1 /*
2 * AMR wideband decoder
3 * Copyright (c) 2010 Marcelo Galvao Povoa
4 *
5 * This file is part of FFmpeg.
6 *
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
11 *
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A particular PURPOSE. See the GNU
15 * Lesser General Public License for more details.
16 *
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 */
21
22 /**
23 * @file
24 * 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 "celp_math.h"
36 #include "acelp_filters.h"
37 #include "acelp_vectors.h"
38 #include "acelp_pitch_delay.h"
39 #include "internal.h"
40
41 #define AMR_USE_16BIT_TABLES
42 #include "amr.h"
43
44 #include "amrwbdata.h"
45 #include "mips/amrwbdec_mips.h"
46
47 typedef struct AMRWBContext {
48 AMRWBFrame frame; ///< AMRWB parameters decoded from bitstream
49 enum Mode fr_cur_mode; ///< mode index of current frame
50 uint8_t fr_quality; ///< frame quality index (FQI)
51 float isf_cur[LP_ORDER]; ///< working ISF vector from current frame
52 float isf_q_past[LP_ORDER]; ///< quantized ISF vector of the previous frame
53 float isf_past_final[LP_ORDER]; ///< final processed ISF vector of the previous frame
54 double isp[4][LP_ORDER]; ///< ISP vectors from current frame
55 double isp_sub4_past[LP_ORDER]; ///< ISP vector for the 4th subframe of the previous frame
56
57 float lp_coef[4][LP_ORDER]; ///< Linear Prediction Coefficients from ISP vector
58
59 uint8_t base_pitch_lag; ///< integer part of pitch lag for the next relative subframe
60 uint8_t pitch_lag_int; ///< integer part of pitch lag of the previous subframe
61
62 float excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 2 + AMRWB_SFR_SIZE]; ///< current excitation and all necessary excitation history
63 float *excitation; ///< points to current excitation in excitation_buf[]
64
65 float pitch_vector[AMRWB_SFR_SIZE]; ///< adaptive codebook (pitch) vector for current subframe
66 float fixed_vector[AMRWB_SFR_SIZE]; ///< algebraic codebook (fixed) vector for current subframe
67
68 float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes
69 float pitch_gain[6]; ///< quantified pitch gains for the current and previous five subframes
70 float fixed_gain[2]; ///< quantified fixed gains for the current and previous subframes
71
72 float tilt_coef; ///< {beta_1} related to the voicing of the previous subframe
73
74 float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness to determine "onset"
75 uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none
76 float prev_tr_gain; ///< previous initial gain used by noise enhancer for threshold
77
78 float samples_az[LP_ORDER + AMRWB_SFR_SIZE]; ///< low-band samples and memory from synthesis at 12.8kHz
79 float samples_up[UPS_MEM_SIZE + AMRWB_SFR_SIZE]; ///< low-band samples and memory processed for upsampling
80 float samples_hb[LP_ORDER_16k + AMRWB_SFR_SIZE_16k]; ///< high-band samples and memory from synthesis at 16kHz
81
82 float hpf_31_mem[2], hpf_400_mem[2]; ///< previous values in the high pass filters
83 float demph_mem[1]; ///< previous value in the de-emphasis filter
84 float bpf_6_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band band pass filter
85 float lpf_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band low pass filter
86
87 AVLFG prng; ///< random number generator for white noise excitation
88 uint8_t first_frame; ///< flag active during decoding of the first frame
89 ACELPFContext acelpf_ctx; ///< context for filters for ACELP-based codecs
90 ACELPVContext acelpv_ctx; ///< context for vector operations for ACELP-based codecs
91 CELPFContext celpf_ctx; ///< context for filters for CELP-based codecs
92 CELPMContext celpm_ctx; ///< context for fixed point math operations
93
94 } AMRWBContext;
95
amrwb_decode_init(AVCodecContext * avctx)96 static av_cold int amrwb_decode_init(AVCodecContext *avctx)
97 {
98 AMRWBContext *ctx = avctx->priv_data;
99 int i;
100
101 if (avctx->channels > 1) {
102 avpriv_report_missing_feature(avctx, "multi-channel AMR");
103 return AVERROR_PATCHWELCOME;
104 }
105
106 avctx->channels = 1;
107 avctx->channel_layout = AV_CH_LAYOUT_MONO;
108 if (!avctx->sample_rate)
109 avctx->sample_rate = 16000;
110 avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
111
112 av_lfg_init(&ctx->prng, 1);
113
114 ctx->excitation = &ctx->excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 1];
115 ctx->first_frame = 1;
116
117 for (i = 0; i < LP_ORDER; i++)
118 ctx->isf_past_final[i] = isf_init[i] * (1.0f / (1 << 15));
119
120 for (i = 0; i < 4; i++)
121 ctx->prediction_error[i] = MIN_ENERGY;
122
123 ff_acelp_filter_init(&ctx->acelpf_ctx);
124 ff_acelp_vectors_init(&ctx->acelpv_ctx);
125 ff_celp_filter_init(&ctx->celpf_ctx);
126 ff_celp_math_init(&ctx->celpm_ctx);
127
128 return 0;
129 }
130
131 /**
132 * Decode the frame header in the "MIME/storage" format. This format
133 * is simpler and does not carry the auxiliary frame information.
134 *
135 * @param[in] ctx The Context
136 * @param[in] buf Pointer to the input buffer
137 *
138 * @return The decoded header length in bytes
139 */
decode_mime_header(AMRWBContext * ctx,const uint8_t * buf)140 static int decode_mime_header(AMRWBContext *ctx, const uint8_t *buf)
141 {
142 /* Decode frame header (1st octet) */
143 ctx->fr_cur_mode = buf[0] >> 3 & 0x0F;
144 ctx->fr_quality = (buf[0] & 0x4) == 0x4;
145
146 return 1;
147 }
148
149 /**
150 * Decode quantized ISF vectors using 36-bit indexes (6K60 mode only).
151 *
152 * @param[in] ind Array of 5 indexes
153 * @param[out] isf_q Buffer for isf_q[LP_ORDER]
154 */
decode_isf_indices_36b(uint16_t * ind,float * isf_q)155 static void decode_isf_indices_36b(uint16_t *ind, float *isf_q)
156 {
157 int i;
158
159 for (i = 0; i < 9; i++)
160 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
161
162 for (i = 0; i < 7; i++)
163 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
164
165 for (i = 0; i < 5; i++)
166 isf_q[i] += dico21_isf_36b[ind[2]][i] * (1.0f / (1 << 15));
167
168 for (i = 0; i < 4; i++)
169 isf_q[i + 5] += dico22_isf_36b[ind[3]][i] * (1.0f / (1 << 15));
170
171 for (i = 0; i < 7; i++)
172 isf_q[i + 9] += dico23_isf_36b[ind[4]][i] * (1.0f / (1 << 15));
173 }
174
175 /**
176 * Decode quantized ISF vectors using 46-bit indexes (except 6K60 mode).
