FFmpeg coverage


Directory: ../../../ffmpeg/
File: src/libavcodec/amrwbdec.c
Date: 2022-01-28 07:56:06
Exec Total Coverage
Lines: 498 520 95.8%
Branches: 230 247 93.1%

Line Branch Exec Source
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
96 typedef struct AMRWBChannelsContext {
97 AMRWBContext ch[2];
98 } AMRWBChannelsContext;
99
100 22 static av_cold int amrwb_decode_init(AVCodecContext *avctx)
101 {
102 22 AMRWBChannelsContext *s = avctx->priv_data;
103 int i;
104
105
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 22 times.
22 if (avctx->channels > 2) {
106 avpriv_report_missing_feature(avctx, ">2 channel AMR");
107 return AVERROR_PATCHWELCOME;
108 }
109
110
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 22 times.
22 if (!avctx->channels) {
111 avctx->channels = 1;
112 avctx->channel_layout = AV_CH_LAYOUT_MONO;
113 }
114
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 22 times.
22 if (!avctx->sample_rate)
115 avctx->sample_rate = 16000;
116 22 avctx->sample_fmt = AV_SAMPLE_FMT_FLTP;
117
118
2/2
✓ Branch 0 taken 22 times.
✓ Branch 1 taken 22 times.
44 for (int ch = 0; ch < avctx->channels; ch++) {
119 22 AMRWBContext *ctx = &s->ch[ch];
120
121 22 av_lfg_init(&ctx->prng, 1);
122
123 22 ctx->excitation = &ctx->excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 1];
124 22 ctx->first_frame = 1;
125
126
2/2
✓ Branch 0 taken 352 times.
✓ Branch 1 taken 22 times.
374 for (i = 0; i < LP_ORDER; i++)
127 352 ctx->isf_past_final[i] = isf_init[i] * (1.0f / (1 << 15));
128
129
2/2
✓ Branch 0 taken 88 times.
✓ Branch 1 taken 22 times.
110 for (i = 0; i < 4; i++)
130 88 ctx->prediction_error[i] = MIN_ENERGY;
131
132 22 ff_acelp_filter_init(&ctx->acelpf_ctx);
133 22 ff_acelp_vectors_init(&ctx->acelpv_ctx);
134 22 ff_celp_filter_init(&ctx->celpf_ctx);
135 22 ff_celp_math_init(&ctx->celpm_ctx);
136 }
137
138 22 return 0;
139 }
140
141 /**
142 * Decode the frame header in the "MIME/storage" format. This format
143 * is simpler and does not carry the auxiliary frame information.
144 *
145 * @param[in] ctx The Context
146 * @param[in] buf Pointer to the input buffer
147 *
148 * @return The decoded header length in bytes
149 */
150 6268 static int decode_mime_header(AMRWBContext *ctx, const uint8_t *buf)
151 {
152 /* Decode frame header (1st octet) */
153 6268 ctx->fr_cur_mode = buf[0] >> 3 & 0x0F;
154 6268 ctx->fr_quality = (buf[0] & 0x4) == 0x4;
155
156 6268 return 1;
157 }
158
159 /**
160 * Decode quantized ISF vectors using 36-bit indexes (6K60 mode only).
161 *
162 * @param[in] ind Array of 5 indexes
163 * @param[out] isf_q Buffer for isf_q[LP_ORDER]
164 */
165 513 static void decode_isf_indices_36b(uint16_t *ind, float *isf_q)
166 {
167 int i;
168
169
2/2
✓ Branch 0 taken 4617 times.
✓ Branch 1 taken 513 times.
5130 for (i = 0; i < 9; i++)
170 4617 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
171
172
2/2
✓ Branch 0 taken 3591 times.
✓ Branch 1 taken 513 times.
4104 for (i = 0; i < 7; i++)
173 3591 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
174
175
2/2
✓ Branch 0 taken 2565 times.
✓ Branch 1 taken 513 times.
3078 for (i = 0; i < 5; i++)
176 2565 isf_q[i] += dico21_isf_36b[ind[2]][i] * (1.0f / (1 << 15));
177
178
2/2
✓ Branch 0 taken 2052 times.
✓ Branch 1 taken 513 times.
2565 for (i = 0; i < 4; i++)
179 2052 isf_q[i + 5] += dico22_isf_36b[ind[3]][i] * (1.0f / (1 << 15));
180
181
2/2
✓ Branch 0 taken 3591 times.
✓ Branch 1 taken 513 times.
4104 for (i = 0; i < 7; i++)
182 3591 isf_q[i + 9] += dico23_isf_36b[ind[4]][i] * (1.0f / (1 << 15));
183 513 }
184
185 /**
186 * Decode quantized ISF vectors using 46-bit indexes (except 6K60 mode).
187 *
188 * @param[in] ind Array of 7 indexes
189 * @param[out] isf_q Buffer for isf_q[LP_ORDER]
190 */
191 5755 static void decode_isf_indices_46b(uint16_t *ind, float *isf_q)
192 {
193 int i;
194
195
2/2
✓ Branch 0 taken 51795 times.
✓ Branch 1 taken 5755 times.
57550 for (i = 0; i < 9; i++)
196 51795 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
197
198
2/2
✓ Branch 0 taken 40285 times.
✓ Branch 1 taken 5755 times.
46040 for (i = 0; i < 7; i++)
199 40285 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
200
201
2/2
✓ Branch 0 taken 17265 times.
✓ Branch 1 taken 5755 times.
23020 for (i = 0; i < 3; i++)
202 17265 isf_q[i] += dico21_isf[ind[2]][i] * (1.0f / (1 << 15));
203
204
2/2
✓ Branch 0 taken 17265 times.
✓ Branch 1 taken 5755 times.
23020 for (i = 0; i < 3; i++)
205 17265 isf_q[i + 3] += dico22_isf[ind[3]][i] * (1.0f / (1 << 15));
206
207
2/2
✓ Branch 0 taken 17265 times.
✓ Branch 1 taken 5755 times.
23020 for (i = 0; i < 3; i++)
208 17265 isf_q[i + 6] += dico23_isf[ind[4]][i] * (1.0f / (1 << 15));
209
210
2/2
✓ Branch 0 taken 17265 times.
✓ Branch 1 taken 5755 times.
23020 for (i = 0; i < 3; i++)
211 17265 isf_q[i + 9] += dico24_isf[ind[5]][i] * (1.0f / (1 << 15));
212
213
2/2
✓ Branch 0 taken 23020 times.
✓ Branch 1 taken 5755 times.
28775 for (i = 0; i < 4; i++)
214 23020 isf_q[i + 12] += dico25_isf[ind[6]][i] * (1.0f / (1 << 15));
215 5755 }
216
217 /**
218 * Apply mean and past ISF values using the prediction factor.
219 * Updates past ISF vector.
220 *
221 * @param[in,out] isf_q Current quantized ISF
222 * @param[in,out] isf_past Past quantized ISF
223 */
224 6268 static void isf_add_mean_and_past(float *isf_q, float *isf_past)
225 {
226 int i;
227 float tmp;
228
229
2/2
✓ Branch 0 taken 100288 times.
✓ Branch 1 taken 6268 times.
106556 for (i = 0; i < LP_ORDER; i++) {
230 100288 tmp = isf_q[i];
231 100288 isf_q[i] += isf_mean[i] * (1.0f / (1 << 15));
232 100288 isf_q[i] += PRED_FACTOR * isf_past[i];
233 100288 isf_past[i] = tmp;
234 }
235 6268 }
236
237 /**
238 * Interpolate the fourth ISP vector from current and past frames
239 * to obtain an ISP vector for each subframe.
240 *
241 * @param[in,out] isp_q ISPs for each subframe
242 * @param[in] isp4_past Past ISP for subframe 4
243 */
244 6268 static void interpolate_isp(double isp_q[4][LP_ORDER], const double *isp4_past)
245 {
246 int i, k;
247
248
2/2
✓ Branch 0 taken 18804 times.
✓ Branch 1 taken 6268 times.
25072 for (k = 0; k < 3; k++) {
249 18804 float c = isfp_inter[k];
250
2/2
✓ Branch 0 taken 300864 times.
