate the high-band gain based on encoded index (23k85 mode) or
839 * on the low-band speech signal and the Voice Activity Detection flag.
840 *
841 * @param[in] ctx The context
842 * @param[in] synth LB speech synthesis at 12.8k
843 * @param[in] hb_idx Gain index for mode 23k85 only
844 * @param[in] vad VAD flag for the frame
845 */
846 static float find_hb_gain(AMRWBContext *ctx, const float *synth,
847 uint16_t hb_idx, uint8_t vad)
848 {
849 int wsp = (vad > 0);
850 float tilt;
851
852 if (ctx->fr_cur_mode == MODE_23k85)
853 return qua_hb_gain[hb_idx] * (1.0f / (1 << 14));
854
855 tilt = avpriv_scalarproduct_float_c(synth, synth + 1, AMRWB_SFR_SIZE - 1) /
856 avpriv_scalarproduct_float_c(synth, synth, AMRWB_SFR_SIZE);
857
858 /* return gain bounded by [0.1, 1.0] */
859 return av_clipf((1.0 - FFMAX(0.0, tilt)) * (1.25 - 0.25 * wsp), 0.1, 1.0);
860 }
861
862 /**
863 * Generate the high-band excitation with the same energy from the lower
864 * one and scaled by the given gain.
865 *
866 * @param[in] ctx The context
867 * @param[out] hb_exc Buffer for the excitation
868 * @param[in] synth_exc Low-band excitation used for synthesis
869 * @param[in] hb_gain Wanted excitation gain
870 */
871 static void scaled_hb_excitation(AMRWBContext *ctx, float *hb_exc,
872 const float *synth_exc, float hb_gain)
873 {
874 int i;
875 float energy = avpriv_scalarproduct_float_c(synth_exc, synth_exc,
876 AMRWB_SFR_SIZE);
877
878 /* Generate a white-noise excitation */
879 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
880 hb_exc[i] = 32768.0 - (uint16_t) av_lfg_get(&ctx->prng);
881
882 ff_scale_vector_to_given_sum_of_squares(hb_exc, hb_exc,
883 energy * hb_gain * hb_gain,
884 AMRWB_SFR_SIZE_16k);
885 }
886
887 /**
888 * Calculate the auto-correlation for the ISF difference vector.
889 */
890 static float auto_correlation(float *diff_isf, float mean, int lag)
891 {
892 int i;
893 float sum = 0.0;
894
895 for (i = 7; i < LP_ORDER - 2; i++) {
896 float prod = (diff_isf[i] - mean) * (diff_isf[i - lag] - mean);
897 sum += prod * prod;
898 }
899 return sum;
900 }
901
902 /**
903 * Extrapolate a ISF vector to the 16kHz range (20th order LP)
904 * used at mode 6k60 LP filter for the high frequency band.
905 *
906 * @param[out] isf Buffer for extrapolated isf; contains LP_ORDER
907 * values on input
908 */
909 static void extrapolate_isf(float isf[LP_ORDER_16k])
910 {
911 float diff_isf[LP_ORDER - 2], diff_mean;
912 float corr_lag[3];
913 float est, scale;
914 int i, j, i_max_corr;
915
916 isf[LP_ORDER_16k - 1] = isf[LP_ORDER - 1];
917
918 /* Calculate the difference vector */
919 for (i = 0; i < LP_ORDER - 2; i++)
920 diff_isf[i] = isf[i + 1] - isf[i];
921
922 diff_mean = 0.0;
923 for (i = 2; i < LP_ORDER - 2; i++)
924 diff_mean += diff_isf[i] * (1.0f / (LP_ORDER - 4));
925
926 /* Find which is the maximum autocorrelation */
927 i_max_corr = 0;
928 for (i = 0; i < 3; i++) {
929 corr_lag[i] = auto_correlation(diff_isf, diff_mean, i + 2);
930
931 if (corr_lag[i] > corr_lag[i_max_corr])
932 i_max_corr = i;
933 }
934 i_max_corr++;