177 *
178 * @param[in] ind Array of 7 indexes
179 * @param[out] isf_q Buffer for isf_q[LP_ORDER]
180 */
decode_isf_indices_46b(uint16_t * ind,float * isf_q)181 static void decode_isf_indices_46b(uint16_t *ind, float *isf_q)
182 {
183 int i;
184
185 for (i = 0; i < 9; i++)
186 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
187
188 for (i = 0; i < 7; i++)
189 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
190
191 for (i = 0; i < 3; i++)
192 isf_q[i] += dico21_isf[ind[2]][i] * (1.0f / (1 << 15));
193
194 for (i = 0; i < 3; i++)
195 isf_q[i + 3] += dico22_isf[ind[3]][i] * (1.0f / (1 << 15));
196
197 for (i = 0; i < 3; i++)
198 isf_q[i + 6] += dico23_isf[ind[4]][i] * (1.0f / (1 << 15));
199
200 for (i = 0; i < 3; i++)
201 isf_q[i + 9] += dico24_isf[ind[5]][i] * (1.0f / (1 << 15));
202
203 for (i = 0; i < 4; i++)
204 isf_q[i + 12] += dico25_isf[ind[6]][i] * (1.0f / (1 << 15));
205 }
206
207 /**
208 * Apply mean and past ISF values using the prediction factor.
209 * Updates past ISF vector.
210 *
211 * @param[in,out] isf_q Current quantized ISF
212 * @param[in,out] isf_past Past quantized ISF
213 */
isf_add_mean_and_past(float * isf_q,float * isf_past)214 static void isf_add_mean_and_past(float *isf_q, float *isf_past)
215 {
216 int i;
217 float tmp;
218
219 for (i = 0; i < LP_ORDER; i++) {
220 tmp = isf_q[i];
221 isf_q[i] += isf_mean[i] * (1.0f / (1 << 15));
222 isf_q[i] += PRED_FACTOR * isf_past[i];
223 isf_past[i] = tmp;
224 }
225 }
226
227 /**
228 * Interpolate the fourth ISP vector from current and past frames
229 * to obtain an ISP vector for each subframe.
230 *
231 * @param[in,out] isp_q ISPs for each subframe
232 * @param[in] isp4_past Past ISP for subframe 4
233 */
interpolate_isp(double isp_q[4][LP_ORDER],const double * isp4_past)234 static void interpolate_isp(double isp_q[4][LP_ORDER], const double *isp4_past)
235 {
236 int i, k;
237
238 for (k = 0; k < 3; k++) {
239 float c = isfp_inter[k];
240 for (i = 0; i < LP_ORDER; i++)
241 isp_q[k][i] = (1.0 - c) * isp4_past[i] + c * isp_q[3][i];
242 }
243 }
244
245 /**
246 * Decode an adaptive codebook index into pitch lag (except 6k60, 8k85 modes).
247 * Calculate integer lag and fractional lag always using 1/4 resolution.
248 * In 1st and 3rd subframes the index is relative to last subframe integer lag.
249 *
250 * @param[out] lag_int Decoded integer pitch lag
251 * @param[out] lag_frac Decoded fractional pitch lag
252 * @param[in] pitch_index Adaptive codebook pitch index
253 * @param[in,out] base_lag_int Base integer lag used in relative subframes
254 * @param[in] subframe Current subframe index (0 to 3)
255 */
decode_pitch_lag_high(int * lag_int,int * lag_frac,int pitch_index,uint8_t * base_lag_int,int subframe)256 static void decode_pitch_lag_high(int *lag_int, int *lag_frac, int pitch_index,
257 uint8_t *base_lag_int, int subframe)
258 {
259 if (subframe == 0 || subframe == 2) {
260 if (pitch_index < 376) {
261 *lag_int = (pitch_index + 137) >> 2;
262 *lag_frac = pitch_index - (*lag_int << 2) + 136;
263 } else if (pitch_index < 440) {
264 *lag_int = (pitch_index + 257 - 376) >> 1;
265 *lag_frac = (pitch_index - (*lag_int << 1) + 256 - 376) * 2;
266 /* the actual resolution is 1/2 but expressed as 1/4 */
267 } else {
268 *lag_int = pitch_index - 280;
269 *lag_frac = 0;
270 }
271 /* minimum lag for next subframe */
272 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
273 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);
274 // XXX: the spec states clearly that *base_lag_int should be
275 // the nearest integer to *lag_int (minus 8), but the ref code
276 // actually always uses its floor, I'm following the latter
277 } else {
278 *lag_int = (pitch_index + 1) >> 2;
279 *lag_frac = pitch_index - (*lag_int << 2);
280 *lag_int += *base_lag_int;
281 }
282 }
283
284 /**
285 * Decode an adaptive codebook index into pitch lag for 8k85 and 6k60 modes.
286 * The description is analogous to decode_pitch_lag_high, but in 6k60 the
287 * relative index is used for all subframes except the first.
288 */
decode_pitch_lag_low(int * lag_int,int * lag_frac,int pitch_index,uint8_t * base_lag_int,int subframe,enum Mode mode)289 static void decode_pitch_lag_low(int *lag_int, int *lag_frac, int pitch_index,
290 uint8_t *base_lag_int, int subframe, enum Mode mode)
291 {
292 if (subframe == 0 || (subframe == 2 && mode != MODE_6k60)) {
293 if (pitch_index < 116) {
294 *lag_int = (pitch_index + 69) >> 1;
295 *lag_frac = (pitch_index - (*lag_int << 1) + 68) * 2;
296 } else {
297 *lag_int = pitch_index - 24;
298 *lag_frac = 0;
299 }
300 // XXX: same problem as before
301 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
302 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);
303 } else {
304 *lag_int = (pitch_index + 1) >> 1;
305 *lag_frac = (pitch_index - (*lag_int << 1)) * 2;
306 *lag_int += *base_lag_int;
307 }
308 }
309
310 /**
311 * Find the pitch vector by interpolating the past excitation at the
312 * pitch delay, which is obtained in this function.
313 *
314 * @param[in,out] ctx The context
315 * @param[in] amr_subframe Current subframe data
316 * @param[in] subframe Current subframe index (0 to 3)
317 */
decode_pitch_vector(AMRWBContext * ctx,const AMRWBSubFrame * amr_subframe,const int subframe)318 static void decode_pitch_vector(AMRWBContext *ctx,
319 const AMRWBSubFrame *amr_subframe,
320 const int subframe)
321 {
322 int pitch_lag_int, pitch_lag_frac;
323 int i;
324 float *exc = ctx->excitation;
325 enum Mode mode = ctx->fr_cur_mode;
326
327 if (mode <= MODE_8k85) {
328 decode_pitch_lag_low(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
329 &ctx->base_pitch_lag, subframe, mode);
330 } else
331 decode_pitch_lag_high(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
332 &ctx->base_pitch_lag, subframe);
333
334 ctx->pitch_lag_int = pitch_lag_int;
335 pitch_lag_int += pitch_lag_frac > 0;
336
337 /* Calculate the pitch vector by interpolating the past excitation at the
338 pitch lag using a hamming windowed sinc function */
339 ctx->acelpf_ctx.acelp_interpolatef(exc,
340 exc + 1 - pitch_lag_int,
341 ac_inter, 4,
342 pitch_lag_frac + (pitch_lag_frac > 0 ? 0 : 4),
343 LP_ORDER, AMRWB_SFR_SIZE + 1);
344
345 /* Check which pitch signal path should be used
346 * 6k60 and 8k85 modes have the ltp flag set to 0 */
347 if (amr_subframe->ltp) {
348 memcpy(ctx->pitch_vector, exc, AMRWB_SFR_SIZE * sizeof(float));
349 } else {
350 for (i = 0; i < AMRWB_SFR_SIZE; i++)
351 ctx->pitch_vector[i] = 0.18 * exc[i - 1] + 0.64 * exc[i] +
352 0.18 * exc[i + 1];
353 memcpy(exc, ctx->pitch_vector, AMRWB_SFR_SIZE * sizeof(float));
354 }
355 }
356
357 /** Get x bits in the index interval [lsb,lsb+len-1] inclusive */
358 #define BIT_STR(x,lsb,len) av_mod_uintp2((x) >> (lsb), (len))
359
360 /** Get the bit at specified position */
361 #define BIT_POS(x, p) (((x) >> (p)) & 1)
362
363 /**
364 * The next six functions decode_[i]p_track decode exactly i pulses
365 * positions and amplitudes (-1 or 1) in a subframe track using
366 * an encoded pulse indexing (TS 26.190 section 5.8.2).