✓ Branch 1 taken 18804 times.
319668 for (i = 0; i < LP_ORDER; i++)
251 300864 isp_q[k][i] = (1.0 - c) * isp4_past[i] + c * isp_q[3][i];
252 }
253 6268 }
254
255 /**
256 * Decode an adaptive codebook index into pitch lag (except 6k60, 8k85 modes).
257 * Calculate integer lag and fractional lag always using 1/4 resolution.
258 * In 1st and 3rd subframes the index is relative to last subframe integer lag.
259 *
260 * @param[out] lag_int Decoded integer pitch lag
261 * @param[out] lag_frac Decoded fractional pitch lag
262 * @param[in] pitch_index Adaptive codebook pitch index
263 * @param[in,out] base_lag_int Base integer lag used in relative subframes
264 * @param[in] subframe Current subframe index (0 to 3)
265 */
266 20968 static void decode_pitch_lag_high(int *lag_int, int *lag_frac, int pitch_index,
267 uint8_t *base_lag_int, int subframe)
268 {
269
4/4
✓ Branch 0 taken 15726 times.
✓ Branch 1 taken 5242 times.
✓ Branch 2 taken 5242 times.
✓ Branch 3 taken 10484 times.
20968 if (subframe == 0 || subframe == 2) {
270
2/2
✓ Branch 0 taken 7630 times.
✓ Branch 1 taken 2854 times.
10484 if (pitch_index < 376) {
271 7630 *lag_int = (pitch_index + 137) >> 2;
272 7630 *lag_frac = pitch_index - (*lag_int << 2) + 136;
273
2/2
✓ Branch 0 taken 1248 times.
✓ Branch 1 taken 1606 times.
2854 } else if (pitch_index < 440) {
274 1248 *lag_int = (pitch_index + 257 - 376) >> 1;
275 1248 *lag_frac = (pitch_index - (*lag_int << 1) + 256 - 376) * 2;
276 /* the actual resolution is 1/2 but expressed as 1/4 */
277 } else {
278 1606 *lag_int = pitch_index - 280;
279 1606 *lag_frac = 0;
280 }
281 /* minimum lag for next subframe */
282 10484 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
283 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);
284 // XXX: the spec states clearly that *base_lag_int should be
285 // the nearest integer to *lag_int (minus 8), but the ref code
286 // actually always uses its floor, I'm following the latter
287 } else {
288 10484 *lag_int = (pitch_index + 1) >> 2;
289 10484 *lag_frac = pitch_index - (*lag_int << 2);
290 10484 *lag_int += *base_lag_int;
291 }
292 20968 }
293
294 /**
295 * Decode an adaptive codebook index into pitch lag for 8k85 and 6k60 modes.
296 * The description is analogous to decode_pitch_lag_high, but in 6k60 the
297 * relative index is used for all subframes except the first.
298 */
299 4104 static void decode_pitch_lag_low(int *lag_int, int *lag_frac, int pitch_index,
300 uint8_t *base_lag_int, int subframe, enum Mode mode)
301 {
302
6/6
✓ Branch 0 taken 3078 times.
✓ Branch 1 taken 1026 times.
✓ Branch 2 taken 1026 times.
✓ Branch 3 taken 2052 times.
✓ Branch 4 taken 513 times.
✓ Branch 5 taken 513 times.
4104 if (subframe == 0 || (subframe == 2 && mode != MODE_6k60)) {
303
2/2
✓ Branch 0 taken 850 times.
✓ Branch 1 taken 689 times.
1539 if (pitch_index < 116) {
304 850 *lag_int = (pitch_index + 69) >> 1;
305 850 *lag_frac = (pitch_index - (*lag_int << 1) + 68) * 2;
306 } else {
307 689 *lag_int = pitch_index - 24;
308 689 *lag_frac = 0;
309 }
310 // XXX: same problem as before
311 1539 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
312 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);
313 } else {
314 2565 *lag_int = (pitch_index + 1) >> 1;
315 2565 *lag_frac = (pitch_index - (*lag_int << 1)) * 2;
316 2565 *lag_int += *base_lag_int;
317 }
318 4104 }
319
320 /**
321 * Find the pitch vector by interpolating the past excitation at the
322 * pitch delay, which is obtained in this function.
323 *
324 * @param[in,out] ctx The context
325 * @param[in] amr_subframe Current subframe data
326 * @param[in] subframe Current subframe index (0 to 3)
327 */
328 25072 static void decode_pitch_vector(AMRWBContext *ctx,
329 const AMRWBSubFrame *amr_subframe,
330 const int subframe)
331 {
332 int pitch_lag_int, pitch_lag_frac;
333 int i;
334 25072 float *exc = ctx->excitation;
335 25072 enum Mode mode = ctx->fr_cur_mode;
336
337
2/2
✓ Branch 0 taken 4104 times.
✓ Branch 1 taken 20968 times.
25072 if (mode <= MODE_8k85) {
338 4104 decode_pitch_lag_low(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
339 &ctx->base_pitch_lag, subframe, mode);
340 } else
341 20968 decode_pitch_lag_high(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
342 &ctx->base_pitch_lag, subframe);
343
344 25072 ctx->pitch_lag_int = pitch_lag_int;
345 25072 pitch_lag_int += pitch_lag_frac > 0;
346
347 /* Calculate the pitch vector by interpolating the past excitation at the
348 pitch lag using a hamming windowed sinc function */
349 50144 ctx->acelpf_ctx.acelp_interpolatef(exc,
350 25072 exc + 1 - pitch_lag_int,
351 ac_inter, 4,
352
2/2
✓ Branch 0 taken 8454 times.
✓ Branch 1 taken 16618 times.
25072 pitch_lag_frac + (pitch_lag_frac > 0 ? 0 : 4),
353 LP_ORDER, AMRWB_SFR_SIZE + 1);
354
355 /* Check which pitch signal path should be used
356 * 6k60 and 8k85 modes have the ltp flag set to 0 */
357
2/2
✓ Branch 0 taken 7140 times.
✓ Branch 1 taken 17932 times.
25072 if (amr_subframe->ltp) {
358 7140 memcpy(ctx->pitch_vector, exc, AMRWB_SFR_SIZE * sizeof(float));
359 } else {
360
2/2
✓ Branch 0 taken 1147648 times.
✓ Branch 1 taken 17932 times.
1165580 for (i = 0; i < AMRWB_SFR_SIZE; i++)
361 1147648 ctx->pitch_vector[i] = 0.18 * exc[i - 1] + 0.64 * exc[i] +
362 1147648 0.18 * exc[i + 1];
363 17932 memcpy(exc, ctx->pitch_vector, AMRWB_SFR_SIZE * sizeof(float));
364 }
365 25072 }
366
367 /** Get x bits in the index interval [lsb,lsb+len-1] inclusive */
368 #define BIT_STR(x,lsb,len) av_mod_uintp2((x) >> (lsb), (len))
369
370 /** Get the bit at specified position */
371 #define BIT_POS(x, p) (((x) >> (p)) & 1)
372
373 /**
374 * The next six functions decode_[i]p_track decode exactly i pulses
375 * positions and amplitudes (-1 or 1) in a subframe track using
376 * an encoded pulse indexing (TS 26.190 section 5.8.2).
377 *
378 * The results are given in out[], in which a negative number means
379 * amplitude -1 and vice versa (i.e., ampl(x) = x / abs(x) ).
380 *
381 * @param[out] out Output buffer (writes i elements)
382 * @param[in] code Pulse index (no. of bits varies, see below)
383 * @param[in] m (log2) Number of potential positions
384 * @param[in] off Offset for decoded positions
385 */
386 106180 static inline void decode_1p_track(int *out, int code, int m, int off)
387 {
388 106180 int pos = BIT_STR(code, 0, m) + off; ///code: m+1 bits
389
390
2/2
✓ Branch 0 taken 53239 times.
✓ Branch 1 taken 52941 times.
106180 out[0] = BIT_POS(code, m) ? -pos : pos;
391 106180 }
392
393 147226 static inline void decode_2p_track(int *out, int code, int m, int off) ///code: 2m+1 bits
394 {
395 147226 int pos0 = BIT_STR(code, m, m) + off;
396 147226 int pos1 = BIT_STR(code, 0, m) + off;
397
398
2/2
✓ Branch 0 taken 74035 times.