935
936 for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
937 isf[i] = isf[i - 1] + isf[i - 1 - i_max_corr]
938 - isf[i - 2 - i_max_corr];
939
940 /* Calculate an estimate for ISF(18) and scale ISF based on the error */
941 est = 7965 + (isf[2] - isf[3] - isf[4]) / 6.0;
942 scale = 0.5 * (FFMIN(est, 7600) - isf[LP_ORDER - 2]) /
943 (isf[LP_ORDER_16k - 2] - isf[LP_ORDER - 2]);
944
945 for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++)
946 diff_isf[j] = scale * (isf[i] - isf[i - 1]);
947
948 /* Stability insurance */
949 for (i = 1; i < LP_ORDER_16k - LP_ORDER; i++)
950 if (diff_isf[i] + diff_isf[i - 1] < 5.0) {
951 if (diff_isf[i] > diff_isf[i - 1]) {
952 diff_isf[i - 1] = 5.0 - diff_isf[i];
953 } else
954 diff_isf[i] = 5.0 - diff_isf[i - 1];
955 }
956
957 for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++)
958 isf[i] = isf[i - 1] + diff_isf[j] * (1.0f / (1 << 15));
959
960 /* Scale the ISF vector for 16000 Hz */
961 for (i = 0; i < LP_ORDER_16k - 1; i++)
962 isf[i] *= 0.8;
963 }
964
965 /**
966 * Spectral expand the LP coefficients using the equation:
967 * y[i] = x[i] * (gamma ** i)
968 *
969 * @param[out] out Output buffer (may use input array)
970 * @param[in] lpc LP coefficients array
971 * @param[in] gamma Weighting factor
972 * @param[in] size LP array size
973 */
974 static void lpc_weighting(float *out, const float *lpc, float gamma, int size)
975 {
976 int i;
977 float fac = gamma;
978
979 for (i = 0; i < size; i++) {
980 out[i] = lpc[i] * fac;
981 fac *= gamma;
982 }
983 }
984
985 /**
986 * Conduct 20th order linear predictive coding synthesis for the high
987 * frequency band excitation at 16kHz.
988 *
989 * @param[in] ctx The context
990 * @param[in] subframe Current subframe index (0 to 3)
991 * @param[in,out] samples Pointer to the output speech samples
992 * @param[in] exc Generated white-noise scaled excitation
993 * @param[in] isf Current frame isf vector
994 * @param[in] isf_past Past frame final isf vector
995 */
996 static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples,
997 const float *exc, const float *isf, const float *isf_past)
998 {
999 float hb_lpc[LP_ORDER_16k];
1000 enum Mode mode = ctx->fr_cur_mode;
1001
1002 if (mode == MODE_6k60) {
1003 float e_isf[LP_ORDER_16k]; // ISF vector for extrapolation
1004 double e_isp[LP_ORDER_16k];
1005
1006 ff_weighted_vector_sumf(e_isf, isf_past, isf, isfp_inter[subframe],
1007 1.0 - isfp_inter[subframe], LP_ORDER);
1008
1009 extrapolate_isf(e_isf);
1010
1011 e_isf[LP_ORDER_16k - 1] *= 2.0;
1012 ff_acelp_lsf2lspd(e_isp, e_isf, LP_ORDER_16k);
1013 ff_amrwb_lsp2lpc(e_isp, hb_lpc, LP_ORDER_16k);
1014
1015 lpc_weighting(hb_lpc, hb_lpc, 0.9, LP_ORDER_16k);
1016 } else {
1017 lpc_weighting(hb_lpc, ctx->lp_coef[subframe], 0.6, LP_ORDER);
1018 }
1019
1020 ff_celp_lp_synthesis_filterf(samples, hb_lpc, exc, AMRWB_SFR_SIZE_16k,
1021 (mode == MODE_6k60) ? LP_ORDER_16k : LP_ORDER);
1022 }
1023
1024 /**
1025 * Apply a 15th order filter to high-band samples.
1026 * The filter characteristic depends on the given coefficients.