367 *
368 * The results are given in out[], in which a negative number means
369 * amplitude -1 and vice versa (i.e., ampl(x) = x / abs(x) ).
370 *
371 * @param[out] out Output buffer (writes i elements)
372 * @param[in] code Pulse index (no. of bits varies, see below)
373 * @param[in] m (log2) Number of potential positions
374 * @param[in] off Offset for decoded positions
375 */
decode_1p_track(int * out,int code,int m,int off)376 static inline void decode_1p_track(int *out, int code, int m, int off)
377 {
378 int pos = BIT_STR(code, 0, m) + off; ///code: m+1 bits
379
380 out[0] = BIT_POS(code, m) ? -pos : pos;
381 }
382
decode_2p_track(int * out,int code,int m,int off)383 static inline void decode_2p_track(int *out, int code, int m, int off) ///code: 2m+1 bits
384 {
385 int pos0 = BIT_STR(code, m, m) + off;
386 int pos1 = BIT_STR(code, 0, m) + off;
387
388 out[0] = BIT_POS(code, 2*m) ? -pos0 : pos0;
389 out[1] = BIT_POS(code, 2*m) ? -pos1 : pos1;
390 out[1] = pos0 > pos1 ? -out[1] : out[1];
391 }
392
decode_3p_track(int * out,int code,int m,int off)393 static void decode_3p_track(int *out, int code, int m, int off) ///code: 3m+1 bits
394 {
395 int half_2p = BIT_POS(code, 2*m - 1) << (m - 1);
396
397 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
398 m - 1, off + half_2p);
399 decode_1p_track(out + 2, BIT_STR(code, 2*m, m + 1), m, off);
400 }
401
decode_4p_track(int * out,int code,int m,int off)402 static void decode_4p_track(int *out, int code, int m, int off) ///code: 4m bits
403 {
404 int half_4p, subhalf_2p;
405 int b_offset = 1 << (m - 1);
406
407 switch (BIT_STR(code, 4*m - 2, 2)) { /* case ID (2 bits) */
408 case 0: /* 0 pulses in A, 4 pulses in B or vice versa */
409 half_4p = BIT_POS(code, 4*m - 3) << (m - 1); // which has 4 pulses
410 subhalf_2p = BIT_POS(code, 2*m - 3) << (m - 2);
411
412 decode_2p_track(out, BIT_STR(code, 0, 2*m - 3),
413 m - 2, off + half_4p + subhalf_2p);
414 decode_2p_track(out + 2, BIT_STR(code, 2*m - 2, 2*m - 1),
415 m - 1, off + half_4p);
416 break;
417 case 1: /* 1 pulse in A, 3 pulses in B */
418 decode_1p_track(out, BIT_STR(code, 3*m - 2, m),
419 m - 1, off);
420 decode_3p_track(out + 1, BIT_STR(code, 0, 3*m - 2),
421 m - 1, off + b_offset);
422 break;
423 case 2: /* 2 pulses in each half */
424 decode_2p_track(out, BIT_STR(code, 2*m - 1, 2*m - 1),
425 m - 1, off);
426 decode_2p_track(out + 2, BIT_STR(code, 0, 2*m - 1),
427 m - 1, off + b_offset);
428 break;
429 case 3: /* 3 pulses in A, 1 pulse in B */
430 decode_3p_track(out, BIT_STR(code, m, 3*m - 2),
431 m - 1, off);
432 decode_1p_track(out + 3, BIT_STR(code, 0, m),
433 m - 1, off + b_offset);
434 break;
435 }
436 }
437
decode_5p_track(int * out,int code,int m,int off)438 static void decode_5p_track(int *out, int code, int m, int off) ///code: 5m bits
439 {
440 int half_3p = BIT_POS(code, 5*m - 1) << (m - 1);
441
442 decode_3p_track(out, BIT_STR(code, 2*m + 1, 3*m - 2),
443 m - 1, off + half_3p);
444
445 decode_2p_track(out + 3, BIT_STR(code, 0, 2*m + 1), m, off);
446 }
447
decode_6p_track(int * out,int code,int m,int off)448 static void decode_6p_track(int *out, int code, int m, int off) ///code: 6m-2 bits
449 {
450 int b_offset = 1 << (m - 1);
451 /* which half has more pulses in cases 0 to 2 */
452 int half_more = BIT_POS(code, 6*m - 5) << (m - 1);
453 int half_other = b_offset - half_more;
454
455 switch (BIT_STR(code, 6*m - 4, 2)) { /* case ID (2 bits) */
456 case 0: /* 0 pulses in A, 6 pulses in B or vice versa */
457 decode_1p_track(out, BIT_STR(code, 0, m),
458 m - 1, off + half_more);
459 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
460 m - 1, off + half_more);
461 break;
462 case 1: /* 1 pulse in A, 5 pulses in B or vice versa */
463 decode_1p_track(out, BIT_STR(code, 0, m),
464 m - 1, off + half_other);
465 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
466 m - 1, off + half_more);
467 break;
468 case 2: /* 2 pulses in A, 4 pulses in B or vice versa */
469 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
470 m - 1, off + half_other);
471 decode_4p_track(out + 2, BIT_STR(code, 2*m - 1, 4*m - 4),
472 m - 1, off + half_more);
473 break;
474 case 3: /* 3 pulses in A, 3 pulses in B */
475 decode_3p_track(out, BIT_STR(code, 3*m - 2, 3*m - 2),
476 m - 1, off);
477 decode_3p_track(out + 3, BIT_STR(code, 0, 3*m - 2),
478 m - 1, off + b_offset);
479 break;
480 }
481 }
482
483 /**
484 * Decode the algebraic codebook index to pulse positions and signs,
485 * then construct the algebraic codebook vector.