✓ Branch 1 taken 73191 times.
147226 out[0] = BIT_POS(code, 2*m) ? -pos0 : pos0;
399
2/2
✓ Branch 0 taken 74035 times.
✓ Branch 1 taken 73191 times.
147226 out[1] = BIT_POS(code, 2*m) ? -pos1 : pos1;
400
2/2
✓ Branch 0 taken 66336 times.
✓ Branch 1 taken 80890 times.
147226 out[1] = pos0 > pos1 ? -out[1] : out[1];
401 147226 }
402
403 68748 static void decode_3p_track(int *out, int code, int m, int off) ///code: 3m+1 bits
404 {
405 68748 int half_2p = BIT_POS(code, 2*m - 1) << (m - 1);
406
407 68748 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
408 m - 1, off + half_2p);
409 68748 decode_1p_track(out + 2, BIT_STR(code, 2*m, m + 1), m, off);
410 68748 }
411
412 32563 static void decode_4p_track(int *out, int code, int m, int off) ///code: 4m bits
413 {
414 int half_4p, subhalf_2p;
415 32563 int b_offset = 1 << (m - 1);
416
417
4/5
✓ Branch 0 taken 3612 times.
✓ Branch 1 taken 7991 times.
✓ Branch 2 taken 12806 times.
✓ Branch 3 taken 8154 times.
✗ Branch 4 not taken.
32563 switch (BIT_STR(code, 4*m - 2, 2)) { /* case ID (2 bits) */
418 3612 case 0: /* 0 pulses in A, 4 pulses in B or vice versa */
419 3612 half_4p = BIT_POS(code, 4*m - 3) << (m - 1); // which has 4 pulses
420 3612 subhalf_2p = BIT_POS(code, 2*m - 3) << (m - 2);
421
422 3612 decode_2p_track(out, BIT_STR(code, 0, 2*m - 3),
423 3612 m - 2, off + half_4p + subhalf_2p);
424 3612 decode_2p_track(out + 2, BIT_STR(code, 2*m - 2, 2*m - 1),
425 m - 1, off + half_4p);
426 3612 break;
427 7991 case 1: /* 1 pulse in A, 3 pulses in B */
428 7991 decode_1p_track(out, BIT_STR(code, 3*m - 2, m),
429 m - 1, off);
430 7991 decode_3p_track(out + 1, BIT_STR(code, 0, 3*m - 2),
431 m - 1, off + b_offset);
432 7991 break;
433 12806 case 2: /* 2 pulses in each half */
434 12806 decode_2p_track(out, BIT_STR(code, 2*m - 1, 2*m - 1),
435 m - 1, off);
436 12806 decode_2p_track(out + 2, BIT_STR(code, 0, 2*m - 1),
437 m - 1, off + b_offset);
438 12806 break;
439 8154 case 3: /* 3 pulses in A, 1 pulse in B */
440 8154 decode_3p_track(out, BIT_STR(code, m, 3*m - 2),
441 m - 1, off);
442 8154 decode_1p_track(out + 3, BIT_STR(code, 0, m),
443 m - 1, off + b_offset);
444 8154 break;
445 }
446 32563 }
447
448 13079 static void decode_5p_track(int *out, int code, int m, int off) ///code: 5m bits
449 {
450 13079 int half_3p = BIT_POS(code, 5*m - 1) << (m - 1);
451
452 13079 decode_3p_track(out, BIT_STR(code, 2*m + 1, 3*m - 2),
453 m - 1, off + half_3p);
454
455 13079 decode_2p_track(out + 3, BIT_STR(code, 0, 2*m + 1), m, off);
456 13079 }
457
458 42832 static void decode_6p_track(int *out, int code, int m, int off) ///code: 6m-2 bits
459 {
460 42832 int b_offset = 1 << (m - 1);
461 /* which half has more pulses in cases 0 to 2 */
462 42832 int half_more = BIT_POS(code, 6*m - 5) << (m - 1);
463 42832 int half_other = b_offset - half_more;
464
465
4/5
✓ Branch 0 taken 1297 times.
✓ Branch 1 taken 7678 times.
✓ Branch 2 taken 20251 times.
✓ Branch 3 taken 13606 times.
✗ Branch 4 not taken.
42832 switch (BIT_STR(code, 6*m - 4, 2)) { /* case ID (2 bits) */
466 1297 case 0: /* 0 pulses in A, 6 pulses in B or vice versa */
467 1297 decode_1p_track(out, BIT_STR(code, 0, m),
468 m - 1, off + half_more);
469 1297 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
470 m - 1, off + half_more);
471 1297 break;
472 7678 case 1: /* 1 pulse in A, 5 pulses in B or vice versa */
473 7678 decode_1p_track(out, BIT_STR(code, 0, m),
474 m - 1, off + half_other);
475 7678 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
476 m - 1, off + half_more);
477 7678 break;
478 20251 case 2: /* 2 pulses in A, 4 pulses in B or vice versa */
479 20251 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
480 m - 1, off + half_other);
481 20251 decode_4p_track(out + 2, BIT_STR(code, 2*m - 1, 4*m - 4),
482 m - 1, off + half_more);
483 20251 break;
484 13606 case 3: /* 3 pulses in A, 3 pulses in B */
485 13606 decode_3p_track(out, BIT_STR(code, 3*m - 2, 3*m - 2),
486 m - 1, off);
487 13606 decode_3p_track(out + 3, BIT_STR(code, 0, 3*m - 2),
488 m - 1, off + b_offset);
489 13606 break;
490 }
491 42832 }
492
493 /**
494 * Decode the algebraic codebook index to pulse positions and signs,
495 * then construct the algebraic codebook vector.
496 *
497 * @param[out] fixed_vector Buffer for the fixed codebook excitation
498 * @param[in] pulse_hi MSBs part of the pulse index array (higher modes only)
499 * @param[in] pulse_lo LSBs part of the pulse index array
500 * @param[in] mode Mode of the current frame
501 */
502 25072 static void decode_fixed_vector(float *fixed_vector, const uint16_t *pulse_hi,
503 const uint16_t *pulse_lo, const enum Mode mode)
504 {
505 /* sig_pos stores for each track the decoded pulse position indexes
506 * (1-based) multiplied by its corresponding amplitude (+1 or -1) */
507 int sig_pos[4][6];
508
2/2
✓ Branch 0 taken 2052 times.
✓ Branch 1 taken 23020 times.
25072 int spacing = (mode == MODE_6k60) ? 2 : 4;
509 int i, j;
510
511
8/9
✓ Branch 0 taken 2052 times.
✓ Branch 1 taken 2052 times.
✓ Branch 2 taken 2052 times.
✓ Branch 3 taken 2052 times.
✓ Branch 4 taken 2052 times.
✓ Branch 5 taken 2052 times.
✓ Branch 6 taken 2052 times.
✓ Branch 7 taken 10708 times.
✗ Branch 8 not taken.
25072 switch (mode) {
512 2052 case MODE_6k60:
513
2/2
✓ Branch 0 taken 4104 times.
✓ Branch 1 taken 2052 times.
6156 for (i = 0; i < 2; i++)
514 4104 decode_1p_track(sig_pos[i], pulse_lo[i], 5, 1);
515 2052 break;
516 2052 case MODE_8k85:
517
2/2
✓ Branch 0 taken 8208 times.
✓ Branch 1 taken 2052 times.
10260 for (i = 0; i < 4; i++)
518 8208 decode_1p_track(sig_pos[i], pulse_lo[i], 4, 1);
519 2052 break;
520 2052 case MODE_12k65:
521
2/2
✓ Branch 0 taken 8208 times.
✓ Branch 1 taken 2052 times.
10260 for (i = 0; i < 4; i++)
522 8208 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
523 2052 break;
524 2052 case MODE_14k25:
525
2/2
✓ Branch 0 taken 4104 times.
✓ Branch 1 taken 2052 times.
6156 for (i = 0; i < 2; i++)
526 4104 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
527
2/2
✓ Branch 0 taken 4104 times.