1027 *
1028 * @param[out] out Buffer for filtered output
1029 * @param[in] fir_coef Filter coefficients
1030 * @param[in,out] mem State from last filtering (updated)
1031 * @param[in] in Input speech data (high-band)
1032 *
1033 * @remark It is safe to pass the same array in in and out parameters
1034 */
1035 static void hb_fir_filter(float *out, const float fir_coef[HB_FIR_SIZE + 1],
1036 float mem[HB_FIR_SIZE], const float *in)
1037 {
1038 int i, j;
1039 float data[AMRWB_SFR_SIZE_16k + HB_FIR_SIZE]; // past and current samples
1040
1041 memcpy(data, mem, HB_FIR_SIZE * sizeof(float));
1042 memcpy(data + HB_FIR_SIZE, in, AMRWB_SFR_SIZE_16k * sizeof(float));
1043
1044 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) {
1045 out[i] = 0.0;
1046 for (j = 0; j <= HB_FIR_SIZE; j++)
1047 out[i] += data[i + j] * fir_coef[j];
1048 }
1049
1050 memcpy(mem, data + AMRWB_SFR_SIZE_16k, HB_FIR_SIZE * sizeof(float));
1051 }
1052
1053 /**
1054 * Update context state before the next subframe.
1055 */
1056 static void update_sub_state(AMRWBContext *ctx)
1057 {
1058 memmove(&ctx->excitation_buf[0], &ctx->excitation_buf[AMRWB_SFR_SIZE],
1059 (AMRWB_P_DELAY_MAX + LP_ORDER + 1) * sizeof(float));
1060
1061 memmove(&ctx->pitch_gain[1], &ctx->pitch_gain[0], 5 * sizeof(float));
1062 memmove(&ctx->fixed_gain[1], &ctx->fixed_gain[0], 1 * sizeof(float));
1063
1064 memmove(&ctx->samples_az[0], &ctx->samples_az[AMRWB_SFR_SIZE],
1065 LP_ORDER * sizeof(float));
1066 memmove(&ctx->samples_up[0], &ctx->samples_up[AMRWB_SFR_SIZE],
1067 UPS_MEM_SIZE * sizeof(float));
1068 memmove(&ctx->samples_hb[0], &ctx->samples_hb[AMRWB_SFR_SIZE_16k],
1069 LP_ORDER_16k * sizeof(float));
1070 }
1071
1072 static int amrwb_decode_frame(AVCodecContext *avctx, void *data,
1073 int *got_frame_ptr, AVPacket *avpkt)
1074 {
1075 AMRWBContext *ctx = avctx->priv_data;
1076 AVFrame *frame = data;
1077 AMRWBFrame *cf = &ctx->frame;
1078 const uint8_t *buf = avpkt->data;
1079 int buf_size = avpkt->size;
1080 int expected_fr_size, header_size;
1081 float *buf_out;
1082 float spare_vector[AMRWB_SFR_SIZE]; // extra stack space to hold result from anti-sparseness processing
1083 float fixed_gain_factor; // fixed gain correction factor (gamma)
1084 float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
1085 float synth_fixed_gain; // the fixed gain that synthesis should use
1086 float voice_fac, stab_fac; // parameters used for gain smoothing
1087 float synth_exc[AMRWB_SFR_SIZE]; // post-processed excitation for synthesis
1088 float hb_exc[AMRWB_SFR_SIZE_16k]; // excitation for the high frequency band
1089 float hb_samples[AMRWB_SFR_SIZE_16k]; // filtered high-band samples from synthesis
1090 float hb_gain;
1091 int sub, i, ret;
1092
1093 /* get output buffer */
1094 frame->nb_samples = 4 * AMRWB_SFR_SIZE_16k;
1095 if ((ret = ff_get_buffer(avctx, frame)) < 0) {
1096 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1097 return ret;
1098 }