486 *
487 * @param[out] fixed_vector Buffer for the fixed codebook excitation
488 * @param[in] pulse_hi MSBs part of the pulse index array (higher modes only)
489 * @param[in] pulse_lo LSBs part of the pulse index array
490 * @param[in] mode Mode of the current frame
491 */
decode_fixed_vector(float * fixed_vector,const uint16_t * pulse_hi,const uint16_t * pulse_lo,const enum Mode mode)492 static void decode_fixed_vector(float *fixed_vector, const uint16_t *pulse_hi,
493 const uint16_t *pulse_lo, const enum Mode mode)
494 {
495 /* sig_pos stores for each track the decoded pulse position indexes
496 * (1-based) multiplied by its corresponding amplitude (+1 or -1) */
497 int sig_pos[4][6];
498 int spacing = (mode == MODE_6k60) ? 2 : 4;
499 int i, j;
500
501 switch (mode) {
502 case MODE_6k60:
503 for (i = 0; i < 2; i++)
504 decode_1p_track(sig_pos[i], pulse_lo[i], 5, 1);
505 break;
506 case MODE_8k85:
507 for (i = 0; i < 4; i++)
508 decode_1p_track(sig_pos[i], pulse_lo[i], 4, 1);
509 break;
510 case MODE_12k65:
511 for (i = 0; i < 4; i++)
512 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
513 break;
514 case MODE_14k25:
515 for (i = 0; i < 2; i++)
516 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
517 for (i = 2; i < 4; i++)
518 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
519 break;
520 case MODE_15k85:
521 for (i = 0; i < 4; i++)
522 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
523 break;
524 case MODE_18k25:
525 for (i = 0; i < 4; i++)
526 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
527 ((int) pulse_hi[i] << 14), 4, 1);
528 break;
529 case MODE_19k85:
530 for (i = 0; i < 2; i++)
531 decode_5p_track(sig_pos[i], (int) pulse_lo[i] +
532 ((int) pulse_hi[i] << 10), 4, 1);
533 for (i = 2; i < 4; i++)
534 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
535 ((int) pulse_hi[i] << 14), 4, 1);
536 break;
537 case MODE_23k05:
538 case MODE_23k85:
539 for (i = 0; i < 4; i++)
540 decode_6p_track(sig_pos[i], (int) pulse_lo[i] +
541 ((int) pulse_hi[i] << 11), 4, 1);
542 break;
543 }
544
545 memset(fixed_vector, 0, sizeof(float) * AMRWB_SFR_SIZE);
546
547 for (i = 0; i < 4; i++)
548 for (j = 0; j < pulses_nb_per_mode_tr[mode][i]; j++) {
549 int pos = (FFABS(sig_pos[i][j]) - 1) * spacing + i;
550
551 fixed_vector[pos] += sig_pos[i][j] < 0 ? -1.0 : 1.0;
552 }
553 }
554
555 /**
556 * Decode pitch gain and fixed gain correction factor.
557 *
558 * @param[in] vq_gain Vector-quantized index for gains
559 * @param[in] mode Mode of the current frame
560 * @param[out] fixed_gain_factor Decoded fixed gain correction factor
561 * @param[out] pitch_gain Decoded pitch gain
562 */
decode_gains(const uint8_t vq_gain,const enum Mode mode,float * fixed_gain_factor,float * pitch_gain)563 static void decode_gains(const uint8_t vq_gain, const enum Mode mode,
564 float *fixed_gain_factor, float *pitch_gain)
565 {
566 const int16_t *gains = (mode <= MODE_8k85 ? qua_gain_6b[vq_gain] :
567 qua_gain_7b[vq_gain]);
568
569 *pitch_gain = gains[0] * (1.0f / (1 << 14));
570 *fixed_gain_factor = gains[1] * (1.0f / (1 << 11));
571 }
572
573 /**
574 * Apply pitch sharpening filters to the fixed codebook vector.
575 *
576 * @param[in] ctx The context
577 * @param[in,out] fixed_vector Fixed codebook excitation
578 */
579 // XXX: Spec states this procedure should be applied when the pitch
580 // lag is less than 64, but this checking seems absent in reference and AMR-NB
pitch_sharpening(AMRWBContext * ctx,float * fixed_vector)581 static void pitch_sharpening(AMRWBContext *ctx, float *fixed_vector)
582 {
583 int i;
584
585 /* Tilt part */
586 for (i = AMRWB_SFR_SIZE - 1; i != 0; i--)
587 fixed_vector[i] -= fixed_vector[i - 1] * ctx->tilt_coef;
588
589 /* Periodicity enhancement part */
590 for (i = ctx->pitch_lag_int; i < AMRWB_SFR_SIZE; i++)
591 fixed_vector[i] += fixed_vector[i - ctx->pitch_lag_int] * 0.85;
592 }
593
594 /**
595 * Calculate the voicing factor (-1.0 = unvoiced to 1.0 = voiced).
596 *
597 * @param[in] p_vector, f_vector Pitch and fixed excitation vectors
598 * @param[in] p_gain, f_gain Pitch and fixed gains
599 * @param[in] ctx The context
600 */
601 // XXX: There is something wrong with the precision here! The magnitudes
602 // of the energies are not correct. Please check the reference code carefully
voice_factor(float * p_vector,float p_gain,float * f_vector,float f_gain,CELPMContext * ctx)603 static float voice_factor(float *p_vector, float p_gain,
604 float *f_vector, float f_gain,
605 CELPMContext *ctx)
606 {
607 double p_ener = (double) ctx->dot_productf(p_vector, p_vector,
608 AMRWB_SFR_SIZE) *
609 p_gain * p_gain;
610 double f_ener = (double) ctx->dot_productf(f_vector, f_vector,
611 AMRWB_SFR_SIZE) *
612 f_gain * f_gain;
613
614 return (p_ener - f_ener) / (p_ener + f_ener + 0.01);
615 }
616
617 /**
618 * Reduce fixed vector sparseness by smoothing with one of three IR filters,
619 * also known as "adaptive phase dispersion".
620 *
621 * @param[in] ctx The context
622 * @param[in,out] fixed_vector Unfiltered fixed vector
623 * @param[out] buf Space for modified vector if necessary
624 *
625 * @return The potentially overwritten filtered fixed vector address
626 */
anti_sparseness(AMRWBContext * ctx,float * fixed_vector,float * buf)627 static float *anti_sparseness(AMRWBContext *ctx,
628 float *fixed_vector, float *buf)
629 {
630 int ir_filter_nr;
631
632 if (ctx->fr_cur_mode > MODE_8k85) // no filtering in higher modes
633 return fixed_vector;
634
635 if (ctx->pitch_gain[0] < 0.6) {
636 ir_filter_nr = 0; // strong filtering
637 } else if (ctx->pitch_gain[0] < 0.9) {
638 ir_filter_nr = 1; // medium filtering
639 } else
640 ir_filter_nr = 2; // no filtering
641
642 /* detect 'onset' */
643 if (ctx->fixed_gain[0] > 3.0 * ctx->fixed_gain[1]) {
644 if (ir_filter_nr < 2)
645 ir_filter_nr++;
646 } else {
647 int i, count = 0;
648
649 for (i = 0; i < 6; i++)
650 if (ctx->pitch_gain[i] < 0.6)
651 count++;
652
653 if (count > 2)
654 ir_filter_nr = 0;
655
656 if (ir_filter_nr > ctx->prev_ir_filter_nr + 1)
657 ir_filter_nr--;
658 }
659
660 /* update ir filter strength history */
661 ctx->prev_ir_filter_nr = ir_filter_nr;
662
663 ir_filter_nr += (ctx->fr_cur_mode == MODE_8k85);
664
665 if (ir_filter_nr < 2) {
666 int i;
667 const float *coef = ir_filters_lookup[ir_filter_nr];
668
669 /* Circular convolution code in the reference
670 * decoder was modified to avoid using one
671 * extra array. The filtered vector is given by:
672 *
673 * c2(n) = sum(i,0,len-1){ c(i) * coef( (n - i + len) % len ) }
674 */
675
676 memset(buf, 0, sizeof(float) * AMRWB_SFR_SIZE);
677 for (i = 0; i < AMRWB_SFR_SIZE; i++)
678 if (fixed_vector[i])
679 ff_celp_circ_addf(buf, buf, coef, i, fixed_vector[i],
680 AMRWB_SFR_SIZE);
681 fixed_vector = buf;
682 }
683
684 return fixed_vector;
685 }
686
687 /**
688 * Calculate a stability factor {teta} based on distance between
689 * current and past isf. A value of 1 shows maximum signal stability.