✓ Branch 1 taken 2052 times.
6156 for (i = 2; i < 4; i++)
528 4104 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
529 2052 break;
530 2052 case MODE_15k85:
531
2/2
✓ Branch 0 taken 8208 times.
✓ Branch 1 taken 2052 times.
10260 for (i = 0; i < 4; i++)
532 8208 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
533 2052 break;
534 2052 case MODE_18k25:
535
2/2
✓ Branch 0 taken 8208 times.
✓ Branch 1 taken 2052 times.
10260 for (i = 0; i < 4; i++)
536 8208 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
537 8208 ((int) pulse_hi[i] << 14), 4, 1);
538 2052 break;
539 2052 case MODE_19k85:
540
2/2
✓ Branch 0 taken 4104 times.
✓ Branch 1 taken 2052 times.
6156 for (i = 0; i < 2; i++)
541 4104 decode_5p_track(sig_pos[i], (int) pulse_lo[i] +
542 4104 ((int) pulse_hi[i] << 10), 4, 1);
543
2/2
✓ Branch 0 taken 4104 times.
✓ Branch 1 taken 2052 times.
6156 for (i = 2; i < 4; i++)
544 4104 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
545 4104 ((int) pulse_hi[i] << 14), 4, 1);
546 2052 break;
547 10708 case MODE_23k05:
548 case MODE_23k85:
549
2/2
✓ Branch 0 taken 42832 times.
✓ Branch 1 taken 10708 times.
53540 for (i = 0; i < 4; i++)
550 42832 decode_6p_track(sig_pos[i], (int) pulse_lo[i] +
551 42832 ((int) pulse_hi[i] << 11), 4, 1);
552 10708 break;
553 }
554
555 25072 memset(fixed_vector, 0, sizeof(float) * AMRWB_SFR_SIZE);
556
557
2/2
✓ Branch 0 taken 100288 times.
✓ Branch 1 taken 25072 times.
125360 for (i = 0; i < 4; i++)
558
2/2
✓ Branch 0 taken 400632 times.
✓ Branch 1 taken 100288 times.
500920 for (j = 0; j < pulses_nb_per_mode_tr[mode][i]; j++) {
559 400632 int pos = (FFABS(sig_pos[i][j]) - 1) * spacing + i;
560
561
2/2
✓ Branch 0 taken 201133 times.
✓ Branch 1 taken 199499 times.
400632 fixed_vector[pos] += sig_pos[i][j] < 0 ? -1.0 : 1.0;
562 }
563 25072 }
564
565 /**
566 * Decode pitch gain and fixed gain correction factor.
567 *
568 * @param[in] vq_gain Vector-quantized index for gains
569 * @param[in] mode Mode of the current frame
570 * @param[out] fixed_gain_factor Decoded fixed gain correction factor
571 * @param[out] pitch_gain Decoded pitch gain
572 */
573 25072 static void decode_gains(const uint8_t vq_gain, const enum Mode mode,
574 float *fixed_gain_factor, float *pitch_gain)
575 {
576
2/2
✓ Branch 0 taken 4104 times.
✓ Branch 1 taken 20968 times.
25072 const int16_t *gains = (mode <= MODE_8k85 ? qua_gain_6b[vq_gain] :
577 20968 qua_gain_7b[vq_gain]);
578
579 25072 *pitch_gain = gains[0] * (1.0f / (1 << 14));
580 25072 *fixed_gain_factor = gains[1] * (1.0f / (1 << 11));
581 25072 }
582
583 /**
584 * Apply pitch sharpening filters to the fixed codebook vector.
585 *
586 * @param[in] ctx The context
587 * @param[in,out] fixed_vector Fixed codebook excitation
588 */
589 // XXX: Spec states this procedure should be applied when the pitch
590 // lag is less than 64, but this checking seems absent in reference and AMR-NB
591 25072 static void pitch_sharpening(AMRWBContext *ctx, float *fixed_vector)
592 {
593 int i;
594
595 /* Tilt part */
596
2/2
✓ Branch 0 taken 1579536 times.
✓ Branch 1 taken 25072 times.
1604608 for (i = AMRWB_SFR_SIZE - 1; i != 0; i--)
597 1579536 fixed_vector[i] -= fixed_vector[i - 1] * ctx->tilt_coef;
598
599 /* Periodicity enhancement part */
600
2/2
✓ Branch 0 taken 161953 times.
✓ Branch 1 taken 25072 times.
187025 for (i = ctx->pitch_lag_int; i < AMRWB_SFR_SIZE; i++)
601 161953 fixed_vector[i] += fixed_vector[i - ctx->pitch_lag_int] * 0.85;
602 25072 }
603
604 /**
605 * Calculate the voicing factor (-1.0 = unvoiced to 1.0 = voiced).
606 *
607 * @param[in] p_vector, f_vector Pitch and fixed excitation vectors
608 * @param[in] p_gain, f_gain Pitch and fixed gains
609 * @param[in] ctx The context
610 */
611 // XXX: There is something wrong with the precision here! The magnitudes
612 // of the energies are not correct. Please check the reference code carefully
613 25072 static float voice_factor(float *p_vector, float p_gain,
614 float *f_vector, float f_gain,
615 CELPMContext *ctx)
616 {
617 25072 double p_ener = (double) ctx->dot_productf(p_vector, p_vector,
618 25072 AMRWB_SFR_SIZE) *
619 25072 p_gain * p_gain;
620 25072 double f_ener = (double) ctx->dot_productf(f_vector, f_vector,
621 25072 AMRWB_SFR_SIZE) *
622 25072 f_gain * f_gain;
623
624 25072 return (p_ener - f_ener) / (p_ener + f_ener + 0.01);
625 }
626
627 /**
628 * Reduce fixed vector sparseness by smoothing with one of three IR filters,
629 * also known as "adaptive phase dispersion".
630 *
631 * @param[in] ctx The context
632 * @param[in,out] fixed_vector Unfiltered fixed vector
633 * @param[out] buf Space for modified vector if necessary
634 *
635 * @return The potentially overwritten filtered fixed vector address
636 */
637 25072 static float *anti_sparseness(AMRWBContext *ctx,
638 float *fixed_vector, float *buf)
639 {
640 int ir_filter_nr;
641
642
2/2
✓ Branch 0 taken 20968 times.
✓ Branch 1 taken 4104 times.
25072 if (ctx->fr_cur_mode > MODE_8k85) // no filtering in higher modes
643 20968 return fixed_vector;
644
645
2/2
✓ Branch 0 taken 2261 times.
✓ Branch 1 taken 1843 times.
4104 if (ctx->pitch_gain[0] < 0.6) {
646 2261 ir_filter_nr = 0; // strong filtering
647
2/2
✓ Branch 0 taken 712 times.
✓ Branch 1 taken 1131 times.
1843 } else if (ctx->pitch_gain[0] < 0.9) {
648 712 ir_filter_nr = 1; // medium filtering
649 } else
650 1131 ir_filter_nr = 2; // no filtering
651
652 /* detect 'onset' */
653
2/2
✓ Branch 0 taken 61 times.
✓ Branch 1 taken 4043 times.
4104 if (ctx->fixed_gain[0] > 3.0 * ctx->fixed_gain[1]) {
654
2/2
✓ Branch 0 taken 39 times.
✓ Branch 1 taken 22 times.
61 if (ir_filter_nr < 2)
655 39 ir_filter_nr++;
656 } else {
657 4043 int i, count = 0;
658
659
2/2
✓ Branch 0 taken 24258 times.
✓ Branch 1 taken 4043 times.
28301 for (i = 0; i < 6; i++)
660
2/2
✓ Branch 0 taken 13373 times.
✓ Branch 1 taken 10885 times.
24258 if (ctx->pitch_gain[i] < 0.6)
661 13373 count++;
662
663
2/2
✓ Branch 0 taken 2724 times.
✓ Branch 1 taken 1319 times.
4043 if (count > 2)
664 2724 ir_filter_nr = 0;
665
666
2/2
✓ Branch 0 taken 83 times.
✓ Branch 1 taken 3960 times.