1099 buf_out = (float *)frame->data[0];
1100
1101 header_size = decode_mime_header(ctx, buf);
1102 if (ctx->fr_cur_mode > MODE_SID) {
1103 av_log(avctx, AV_LOG_ERROR,
1104 "Invalid mode %d\n", ctx->fr_cur_mode);
1105 return AVERROR_INVALIDDATA;
1106 }
1107 expected_fr_size = ((cf_sizes_wb[ctx->fr_cur_mode] + 7) >> 3) + 1;
1108
1109 if (buf_size < expected_fr_size) {
1110 av_log(avctx, AV_LOG_ERROR,
1111 "Frame too small (%d bytes). Truncated file?\n", buf_size);
1112 *got_frame_ptr = 0;
1113 return AVERROR_INVALIDDATA;
1114 }
1115
1116 if (!ctx->fr_quality || ctx->fr_cur_mode > MODE_SID)
1117 av_log(avctx, AV_LOG_ERROR, "Encountered a bad or corrupted frame\n");
1118
1119 if (ctx->fr_cur_mode == MODE_SID) { /* Comfort noise frame */
1120 av_log_missing_feature(avctx, "SID mode", 1);
1121 return AVERROR_PATCHWELCOME;
1122 }
1123
1124 ff_amr_bit_reorder((uint16_t *) &ctx->frame, sizeof(AMRWBFrame),
1125 buf + header_size, amr_bit_orderings_by_mode[ctx->fr_cur_mode]);
1126
1127 /* Decode the quantized ISF vector */
1128 if (ctx->fr_cur_mode == MODE_6k60) {
1129 decode_isf_indices_36b(cf->isp_id, ctx->isf_cur);
1130 } else {
1131 decode_isf_indices_46b(cf->isp_id, ctx->isf_cur);
1132 }
1133
1134 isf_add_mean_and_past(ctx->isf_cur, ctx->isf_q_past);
1135 ff_set_min_dist_lsf(ctx->isf_cur, MIN_ISF_SPACING, LP_ORDER - 1);
1136
1137 stab_fac = stability_factor(ctx->isf_cur, ctx->isf_past_final);
1138
1139 ctx->isf_cur[LP_ORDER - 1] *= 2.0;
1140 ff_acelp_lsf2lspd(ctx->isp[3], ctx->isf_cur, LP_ORDER);
1141
1142 /* Generate a ISP vector for each subframe */
1143 if (ctx->first_frame) {
1144 ctx->first_frame = 0;
1145 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(double));
1146 }
1147 interpolate_isp(ctx->isp, ctx->isp_sub4_past);
1148
1149 for (sub = 0; sub < 4; sub++)
1150 ff_amrwb_lsp2lpc(ctx->isp[sub], ctx->lp_coef[sub], LP_ORDER);
1151
1152 for (sub = 0; sub < 4; sub++) {
1153 const AMRWBSubFrame *cur_subframe = &cf->subframe[sub];
1154 float *sub_buf = buf_out + sub * AMRWB_SFR_SIZE_16k;
1155
1156 /* Decode adaptive codebook (pitch vector) */
1157 decode_pitch_vector(ctx, cur_subframe, sub);
1158 /* Decode innovative codebook (fixed vector) */
1159 decode_fixed_vector(ctx->fixed_vector, cur_subframe->pul_ih,
1160 cur_subframe->pul_il, ctx->fr_cur_mode);
1161
1162 pitch_sharpening(ctx, ctx->fixed_vector);
1163
1164 decode_gains(cur_subframe->vq_gain, ctx->fr_cur_mode,
1165 &fixed_gain_factor, &ctx->pitch_gain[0]);
1166
1167 ctx->fixed_gain[0] =
1168 ff_amr_set_fixed_gain(fixed_gain_factor,
1169 avpriv_scalarproduct_float_c(ctx->fixed_vector,
1170 ctx->fixed_vector,
1171 AMRWB_SFR_SIZE) /
1172 AMRWB_SFR_SIZE,
1173 ctx->prediction_error,
1174 ENERGY_MEAN, energy_pred_fac);
1175
1176 /* Calculate voice factor and store tilt for next subframe */
1177 voice_fac = voice_factor(ctx->pitch_vector, ctx->pitch_gain[0],
1178 ctx->fixed_vector, ctx->fixed_gain[0]);
1179 ctx->tilt_coef = voice_fac * 0.25 + 0.25;