690 */
stability_factor(const float * isf,const float * isf_past)691 static float stability_factor(const float *isf, const float *isf_past)
692 {
693 int i;
694 float acc = 0.0;
695
696 for (i = 0; i < LP_ORDER - 1; i++)
697 acc += (isf[i] - isf_past[i]) * (isf[i] - isf_past[i]);
698
699 // XXX: This part is not so clear from the reference code
700 // the result is more accurate changing the "/ 256" to "* 512"
701 return FFMAX(0.0, 1.25 - acc * 0.8 * 512);
702 }
703
704 /**
705 * Apply a non-linear fixed gain smoothing in order to reduce
706 * fluctuation in the energy of excitation.
707 *
708 * @param[in] fixed_gain Unsmoothed fixed gain
709 * @param[in,out] prev_tr_gain Previous threshold gain (updated)
710 * @param[in] voice_fac Frame voicing factor
711 * @param[in] stab_fac Frame stability factor
712 *
713 * @return The smoothed gain
714 */
noise_enhancer(float fixed_gain,float * prev_tr_gain,float voice_fac,float stab_fac)715 static float noise_enhancer(float fixed_gain, float *prev_tr_gain,
716 float voice_fac, float stab_fac)
717 {
718 float sm_fac = 0.5 * (1 - voice_fac) * stab_fac;
719 float g0;
720
721 // XXX: the following fixed-point constants used to in(de)crement
722 // gain by 1.5dB were taken from the reference code, maybe it could
723 // be simpler
724 if (fixed_gain < *prev_tr_gain) {
725 g0 = FFMIN(*prev_tr_gain, fixed_gain + fixed_gain *
726 (6226 * (1.0f / (1 << 15)))); // +1.5 dB
727 } else
728 g0 = FFMAX(*prev_tr_gain, fixed_gain *
729 (27536 * (1.0f / (1 << 15)))); // -1.5 dB
730
731 *prev_tr_gain = g0; // update next frame threshold
732
733 return sm_fac * g0 + (1 - sm_fac) * fixed_gain;
734 }
735
736 /**
737 * Filter the fixed_vector to emphasize the higher frequencies.
738 *
739 * @param[in,out] fixed_vector Fixed codebook vector
740 * @param[in] voice_fac Frame voicing factor
741 */
pitch_enhancer(float * fixed_vector,float voice_fac)742 static void pitch_enhancer(float *fixed_vector, float voice_fac)
743 {
744 int i;
745 float cpe = 0.125 * (1 + voice_fac);
746 float last = fixed_vector[0]; // holds c(i - 1)
747
748 fixed_vector[0] -= cpe * fixed_vector[1];
749
750 for (i = 1; i < AMRWB_SFR_SIZE - 1; i++) {
751 float cur = fixed_vector[i];
752
753 fixed_vector[i] -= cpe * (last + fixed_vector[i + 1]);
754 last = cur;
755 }
756
757 fixed_vector[AMRWB_SFR_SIZE - 1] -= cpe * last;
758 }
759
760 /**
761 * Conduct 16th order linear predictive coding synthesis from excitation.
762 *
763 * @param[in] ctx Pointer to the AMRWBContext
764 * @param[in] lpc Pointer to the LPC coefficients
765 * @param[out] excitation Buffer for synthesis final excitation
766 * @param[in] fixed_gain Fixed codebook gain for synthesis
767 * @param[in] fixed_vector Algebraic codebook vector
768 * @param[in,out] samples Pointer to the output samples and memory
769 */
synthesis(AMRWBContext * ctx,float * lpc,float * excitation,float fixed_gain,const float * fixed_vector,float * samples)770 static void synthesis(AMRWBContext *ctx, float *lpc, float *excitation,
771 float fixed_gain, const float *fixed_vector,
772 float *samples)
773 {
774 ctx->acelpv_ctx.weighted_vector_sumf(excitation, ctx->pitch_vector, fixed_vector,
775 ctx->pitch_gain[0], fixed_gain, AMRWB_SFR_SIZE);
776
777 /* emphasize pitch vector contribution in low bitrate modes */
778 if (ctx->pitch_gain[0] > 0.5 && ctx->fr_cur_mode <= MODE_8k85) {
779 int i;
780 float energy = ctx->celpm_ctx.dot_productf(excitation, excitation,
781 AMRWB_SFR_SIZE);
782
783 // XXX: Weird part in both ref code and spec. A unknown parameter
784 // {beta} seems to be identical to the current pitch gain
785 float pitch_factor = 0.25 * ctx->pitch_gain[0] * ctx->pitch_gain[0];
786
787 for (i = 0; i < AMRWB_SFR_SIZE; i++)
788 excitation[i] += pitch_factor * ctx->pitch_vector[i];
789
790 ff_scale_vector_to_given_sum_of_squares(excitation, excitation,
791 energy, AMRWB_SFR_SIZE);
792 }
793
794 ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation,
795 AMRWB_SFR_SIZE, LP_ORDER);
796 }
797
798 /**
799 * Apply to synthesis a de-emphasis filter of the form:
800 * H(z) = 1 / (1 - m * z^-1)
801 *
802 * @param[out] out Output buffer
803 * @param[in] in Input samples array with in[-1]
804 * @param[in] m Filter coefficient
805 * @param[in,out] mem State from last filtering
806 */
de_emphasis(float * out,float * in,float m,float mem[1])807 static void de_emphasis(float *out, float *in, float m, float mem[1])
808 {
809 int i;
810
811 out[0] = in[0] + m * mem[0];
812
813 for (i = 1; i < AMRWB_SFR_SIZE; i++)
814 out[i] = in[i] + out[i - 1] * m;
815
816 mem[0] = out[AMRWB_SFR_SIZE - 1];
817 }
818
819 /**
820 * Upsample a signal by 5/4 ratio (from 12.8kHz to 16kHz) using
821 * a FIR interpolation filter. Uses past data from before *in address.
822 *
823 * @param[out] out Buffer for interpolated signal
824 * @param[in] in Current signal data (length 0.8*o_size)
825 * @param[in] o_size Output signal length
826 * @param[in] ctx The context
827 */
upsample_5_4(float * out,const float * in,int o_size,CELPMContext * ctx)828 static void upsample_5_4(float *out, const float *in, int o_size, CELPMContext *ctx)
829 {
830 const float *in0 = in - UPS_FIR_SIZE + 1;
831 int i, j, k;
832 int int_part = 0, frac_part;
833
834 i = 0;
835 for (j = 0; j < o_size / 5; j++) {
836 out[i] = in[int_part];
837 frac_part = 4;
838 i++;
839
840 for (k = 1; k < 5; k++) {
841 out[i] = ctx->dot_productf(in0 + int_part,
842 upsample_fir[4 - frac_part],
843 UPS_MEM_SIZE);
844 int_part++;
845 frac_part--;
846 i++;
847 }
848 }
849 }
850
851 /**
852 * Calculate the high-band gain based on encoded index (23k85 mode) or
853 * on the low-band speech signal and the Voice Activity Detection flag.