4043 if (ir_filter_nr > ctx->prev_ir_filter_nr + 1)
667 83 ir_filter_nr--;
668 }
669
670 /* update ir filter strength history */
671 4104 ctx->prev_ir_filter_nr = ir_filter_nr;
672
673 4104 ir_filter_nr += (ctx->fr_cur_mode == MODE_8k85);
674
675
2/2
✓ Branch 0 taken 3197 times.
✓ Branch 1 taken 907 times.
4104 if (ir_filter_nr < 2) {
676 int i;
677 3197 const float *coef = ir_filters_lookup[ir_filter_nr];
678
679 /* Circular convolution code in the reference
680 * decoder was modified to avoid using one
681 * extra array. The filtered vector is given by:
682 *
683 * c2(n) = sum(i,0,len-1){ c(i) * coef( (n - i + len) % len ) }
684 */
685
686 3197 memset(buf, 0, sizeof(float) * AMRWB_SFR_SIZE);
687
2/2
✓ Branch 0 taken 204608 times.
✓ Branch 1 taken 3197 times.
207805 for (i = 0; i < AMRWB_SFR_SIZE; i++)
688
2/2
✓ Branch 0 taken 21471 times.
✓ Branch 1 taken 183137 times.
204608 if (fixed_vector[i])
689 21471 ff_celp_circ_addf(buf, buf, coef, i, fixed_vector[i],
690 AMRWB_SFR_SIZE);
691 3197 fixed_vector = buf;
692 }
693
694 4104 return fixed_vector;
695 }
696
697 /**
698 * Calculate a stability factor {teta} based on distance between
699 * current and past isf. A value of 1 shows maximum signal stability.
700 */
701 6268 static float stability_factor(const float *isf, const float *isf_past)
702 {
703 int i;
704 6268 float acc = 0.0;
705
706
2/2
✓ Branch 0 taken 94020 times.
✓ Branch 1 taken 6268 times.
100288 for (i = 0; i < LP_ORDER - 1; i++)
707 94020 acc += (isf[i] - isf_past[i]) * (isf[i] - isf_past[i]);
708
709 // XXX: This part is not so clear from the reference code
710 // the result is more accurate changing the "/ 256" to "* 512"
711
2/2
✓ Branch 0 taken 511 times.
✓ Branch 1 taken 5757 times.
6268 return FFMAX(0.0, 1.25 - acc * 0.8 * 512);
712 }
713
714 /**
715 * Apply a non-linear fixed gain smoothing in order to reduce
716 * fluctuation in the energy of excitation.
717 *
718 * @param[in] fixed_gain Unsmoothed fixed gain
719 * @param[in,out] prev_tr_gain Previous threshold gain (updated)
720 * @param[in] voice_fac Frame voicing factor
721 * @param[in] stab_fac Frame stability factor
722 *
723 * @return The smoothed gain
724 */
725 25072 static float noise_enhancer(float fixed_gain, float *prev_tr_gain,
726 float voice_fac, float stab_fac)
727 {
728 25072 float sm_fac = 0.5 * (1 - voice_fac) * stab_fac;
729 float g0;
730
731 // XXX: the following fixed-point constants used to in(de)crement
732 // gain by 1.5dB were taken from the reference code, maybe it could
733 // be simpler
734
2/2
✓ Branch 0 taken 13115 times.
✓ Branch 1 taken 11957 times.
25072 if (fixed_gain < *prev_tr_gain) {
735
2/2
✓ Branch 0 taken 7438 times.
✓ Branch 1 taken 5677 times.
13115 g0 = FFMIN(*prev_tr_gain, fixed_gain + fixed_gain *
736 (6226 * (1.0f / (1 << 15)))); // +1.5 dB
737 } else
738
2/2
✓ Branch 0 taken 5835 times.
✓ Branch 1 taken 6122 times.
11957 g0 = FFMAX(*prev_tr_gain, fixed_gain *
739 (27536 * (1.0f / (1 << 15)))); // -1.5 dB
740
741 25072 *prev_tr_gain = g0; // update next frame threshold
742
743 25072 return sm_fac * g0 + (1 - sm_fac) * fixed_gain;
744 }
745
746 /**
747 * Filter the fixed_vector to emphasize the higher frequencies.
748 *
749 * @param[in,out] fixed_vector Fixed codebook vector
750 * @param[in] voice_fac Frame voicing factor
751 */
752 25072 static void pitch_enhancer(float *fixed_vector, float voice_fac)
753 {
754 int i;
755 25072 float cpe = 0.125 * (1 + voice_fac);
756 25072 float last = fixed_vector[0]; // holds c(i - 1)
757
758 25072 fixed_vector[0] -= cpe * fixed_vector[1];
759
760
2/2
✓ Branch 0 taken 1554464 times.
✓ Branch 1 taken 25072 times.
1579536 for (i = 1; i < AMRWB_SFR_SIZE - 1; i++) {
761 1554464 float cur = fixed_vector[i];
762
763 1554464 fixed_vector[i] -= cpe * (last + fixed_vector[i + 1]);
764 1554464 last = cur;
765 }
766
767 25072 fixed_vector[AMRWB_SFR_SIZE - 1] -= cpe * last;
768 25072 }
769
770 /**
771 * Conduct 16th order linear predictive coding synthesis from excitation.
772 *
773 * @param[in] ctx Pointer to the AMRWBContext
774 * @param[in] lpc Pointer to the LPC coefficients
775 * @param[out] excitation Buffer for synthesis final excitation
776 * @param[in] fixed_gain Fixed codebook gain for synthesis
777 * @param[in] fixed_vector Algebraic codebook vector
778 * @param[in,out] samples Pointer to the output samples and memory
779 */
780 25072 static void synthesis(AMRWBContext *ctx, float *lpc, float *excitation,
781 float fixed_gain, const float *fixed_vector,
782 float *samples)
783 {
784 25072 ctx->acelpv_ctx.weighted_vector_sumf(excitation, ctx->pitch_vector, fixed_vector,
785 ctx->pitch_gain[0], fixed_gain, AMRWB_SFR_SIZE);
786
787 /* emphasize pitch vector contribution in low bitrate modes */
788
4/4
✓ Branch 0 taken 14894 times.
✓ Branch 1 taken 10178 times.
✓ Branch 2 taken 1897 times.
✓ Branch 3 taken 12997 times.
25072 if (ctx->pitch_gain[0] > 0.5 && ctx->fr_cur_mode <= MODE_8k85) {
789 int i;
790 1897 float energy = ctx->celpm_ctx.dot_productf(excitation, excitation,
791 AMRWB_SFR_SIZE);
792
793 // XXX: Weird part in both ref code and spec. A unknown parameter
794 // {beta} seems to be identical to the current pitch gain
795 1897 float pitch_factor = 0.25 * ctx->pitch_gain[0] * ctx->pitch_gain[0];
796
797
2/2
✓ Branch 0 taken 121408 times.
✓ Branch 1 taken 1897 times.
123305 for (i = 0; i < AMRWB_SFR_SIZE; i++)
798 121408 excitation[i] += pitch_factor * ctx->pitch_vector[i];
799
800 1897 ff_scale_vector_to_given_sum_of_squares(excitation, excitation,
801 energy, AMRWB_SFR_SIZE);
802 }
803
804 25072 ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation,
805 AMRWB_SFR_SIZE, LP_ORDER);
806 25072 }
807
808 /**
809 * Apply to synthesis a de-emphasis filter of the form:
810 * H(z) = 1 / (1 - m * z^-1)
811 *
812 * @param[out] out Output buffer
813 * @param[in] in Input samples array with in[-1]
814 * @param[in] m Filter coefficient
815 * @param[in,out] mem State from last filtering
816 */
817 25072 static void de_emphasis(float *out, float *in, float m, float mem[1])
818 {
819 int i;
820
821 25072 out[0] = in[0] + m * mem[0];
822
823
2/2
✓ Branch 0 taken 1579536 times.
✓ Branch 1 taken 25072 times.
1604608 for (i = 1; i < AMRWB_SFR_SIZE; i++)
824 1579536 out[i] = in[i] + out[i - 1] * m;
825
826 25072 mem[0] = out[AMRWB_SFR_SIZE - 1];
827 25072 }
828
829 /**
830 * Upsample a signal by 5/4 ratio (from 12.8kHz to 16kHz) using
831 * a FIR interpolation filter. Uses past data from before *in address.