1180
1181 /* Construct current excitation */
1182 for (i = 0; i < AMRWB_SFR_SIZE; i++) {
1183 ctx->excitation[i] *= ctx->pitch_gain[0];
1184 ctx->excitation[i] += ctx->fixed_gain[0] * ctx->fixed_vector[i];
1185 ctx->excitation[i] = truncf(ctx->excitation[i]);
1186 }
1187
1188 /* Post-processing of excitation elements */
1189 synth_fixed_gain = noise_enhancer(ctx->fixed_gain[0], &ctx->prev_tr_gain,
1190 voice_fac, stab_fac);
1191
1192 synth_fixed_vector = anti_sparseness(ctx, ctx->fixed_vector,
1193 spare_vector);
1194
1195 pitch_enhancer(synth_fixed_vector, voice_fac);
1196
1197 synthesis(ctx, ctx->lp_coef[sub], synth_exc, synth_fixed_gain,
1198 synth_fixed_vector, &ctx->samples_az[LP_ORDER]);
1199
1200 /* Synthesis speech post-processing */
1201 de_emphasis(&ctx->samples_up[UPS_MEM_SIZE],
1202 &ctx->samples_az[LP_ORDER], PREEMPH_FAC, ctx->demph_mem);
1203
1204 ff_acelp_apply_order_2_transfer_function(&ctx->samples_up[UPS_MEM_SIZE],
1205 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_31_poles,
1206 hpf_31_gain, ctx->hpf_31_mem, AMRWB_SFR_SIZE);
1207
1208 upsample_5_4(sub_buf, &ctx->samples_up[UPS_FIR_SIZE],
1209 AMRWB_SFR_SIZE_16k);
1210
1211 /* High frequency band (6.4 - 7.0 kHz) generation part */
1212 ff_acelp_apply_order_2_transfer_function(hb_samples,
1213 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_400_poles,
1214 hpf_400_gain, ctx->hpf_400_mem, AMRWB_SFR_SIZE);
1215
1216 hb_gain = find_hb_gain(ctx, hb_samples,
1217 cur_subframe->hb_gain, cf->vad);
1218
1219 scaled_hb_excitation(ctx, hb_exc, synth_exc, hb_gain);
1220
1221 hb_synthesis(ctx, sub, &ctx->samples_hb[LP_ORDER_16k],
1222 hb_exc, ctx->isf_cur, ctx->isf_past_final);
1223
1224 /* High-band post-processing filters */
1225 hb_fir_filter(hb_samples, bpf_6_7_coef, ctx->bpf_6_7_mem,
1226 &ctx->samples_hb[LP_ORDER_16k]);
1227
1228 if (ctx->fr_cur_mode == MODE_23k85)
1229 hb_fir_filter(hb_samples, lpf_7_coef, ctx->lpf_7_mem,
1230 hb_samples);
1231
1232 /* Add the low and high frequency bands */
1233 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
1234 sub_buf[i] = (sub_buf[i] + hb_samples[i]) * (1.0f / (1 << 15));
1235
1236 /* Update buffers and history */
1237 update_sub_state(ctx);
1238 }
1239
1240 /* update state for next frame */
1241 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(ctx->isp[3][0]));
1242 memcpy(ctx->isf_past_final, ctx->isf_cur, LP_ORDER * sizeof(float));
1243
1244 *got_frame_ptr = 1;
1245
1246 return expected_fr_size;
1247 }
1248
1249 AVCodec ff_amrwb_decoder = {
1250 .name = "amrwb",
1251 .type = AVMEDIA_TYPE_AUDIO,
1252 .id = AV_CODEC_ID_AMR_WB,
1253 .priv_data_size = sizeof(AMRWBContext),
1254 .init = amrwb_decode_init,
1255 .decode = amrwb_decode_frame,
1256 .capabilities = CODEC_CAP_DR1,
1257 .long_name = NULL_IF_CONFIG_SMALL("AMR-WB (Adaptive Multi-Rate WideBand)"),
1258 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT,
1259 AV_SAMPLE_FMT_NONE },
1260 };