854 *
855 * @param[in] ctx The context
856 * @param[in] synth LB speech synthesis at 12.8k
857 * @param[in] hb_idx Gain index for mode 23k85 only
858 * @param[in] vad VAD flag for the frame
859 */
find_hb_gain(AMRWBContext * ctx,const float * synth,uint16_t hb_idx,uint8_t vad)860 static float find_hb_gain(AMRWBContext *ctx, const float *synth,
861 uint16_t hb_idx, uint8_t vad)
862 {
863 int wsp = (vad > 0);
864 float tilt;
865 float tmp;
866
867 if (ctx->fr_cur_mode == MODE_23k85)
868 return qua_hb_gain[hb_idx] * (1.0f / (1 << 14));
869
870 tmp = ctx->celpm_ctx.dot_productf(synth, synth + 1, AMRWB_SFR_SIZE - 1);
871
872 if (tmp > 0) {
873 tilt = tmp / ctx->celpm_ctx.dot_productf(synth, synth, AMRWB_SFR_SIZE);
874 } else
875 tilt = 0;
876
877 /* return gain bounded by [0.1, 1.0] */
878 return av_clipf((1.0 - tilt) * (1.25 - 0.25 * wsp), 0.1, 1.0);
879 }
880
881 /**
882 * Generate the high-band excitation with the same energy from the lower
883 * one and scaled by the given gain.
884 *
885 * @param[in] ctx The context
886 * @param[out] hb_exc Buffer for the excitation
887 * @param[in] synth_exc Low-band excitation used for synthesis
888 * @param[in] hb_gain Wanted excitation gain
889 */
scaled_hb_excitation(AMRWBContext * ctx,float * hb_exc,const float * synth_exc,float hb_gain)890 static void scaled_hb_excitation(AMRWBContext *ctx, float *hb_exc,
891 const float *synth_exc, float hb_gain)
892 {
893 int i;
894 float energy = ctx->celpm_ctx.dot_productf(synth_exc, synth_exc,
895 AMRWB_SFR_SIZE);
896
897 /* Generate a white-noise excitation */
898 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
899 hb_exc[i] = 32768.0 - (uint16_t) av_lfg_get(&ctx->prng);
900
901 ff_scale_vector_to_given_sum_of_squares(hb_exc, hb_exc,
902 energy * hb_gain * hb_gain,
903 AMRWB_SFR_SIZE_16k);
904 }
905
906 /**
907 * Calculate the auto-correlation for the ISF difference vector.
908 */
auto_correlation(float * diff_isf,float mean,int lag)909 static float auto_correlation(float *diff_isf, float mean, int lag)
910 {
911 int i;
912 float sum = 0.0;
913
914 for (i = 7; i < LP_ORDER - 2; i++) {
915 float prod = (diff_isf[i] - mean) * (diff_isf[i - lag] - mean);
916 sum += prod * prod;
917 }
918 return sum;
919 }
920
921 /**
922 * Extrapolate a ISF vector to the 16kHz range (20th order LP)
923 * used at mode 6k60 LP filter for the high frequency band.
924 *
925 * @param[out] isf Buffer for extrapolated isf; contains LP_ORDER
926 * values on input
927 */
extrapolate_isf(float isf[LP_ORDER_16k])928 static void extrapolate_isf(float isf[LP_ORDER_16k])
929 {
930 float diff_isf[LP_ORDER - 2], diff_mean;
931 float corr_lag[3];
932 float est, scale;
933 int i, j, i_max_corr;
934
935 isf[LP_ORDER_16k - 1] = isf[LP_ORDER - 1];
936
937 /* Calculate the difference vector */
938 for (i = 0; i < LP_ORDER - 2; i++)
939 diff_isf[i] = isf[i + 1] - isf[i];
940
941 diff_mean = 0.0;
942 for (i = 2; i < LP_ORDER - 2; i++)
943 diff_mean += diff_isf[i] * (1.0f / (LP_ORDER - 4));
944
945 /* Find which is the maximum autocorrelation */
946 i_max_corr = 0;
947 for (i = 0; i < 3; i++) {
948 corr_lag[i] = auto_correlation(diff_isf, diff_mean, i + 2);
949
950 if (corr_lag[i] > corr_lag[i_max_corr])
951 i_max_corr = i;
952 }
953 i_max_corr++;
954
955 for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
956 isf[i] = isf[i - 1] + isf[i - 1 - i_max_corr]
957 - isf[i - 2 - i_max_corr];
958
959 /* Calculate an estimate for ISF(18) and scale ISF based on the error */
960 est = 7965 + (isf[2] - isf[3] - isf[4]) / 6.0;
961 scale = 0.5 * (FFMIN(est, 7600) - isf[LP_ORDER - 2]) /
962 (isf[LP_ORDER_16k - 2] - isf[LP_ORDER - 2]);
963
964 for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++)
965 diff_isf[j] = scale * (isf[i] - isf[i - 1]);
966
967 /* Stability insurance */
968 for (i = 1; i < LP_ORDER_16k - LP_ORDER; i++)
969 if (diff_isf[i] + diff_isf[i - 1] < 5.0) {
970 if (diff_isf[i] > diff_isf[i - 1]) {
971 diff_isf[i - 1] = 5.0 - diff_isf[i];
972 } else
973 diff_isf[i] = 5.0 - diff_isf[i - 1];
974 }
975
976 for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++)
977 isf[i] = isf[i - 1] + diff_isf[j] * (1.0f / (1 << 15));
978
979 /* Scale the ISF vector for 16000 Hz */
980 for (i = 0; i < LP_ORDER_16k - 1; i++)
981 isf[i] *= 0.8;
982 }
983
984 /**
985 * Spectral expand the LP coefficients using the equation:
986 * y[i] = x[i] * (gamma ** i)
987 *
988 * @param[out] out Output buffer (may use input array)
989 * @param[in] lpc LP coefficients array
990 * @param[in] gamma Weighting factor
991 * @param[in] size LP array size
992 */
lpc_weighting(float * out,const float * lpc,float gamma,int size)993 static void lpc_weighting(float *out, const float *lpc, float gamma, int size)
994 {
995 int i;
996 float fac = gamma;
997
998 for (i = 0; i < size; i++) {
999 out[i] = lpc[i] * fac;
1000 fac *= gamma;
1001 }
1002 }
1003
1004 /**
1005 * Conduct 20th order linear predictive coding synthesis for the high
1006 * frequency band excitation at 16kHz.