832 *
833 * @param[out] out Buffer for interpolated signal
834 * @param[in] in Current signal data (length 0.8*o_size)
835 * @param[in] o_size Output signal length
836 * @param[in] ctx The context
837 */
838 25072 static void upsample_5_4(float *out, const float *in, int o_size, CELPMContext *ctx)
839 {
840 25072 const float *in0 = in - UPS_FIR_SIZE + 1;
841 int i, j, k;
842 25072 int int_part = 0, frac_part;
843
844 25072 i = 0;
845
2/2
✓ Branch 0 taken 401152 times.
✓ Branch 1 taken 25072 times.
426224 for (j = 0; j < o_size / 5; j++) {
846 401152 out[i] = in[int_part];
847 401152 frac_part = 4;
848 401152 i++;
849
850
2/2
✓ Branch 0 taken 1604608 times.
✓ Branch 1 taken 401152 times.
2005760 for (k = 1; k < 5; k++) {
851 3209216 out[i] = ctx->dot_productf(in0 + int_part,
852 1604608 upsample_fir[4 - frac_part],
853 UPS_MEM_SIZE);
854 1604608 int_part++;
855 1604608 frac_part--;
856 1604608 i++;
857 }
858 }
859 25072 }
860
861 /**
862 * Calculate the high-band gain based on encoded index (23k85 mode) or
863 * on the low-band speech signal and the Voice Activity Detection flag.
864 *
865 * @param[in] ctx The context
866 * @param[in] synth LB speech synthesis at 12.8k
867 * @param[in] hb_idx Gain index for mode 23k85 only
868 * @param[in] vad VAD flag for the frame
869 */
870 25072 static float find_hb_gain(AMRWBContext *ctx, const float *synth,
871 uint16_t hb_idx, uint8_t vad)
872 {
873 25072 int wsp = (vad > 0);
874 float tilt;
875 float tmp;
876
877
2/2
✓ Branch 0 taken 8656 times.
✓ Branch 1 taken 16416 times.
25072 if (ctx->fr_cur_mode == MODE_23k85)
878 8656 return qua_hb_gain[hb_idx] * (1.0f / (1 << 14));
879
880 16416 tmp = ctx->celpm_ctx.dot_productf(synth, synth + 1, AMRWB_SFR_SIZE - 1);
881
882
2/2
✓ Branch 0 taken 15653 times.
✓ Branch 1 taken 763 times.
16416 if (tmp > 0) {
883 15653 tilt = tmp / ctx->celpm_ctx.dot_productf(synth, synth, AMRWB_SFR_SIZE);
884 } else
885 763 tilt = 0;
886
887 /* return gain bounded by [0.1, 1.0] */
888 16416 return av_clipf((1.0 - tilt) * (1.25 - 0.25 * wsp), 0.1, 1.0);
889 }
890
891 /**
892 * Generate the high-band excitation with the same energy from the lower
893 * one and scaled by the given gain.
894 *
895 * @param[in] ctx The context
896 * @param[out] hb_exc Buffer for the excitation
897 * @param[in] synth_exc Low-band excitation used for synthesis
898 * @param[in] hb_gain Wanted excitation gain
899 */
900 25072 static void scaled_hb_excitation(AMRWBContext *ctx, float *hb_exc,
901 const float *synth_exc, float hb_gain)
902 {
903 int i;
904 25072 float energy = ctx->celpm_ctx.dot_productf(synth_exc, synth_exc,
905 AMRWB_SFR_SIZE);
906
907 /* Generate a white-noise excitation */
908
2/2
✓ Branch 0 taken 2005760 times.
✓ Branch 1 taken 25072 times.
2030832 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
909 2005760 hb_exc[i] = 32768.0 - (uint16_t) av_lfg_get(&ctx->prng);
910
911 25072 ff_scale_vector_to_given_sum_of_squares(hb_exc, hb_exc,
912 25072 energy * hb_gain * hb_gain,
913 AMRWB_SFR_SIZE_16k);
914 25072 }
915
916 /**
917 * Calculate the auto-correlation for the ISF difference vector.
918 */
919 6156 static float auto_correlation(float *diff_isf, float mean, int lag)
920 {
921 int i;
922 6156 float sum = 0.0;
923
924
2/2
✓ Branch 0 taken 43092 times.
✓ Branch 1 taken 6156 times.
49248 for (i = 7; i < LP_ORDER - 2; i++) {
925 43092 float prod = (diff_isf[i] - mean) * (diff_isf[i - lag] - mean);
926 43092 sum += prod * prod;
927 }
928 6156 return sum;
929 }
930
931 /**
932 * Extrapolate a ISF vector to the 16kHz range (20th order LP)
933 * used at mode 6k60 LP filter for the high frequency band.
934 *
935 * @param[out] isf Buffer for extrapolated isf; contains LP_ORDER
936 * values on input
937 */
938 2052 static void extrapolate_isf(float isf[LP_ORDER_16k])
939 {
940 float diff_isf[LP_ORDER - 2], diff_mean;
941 float corr_lag[3];
942 float est, scale;
943 int i, j, i_max_corr;
944
945 2052 isf[LP_ORDER_16k - 1] = isf[LP_ORDER - 1];
946
947 /* Calculate the difference vector */
948
2/2
✓ Branch 0 taken 28728 times.
✓ Branch 1 taken 2052 times.
30780 for (i = 0; i < LP_ORDER - 2; i++)
949 28728 diff_isf[i] = isf[i + 1] - isf[i];
950
951 2052 diff_mean = 0.0;
952
2/2
✓ Branch 0 taken 24624 times.
✓ Branch 1 taken 2052 times.
26676 for (i = 2; i < LP_ORDER - 2; i++)
953 24624 diff_mean += diff_isf[i] * (1.0f / (LP_ORDER - 4));
954
955 /* Find which is the maximum autocorrelation */
956 2052 i_max_corr = 0;
957
2/2
✓ Branch 0 taken 6156 times.
✓ Branch 1 taken 2052 times.
8208 for (i = 0; i < 3; i++) {
958 6156 corr_lag[i] = auto_correlation(diff_isf, diff_mean, i + 2);
959
960
2/2
✓ Branch 0 taken 2154 times.
✓ Branch 1 taken 4002 times.
6156 if (corr_lag[i] > corr_lag[i_max_corr])
961 2154 i_max_corr = i;
962 }
963 2052 i_max_corr++;
964
965
2/2
✓ Branch 0 taken 8208 times.
✓ Branch 1 taken 2052 times.
10260 for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
966 8208 isf[i] = isf[i - 1] + isf[i - 1 - i_max_corr]
967 8208 - isf[i - 2 - i_max_corr];
968
969 /* Calculate an estimate for ISF(18) and scale ISF based on the error */
970 2052 est = 7965 + (isf[2] - isf[3] - isf[4]) / 6.0;
971
1/2
✓ Branch 0 taken 2052 times.
✗ Branch 1 not taken.
2052 scale = 0.5 * (FFMIN(est, 7600) - isf[LP_ORDER - 2]) /
972 2052 (isf[LP_ORDER_16k - 2] - isf[LP_ORDER - 2]);
973
974
2/2
✓ Branch 0 taken 8208 times.
✓ Branch 1 taken 2052 times.
10260 for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++)
975 8208 diff_isf[j] = scale * (isf[i] - isf[i - 1]);
976
977 /* Stability insurance */
978
2/2
✓ Branch 0 taken 6156 times.
✓ Branch 1 taken 2052 times.
8208 for (i = 1; i < LP_ORDER_16k - LP_ORDER; i++)
979
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 6156 times.
6156 if (diff_isf[i] + diff_isf[i - 1] < 5.0) {
980 if (diff_isf[i] > diff_isf[i - 1]) {
981 diff_isf[i - 1] = 5.0 - diff_isf[i];
982 } else
983 diff_isf[i] = 5.0 - diff_isf[i - 1];
984 }
985
986
2/2
✓ Branch 0 taken 8208 times.