1007 *
1008 * @param[in] ctx The context
1009 * @param[in] subframe Current subframe index (0 to 3)
1010 * @param[in,out] samples Pointer to the output speech samples
1011 * @param[in] exc Generated white-noise scaled excitation
1012 * @param[in] isf Current frame isf vector
1013 * @param[in] isf_past Past frame final isf vector
1014 */
hb_synthesis(AMRWBContext * ctx,int subframe,float * samples,const float * exc,const float * isf,const float * isf_past)1015 static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples,
1016 const float *exc, const float *isf, const float *isf_past)
1017 {
1018 float hb_lpc[LP_ORDER_16k];
1019 enum Mode mode = ctx->fr_cur_mode;
1020
1021 if (mode == MODE_6k60) {
1022 float e_isf[LP_ORDER_16k]; // ISF vector for extrapolation
1023 double e_isp[LP_ORDER_16k];
1024
1025 ctx->acelpv_ctx.weighted_vector_sumf(e_isf, isf_past, isf, isfp_inter[subframe],
1026 1.0 - isfp_inter[subframe], LP_ORDER);
1027
1028 extrapolate_isf(e_isf);
1029
1030 e_isf[LP_ORDER_16k - 1] *= 2.0;
1031 ff_acelp_lsf2lspd(e_isp, e_isf, LP_ORDER_16k);
1032 ff_amrwb_lsp2lpc(e_isp, hb_lpc, LP_ORDER_16k);
1033
1034 lpc_weighting(hb_lpc, hb_lpc, 0.9, LP_ORDER_16k);
1035 } else {
1036 lpc_weighting(hb_lpc, ctx->lp_coef[subframe], 0.6, LP_ORDER);
1037 }
1038
1039 ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, hb_lpc, exc, AMRWB_SFR_SIZE_16k,
1040 (mode == MODE_6k60) ? LP_ORDER_16k : LP_ORDER);
1041 }
1042
1043 /**
1044 * Apply a 15th order filter to high-band samples.
1045 * The filter characteristic depends on the given coefficients.
1046 *
1047 * @param[out] out Buffer for filtered output
1048 * @param[in] fir_coef Filter coefficients
1049 * @param[in,out] mem State from last filtering (updated)
1050 * @param[in] in Input speech data (high-band)
1051 *
1052 * @remark It is safe to pass the same array in in and out parameters
1053 */
1054
1055 #ifndef hb_fir_filter
hb_fir_filter(float * out,const float fir_coef[HB_FIR_SIZE+1],float mem[HB_FIR_SIZE],const float * in)1056 static void hb_fir_filter(float *out, const float fir_coef[HB_FIR_SIZE + 1],
1057 float mem[HB_FIR_SIZE], const float *in)
1058 {
1059 int i, j;
1060 float data[AMRWB_SFR_SIZE_16k + HB_FIR_SIZE]; // past and current samples
1061
1062 memcpy(data, mem, HB_FIR_SIZE * sizeof(float));
1063 memcpy(data + HB_FIR_SIZE, in, AMRWB_SFR_SIZE_16k * sizeof(float));
1064
1065 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) {
1066 out[i] = 0.0;
1067 for (j = 0; j <= HB_FIR_SIZE; j++)
1068 out[i] += data[i + j] * fir_coef[j];
1069 }
1070
1071 memcpy(mem, data + AMRWB_SFR_SIZE_16k, HB_FIR_SIZE * sizeof(float));
1072 }
1073 #endif /* hb_fir_filter */
1074
1075 /**
1076 * Update context state before the next subframe.
1077 */
update_sub_state(AMRWBContext * ctx)1078 static void update_sub_state(AMRWBContext *ctx)
1079 {
1080 memmove(&ctx->excitation_buf[0], &ctx->excitation_buf[AMRWB_SFR_SIZE],
1081 (AMRWB_P_DELAY_MAX + LP_ORDER + 1) * sizeof(float));
1082
1083 memmove(&ctx->pitch_gain[1], &ctx->pitch_gain[0], 5 * sizeof(float));
1084 memmove(&ctx->fixed_gain[1], &ctx->fixed_gain[0], 1 * sizeof(float));
1085
1086 memmove(&ctx->samples_az[0], &ctx->samples_az[AMRWB_SFR_SIZE],
1087 LP_ORDER * sizeof(float));
1088 memmove(&ctx->samples_up[0], &ctx->samples_up[AMRWB_SFR_SIZE],
1089 UPS_MEM_SIZE * sizeof(float));
1090 memmove(&ctx->samples_hb[0], &ctx->samples_hb[AMRWB_SFR_SIZE_16k],
1091 LP_ORDER_16k * sizeof(float));
1092 }
1093
amrwb_decode_frame(AVCodecContext * avctx,void * data,int * got_frame_ptr,AVPacket * avpkt)1094 static int amrwb_decode_frame(AVCodecContext *avctx, void *data,
1095 int *got_frame_ptr, AVPacket *avpkt)
1096 {
1097 AMRWBContext *ctx = avctx->priv_data;
1098 AVFrame *frame = data;
1099 AMRWBFrame *cf = &ctx->frame;
1100 const uint8_t *buf = avpkt->data;
1101 int buf_size = avpkt->size;
1102 int expected_fr_size, header_size;
1103 float *buf_out;
1104 float spare_vector[AMRWB_SFR_SIZE]; // extra stack space to hold result from anti-sparseness processing
1105 float fixed_gain_factor; // fixed gain correction factor (gamma)
1106 float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
1107 float synth_fixed_gain; // the fixed gain that synthesis should use
1108 float voice_fac, stab_fac; // parameters used for gain smoothing
1109 float synth_exc[AMRWB_SFR_SIZE]; // post-processed excitation for synthesis
1110 float hb_exc[AMRWB_SFR_SIZE_16k]; // excitation for the high frequency band
1111 float hb_samples[AMRWB_SFR_SIZE_16k]; // filtered high-band samples from synthesis
1112 float hb_gain;
1113 int sub, i, ret;
1114
1115 /* get output buffer */
1116 frame->nb_samples = 4 * AMRWB_SFR_SIZE_16k;
1117 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
1118 return ret;
1119 buf_out = (float *)frame->data[0];
1120
1121 header_size = decode_mime_header(ctx, buf);
1122 expected_fr_size = ((cf_sizes_wb[ctx->fr_cur_mode] + 7) >> 3) + 1;
1123
1124 if (!ctx->fr_quality)
1125 av_log(avctx, AV_LOG_ERROR, "Encountered a bad or corrupted frame\n");
1126
1127 if (ctx->fr_cur_mode == NO_DATA || !ctx->fr_quality) {
1128 /* The specification suggests a "random signal" and
1129 "a muting technique" to "gradually decrease the output level". */
1130 av_samples_set_silence(&frame->data[0], 0, frame->nb_samples, 1, AV_SAMPLE_FMT_FLT);
1131 *got_frame_ptr = 1;
1132 return expected_fr_size;
1133 }
1134 if (ctx->fr_cur_mode > MODE_SID) {
1135 av_log(avctx, AV_LOG_ERROR,
1136 "Invalid mode %d\n", ctx->fr_cur_mode);
1137 return AVERROR_INVALIDDATA;
1138 }
1139
1140 if (buf_size < expected_fr_size) {
1141 av_log(avctx, AV_LOG_ERROR,
1142 "Frame too small (%d bytes). Truncated file?\n", buf_size);
1143 *got_frame_ptr = 0;
1144 return AVERROR_INVALIDDATA;
1145 }
1146
1147 if (ctx->fr_cur_mode == MODE_SID) { /* Comfort noise frame */
1148 avpriv_request_sample(avctx, "SID mode");
1149 return AVERROR_PATCHWELCOME;
1150 }
1151
1152 ff_amr_bit_reorder((uint16_t *) &ctx->frame, sizeof(AMRWBFrame),
1153 buf + header_size, amr_bit_orderings_by_mode[ctx->fr_cur_mode]);
1154
1155 /* Decode the quantized ISF vector */
1156 if (ctx->fr_cur_mode == MODE_6k60) {
1157 decode_isf_indices_36b(cf->isp_id, ctx->isf_cur);
1158 } else {
1159 decode_isf_indices_46b(cf->isp_id, ctx->isf_cur);
1160 }
1161
1162 isf_add_mean_and_past(ctx->isf_cur, ctx->isf_q_past);
1163 ff_set_min_dist_lsf(ctx->isf_cur, MIN_ISF_SPACING, LP_ORDER - 1);
1164
1165 stab_fac = stability_factor(ctx->isf_cur, ctx->isf_past_final);
1166
1167 ctx->isf_cur[LP_ORDER - 1] *= 2.0;
1168 ff_acelp_lsf2lspd(ctx->isp[3], ctx->isf_cur, LP_ORDER);
1169
1170 /* Generate a ISP vector for each subframe */
1171 if (ctx->first_frame) {
1172 ctx->first_frame = 0;
1173 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(double));
1174 }
1175 interpolate_isp(ctx->isp, ctx->isp_sub4_past);
1176
1177 for (sub = 0; sub < 4; sub++)
1178 ff_amrwb_lsp2lpc(ctx->isp[sub], ctx->lp_coef[sub], LP_ORDER);
1179
1180 for (sub = 0; sub < 4; sub++) {
1181 const AMRWBSubFrame *cur_subframe = &cf->subframe[sub];
1182 float *sub_buf = buf_out + sub * AMRWB_SFR_SIZE_16k;
1183
1184 /* Decode adaptive codebook (pitch vector) */
1185 decode_pitch_vector(ctx, cur_subframe, sub);
1186 /* Decode innovative codebook (fixed vector) */
1187 decode_fixed_vector(ctx->fixed_vector, cur_subframe->pul_ih,
1188 cur_subframe->pul_il, ctx->fr_cur_mode);
1189
1190 pitch_sharpening(ctx, ctx->fixed_vector);
1191
1192 decode_gains(cur_subframe->vq_gain, ctx->fr_cur_mode,
1193 &fixed_gain_factor, &ctx->pitch_gain[0]);
1194
1195 ctx->fixed_gain[0] =
1196 ff_amr_set_fixed_gain(fixed_gain_factor,
1197 ctx->celpm_ctx.dot_productf(ctx->fixed_vector,
1198 ctx->fixed_vector,
1199 AMRWB_SFR_SIZE) /
1200 AMRWB_SFR_SIZE,
1201 ctx->prediction_error,
1202 ENERGY_MEAN, energy_pred_fac);
1203
1204 /* Calculate voice factor and store tilt for next subframe */
1205 voice_fac = voice_factor(ctx->pitch_vector, ctx->pitch_gain[0],
1206 ctx->fixed_vector, ctx->fixed_gain[0],
1207 &ctx->celpm_ctx);
1208 ctx->tilt_coef = voice_fac * 0.25 + 0.25;
1209
1210 /* Construct current excitation */
1211 for (i = 0; i < AMRWB_SFR_SIZE; i++) {
1212 ctx->excitation[i] *= ctx->pitch_gain[0];
1213 ctx->excitation[i] += ctx->fixed_gain[0] * ctx->fixed_vector[i];
1214 ctx->excitation[i] = truncf(ctx->excitation[i]);
1215 }
1216
1217 /* Post-processing of excitation elements */
1218 synth_fixed_gain = noise_enhancer(ctx->fixed_gain[0], &ctx->prev_tr_gain,
1219 voice_fac, stab_fac);
1220
1221 synth_fixed_vector = anti_sparseness(ctx, ctx->fixed_vector,
1222 spare_vector);
1223
1224 pitch_enhancer(synth_fixed_vector, voice_fac);
1225
1226 synthesis(ctx, ctx->lp_coef[sub], synth_exc, synth_fixed_gain,
1227 synth_fixed_vector, &ctx->samples_az[LP_ORDER]);
1228
1229 /* Synthesis speech post-processing */
1230 de_emphasis(&ctx->samples_up[UPS_MEM_SIZE],
1231 &ctx->samples_az[LP_ORDER], PREEMPH_FAC, ctx->demph_mem);
1232
1233 ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(&ctx->samples_up[UPS_MEM_SIZE],
1234 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_31_poles,
1235 hpf_31_gain, ctx->hpf_31_mem, AMRWB_SFR_SIZE);
1236
1237 upsample_5_4(sub_buf, &ctx->samples_up[UPS_FIR_SIZE],
1238 AMRWB_SFR_SIZE_16k, &ctx->celpm_ctx);
1239
1240 /* High frequency band (6.4 - 7.0 kHz) generation part */
1241 ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(hb_samples,
1242 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_400_poles,
1243 hpf_400_gain, ctx->hpf_400_mem, AMRWB_SFR_SIZE);
1244
1245 hb_gain = find_hb_gain(ctx, hb_samples,
1246 cur_subframe->hb_gain, cf->vad);
1247
1248 scaled_hb_excitation(ctx, hb_exc, synth_exc, hb_gain);
1249
1250 hb_synthesis(ctx, sub, &ctx->samples_hb[LP_ORDER_16k],
1251 hb_exc, ctx->isf_cur, ctx->isf_past_final);
1252
1253 /* High-band post-processing filters */
1254 hb_fir_filter(hb_samples, bpf_6_7_coef, ctx->bpf_6_7_mem,
1255 &ctx->samples_hb[LP_ORDER_16k]);
1256
1257 if (ctx->fr_cur_mode == MODE_23k85)
1258 hb_fir_filter(hb_samples, lpf_7_coef, ctx->lpf_7_mem,
1259 hb_samples);
1260
1261 /* Add the low and high frequency bands */
1262 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
1263 sub_buf[i] = (sub_buf[i] + hb_samples[i]) * (1.0f / (1 << 15));
1264
1265 /* Update buffers and history */
1266 update_sub_state(ctx);
1267 }
1268
1269 /* update state for next frame */
1270 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(ctx->isp[3][0]));
1271 memcpy(ctx->isf_past_final, ctx->isf_cur, LP_ORDER * sizeof(float));
1272
1273 *got_frame_ptr = 1;
1274
1275 return expected_fr_size;
1276 }
1277
1278 AVCodec ff_amrwb_decoder = {
1279 .name = "amrwb",
1280 .long_name = NULL_IF_CONFIG_SMALL("AMR-WB (Adaptive Multi-Rate WideBand)"),
1281 .type = AVMEDIA_TYPE_AUDIO,
1282 .id = AV_CODEC_ID_AMR_WB,
1283 .priv_data_size = sizeof(AMRWBContext),
1284 .init = amrwb_decode_init,
1285 .decode = amrwb_decode_frame,
1286 .capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF,
1287 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT,
1288 AV_SAMPLE_FMT_NONE },
1289 };
1290