✓ Branch 1 taken 2052 times.
10260 for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++)
987 8208 isf[i] = isf[i - 1] + diff_isf[j] * (1.0f / (1 << 15));
988
989 /* Scale the ISF vector for 16000 Hz */
990
2/2
✓ Branch 0 taken 38988 times.
✓ Branch 1 taken 2052 times.
41040 for (i = 0; i < LP_ORDER_16k - 1; i++)
991 38988 isf[i] *= 0.8;
992 2052 }
993
994 /**
995 * Spectral expand the LP coefficients using the equation:
996 * y[i] = x[i] * (gamma ** i)
997 *
998 * @param[out] out Output buffer (may use input array)
999 * @param[in] lpc LP coefficients array
1000 * @param[in] gamma Weighting factor
1001 * @param[in] size LP array size
1002 */
1003 25072 static void lpc_weighting(float *out, const float *lpc, float gamma, int size)
1004 {
1005 int i;
1006 25072 float fac = gamma;
1007
1008
2/2
✓ Branch 0 taken 409360 times.
✓ Branch 1 taken 25072 times.
434432 for (i = 0; i < size; i++) {
1009 409360 out[i] = lpc[i] * fac;
1010 409360 fac *= gamma;
1011 }
1012 25072 }
1013
1014 /**
1015 * Conduct 20th order linear predictive coding synthesis for the high
1016 * frequency band excitation at 16kHz.
1017 *
1018 * @param[in] ctx The context
1019 * @param[in] subframe Current subframe index (0 to 3)
1020 * @param[in,out] samples Pointer to the output speech samples
1021 * @param[in] exc Generated white-noise scaled excitation
1022 * @param[in] isf Current frame isf vector
1023 * @param[in] isf_past Past frame final isf vector
1024 */
1025 25072 static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples,
1026 const float *exc, const float *isf, const float *isf_past)
1027 {
1028 float hb_lpc[LP_ORDER_16k];
1029 25072 enum Mode mode = ctx->fr_cur_mode;
1030
1031
2/2
✓ Branch 0 taken 2052 times.
✓ Branch 1 taken 23020 times.
25072 if (mode == MODE_6k60) {
1032 float e_isf[LP_ORDER_16k]; // ISF vector for extrapolation
1033 double e_isp[LP_ORDER_16k];
1034
1035 2052 ctx->acelpv_ctx.weighted_vector_sumf(e_isf, isf_past, isf, isfp_inter[subframe],
1036 2052 1.0 - isfp_inter[subframe], LP_ORDER);
1037
1038 2052 extrapolate_isf(e_isf);
1039
1040 2052 e_isf[LP_ORDER_16k - 1] *= 2.0;
1041 2052 ff_acelp_lsf2lspd(e_isp, e_isf, LP_ORDER_16k);
1042 2052 ff_amrwb_lsp2lpc(e_isp, hb_lpc, LP_ORDER_16k);
1043
1044 2052 lpc_weighting(hb_lpc, hb_lpc, 0.9, LP_ORDER_16k);
1045 } else {
1046 23020 lpc_weighting(hb_lpc, ctx->lp_coef[subframe], 0.6, LP_ORDER);
1047 }
1048
1049
2/2
✓ Branch 0 taken 2052 times.
✓ Branch 1 taken 23020 times.
25072 ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, hb_lpc, exc, AMRWB_SFR_SIZE_16k,
1050 (mode == MODE_6k60) ? LP_ORDER_16k : LP_ORDER);
1051 25072 }
1052
1053 /**
1054 * Apply a 15th order filter to high-band samples.
1055 * The filter characteristic depends on the given coefficients.
1056 *
1057 * @param[out] out Buffer for filtered output
1058 * @param[in] fir_coef Filter coefficients
1059 * @param[in,out] mem State from last filtering (updated)
1060 * @param[in] in Input speech data (high-band)
1061 *
1062 * @remark It is safe to pass the same array in in and out parameters
1063 */
1064
1065 #ifndef hb_fir_filter
1066 33728 static void hb_fir_filter(float *out, const float fir_coef[HB_FIR_SIZE + 1],
1067 float mem[HB_FIR_SIZE], const float *in)
1068 {
1069 int i, j;
1070 float data[AMRWB_SFR_SIZE_16k + HB_FIR_SIZE]; // past and current samples
1071
1072 33728 memcpy(data, mem, HB_FIR_SIZE * sizeof(float));
1073 33728 memcpy(data + HB_FIR_SIZE, in, AMRWB_SFR_SIZE_16k * sizeof(float));
1074
1075
2/2
✓ Branch 0 taken 2698240 times.
✓ Branch 1 taken 33728 times.
2731968 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) {
1076 2698240 out[i] = 0.0;
1077
2/2
✓ Branch 0 taken 83645440 times.
✓ Branch 1 taken 2698240 times.
86343680 for (j = 0; j <= HB_FIR_SIZE; j++)
1078 83645440 out[i] += data[i + j] * fir_coef[j];
1079 }
1080
1081 33728 memcpy(mem, data + AMRWB_SFR_SIZE_16k, HB_FIR_SIZE * sizeof(float));
1082 33728 }
1083 #endif /* hb_fir_filter */
1084
1085 /**
1086 * Update context state before the next subframe.
1087 */
1088 25072 static void update_sub_state(AMRWBContext *ctx)
1089 {
1090 25072 memmove(&ctx->excitation_buf[0], &ctx->excitation_buf[AMRWB_SFR_SIZE],
1091 (AMRWB_P_DELAY_MAX + LP_ORDER + 1) * sizeof(float));
1092
1093 25072 memmove(&ctx->pitch_gain[1], &ctx->pitch_gain[0], 5 * sizeof(float));
1094 25072 memmove(&ctx->fixed_gain[1], &ctx->fixed_gain[0], 1 * sizeof(float));
1095
1096 25072 memmove(&ctx->samples_az[0], &ctx->samples_az[AMRWB_SFR_SIZE],
1097 LP_ORDER * sizeof(float));
1098 25072 memmove(&ctx->samples_up[0], &ctx->samples_up[AMRWB_SFR_SIZE],
1099 UPS_MEM_SIZE * sizeof(float));
1100 25072 memmove(&ctx->samples_hb[0], &ctx->samples_hb[AMRWB_SFR_SIZE_16k],
1101 LP_ORDER_16k * sizeof(float));
1102 25072 }
1103
1104 6268 static int amrwb_decode_frame(AVCodecContext *avctx, void *data,
1105 int *got_frame_ptr, AVPacket *avpkt)
1106 {
1107 6268 AMRWBChannelsContext *s = avctx->priv_data;
1108 6268 AVFrame *frame = data;
1109 6268 const uint8_t *buf = avpkt->data;
1110 6268 int buf_size = avpkt->size;
1111 int sub, i, ret;
1112
1113 /* get output buffer */
1114 6268 frame->nb_samples = 4 * AMRWB_SFR_SIZE_16k;
1115
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 6268 times.
6268 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
1116 return ret;
1117
1118
2/2
✓ Branch 0 taken 6268 times.
✓ Branch 1 taken 6268 times.
12536 for (int ch = 0; ch < avctx->channels; ch++) {
1119 6268 AMRWBContext *ctx = &s->ch[ch];
1120 6268 AMRWBFrame *cf = &ctx->frame;
1121 int expected_fr_size, header_size;
1122 float spare_vector[AMRWB_SFR_SIZE]; // extra stack space to hold result from anti-sparseness processing
1123 float fixed_gain_factor; // fixed gain correction factor (gamma)
1124 float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
1125 float synth_fixed_gain; // the fixed gain that synthesis should use
1126 float voice_fac, stab_fac; // parameters used for gain smoothing
1127 float synth_exc[AMRWB_SFR_SIZE]; // post-processed excitation for synthesis
1128 float hb_exc[AMRWB_SFR_SIZE_16k]; // excitation for the high frequency band
1129 float hb_samples[AMRWB_SFR_SIZE_16k]; // filtered high-band samples from synthesis
1130 float hb_gain;
1131 6268 float *buf_out = (float *)frame->extended_data[ch];
1132
1133 6268 header_size = decode_mime_header(ctx, buf);
1134 6268 expected_fr_size = ((cf_sizes_wb[ctx->fr_cur_mode] + 7) >> 3) + 1;
1135
1136
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 6268 times.
6268 if (!ctx->fr_quality)
1137 av_log(avctx, AV_LOG_ERROR, "Encountered a bad or corrupted frame\n");
1138
1139
2/4
✓ Branch 0 taken 6268 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 6268 times.
6268 if (ctx->fr_cur_mode == NO_DATA || !ctx->fr_quality) {
1140 /* The specification suggests a "random signal" and
1141 "a muting technique" to "gradually decrease the output level". */
1142 av_samples_set_silence(&frame->extended_data[ch], 0, frame->nb_samples, 1, AV_SAMPLE_FMT_FLT);
1143 buf += expected_fr_size;
1144 buf_size -= expected_fr_size;
1145 continue;
1146 }
1147
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 6268 times.
6268 if (ctx->fr_cur_mode > MODE_SID) {
1148 av_log(avctx, AV_LOG_ERROR,
1149 "Invalid mode %d\n", ctx->fr_cur_mode);
1150 return AVERROR_INVALIDDATA;
1151 }
1152
1153
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 6268 times.
6268 if (buf_size < expected_fr_size) {
1154 av_log(avctx, AV_LOG_ERROR,
1155 "Frame too small (%d bytes). Truncated file?\n", buf_size);
1156 *got_frame_ptr = 0;
1157 return AVERROR_INVALIDDATA;
1158 }
1159
1160
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 6268 times.
6268 if (ctx->fr_cur_mode == MODE_SID) { /* Comfort noise frame */
1161 avpriv_request_sample(avctx, "SID mode");
1162 return AVERROR_PATCHWELCOME;
1163 }
1164
1165 6268 ff_amr_bit_reorder((uint16_t *) &ctx->frame, sizeof(AMRWBFrame),
1166 6268 buf + header_size, amr_bit_orderings_by_mode[ctx->fr_cur_mode]);
1167
1168 /* Decode the quantized ISF vector */
1169
2/2
✓ Branch 0 taken 513 times.
✓ Branch 1 taken 5755 times.
6268 if (ctx->fr_cur_mode == MODE_6k60) {
1170 513 decode_isf_indices_36b(cf->isp_id, ctx->isf_cur);
1171 } else {
1172 5755 decode_isf_indices_46b(cf->isp_id, ctx->isf_cur);
1173 }
1174
1175 6268 isf_add_mean_and_past(ctx->isf_cur, ctx->isf_q_past);
1176 6268 ff_set_min_dist_lsf(ctx->isf_cur, MIN_ISF_SPACING, LP_ORDER - 1);
1177
1178 6268 stab_fac = stability_factor(ctx->isf_cur, ctx->isf_past_final);
1179
1180 6268 ctx->isf_cur[LP_ORDER - 1] *= 2.0;
1181 6268 ff_acelp_lsf2lspd(ctx->isp[3], ctx->isf_cur, LP_ORDER);
1182
1183 /* Generate a ISP vector for each subframe */
1184
2/2
✓ Branch 0 taken 22 times.
✓ Branch 1 taken 6246 times.
6268 if (ctx->first_frame) {
1185 22 ctx->first_frame = 0;
1186 22 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(double));
1187 }
1188 6268 interpolate_isp(ctx->isp, ctx->isp_sub4_past);
1189
1190
2/2
✓ Branch 0 taken 25072 times.
✓ Branch 1 taken 6268 times.
31340 for (sub = 0; sub < 4; sub++)
1191 25072 ff_amrwb_lsp2lpc(ctx->isp[sub], ctx->lp_coef[sub], LP_ORDER);
1192
1193
2/2
✓ Branch 0 taken 25072 times.
✓ Branch 1 taken 6268 times.
31340 for (sub = 0; sub < 4; sub++) {
1194 25072 const AMRWBSubFrame *cur_subframe = &cf->subframe[sub];
1195 25072 float *sub_buf = buf_out + sub * AMRWB_SFR_SIZE_16k;
1196
1197 /* Decode adaptive codebook (pitch vector) */
1198 25072 decode_pitch_vector(ctx, cur_subframe, sub);
1199 /* Decode innovative codebook (fixed vector) */
1200 25072 decode_fixed_vector(ctx->fixed_vector, cur_subframe->pul_ih,
1201 25072 cur_subframe->pul_il, ctx->fr_cur_mode);
1202
1203 25072 pitch_sharpening(ctx, ctx->fixed_vector);
1204
1205 25072 decode_gains(cur_subframe->vq_gain, ctx->fr_cur_mode,
1206 &fixed_gain_factor, &ctx->pitch_gain[0]);
1207
1208 25072 ctx->fixed_gain[0] =
1209 25072 ff_amr_set_fixed_gain(fixed_gain_factor,
1210 25072 ctx->celpm_ctx.dot_productf(ctx->fixed_vector,
1211 25072 ctx->fixed_vector,
1212 AMRWB_SFR_SIZE) /
1213 AMRWB_SFR_SIZE,
1214 25072 ctx->prediction_error,
1215 ENERGY_MEAN, energy_pred_fac);
1216
1217 /* Calculate voice factor and store tilt for next subframe */
1218 25072 voice_fac = voice_factor(ctx->pitch_vector, ctx->pitch_gain[0],
1219 25072 ctx->fixed_vector, ctx->fixed_gain[0],
1220 &ctx->celpm_ctx);
1221 25072 ctx->tilt_coef = voice_fac * 0.25 + 0.25;
1222
1223 /* Construct current excitation */
1224
2/2
✓ Branch 0 taken 1604608 times.
✓ Branch 1 taken 25072 times.
1629680 for (i = 0; i < AMRWB_SFR_SIZE; i++) {
1225 1604608 ctx->excitation[i] *= ctx->pitch_gain[0];
1226 1604608 ctx->excitation[i] += ctx->fixed_gain[0] * ctx->fixed_vector[i];
1227 1604608 ctx->excitation[i] = truncf(ctx->excitation[i]);
1228 }
1229
1230 /* Post-processing of excitation elements */
1231 25072 synth_fixed_gain = noise_enhancer(ctx->fixed_gain[0], &ctx->prev_tr_gain,
1232 voice_fac, stab_fac);
1233
1234 25072 synth_fixed_vector = anti_sparseness(ctx, ctx->fixed_vector,
1235 spare_vector);
1236
1237 25072 pitch_enhancer(synth_fixed_vector, voice_fac);
1238
1239 25072 synthesis(ctx, ctx->lp_coef[sub], synth_exc, synth_fixed_gain,
1240 synth_fixed_vector, &ctx->samples_az[LP_ORDER]);
1241
1242 /* Synthesis speech post-processing */
1243 25072 de_emphasis(&ctx->samples_up[UPS_MEM_SIZE],
1244 25072 &ctx->samples_az[LP_ORDER], PREEMPH_FAC, ctx->demph_mem);
1245
1246 25072 ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(&ctx->samples_up[UPS_MEM_SIZE],
1247 25072 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_31_poles,
1248 25072 hpf_31_gain, ctx->hpf_31_mem, AMRWB_SFR_SIZE);
1249
1250 25072 upsample_5_4(sub_buf, &ctx->samples_up[UPS_FIR_SIZE],
1251 AMRWB_SFR_SIZE_16k, &ctx->celpm_ctx);
1252
1253 /* High frequency band (6.4 - 7.0 kHz) generation part */
1254 25072 ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(hb_samples,
1255 25072 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_400_poles,
1256 25072 hpf_400_gain, ctx->hpf_400_mem, AMRWB_SFR_SIZE);
1257
1258 25072 hb_gain = find_hb_gain(ctx, hb_samples,
1259 25072 cur_subframe->hb_gain, cf->vad);
1260
1261 25072 scaled_hb_excitation(ctx, hb_exc, synth_exc, hb_gain);
1262
1263 25072 hb_synthesis(ctx, sub, &ctx->samples_hb[LP_ORDER_16k],
1264 25072 hb_exc, ctx->isf_cur, ctx->isf_past_final);
1265
1266 /* High-band post-processing filters */
1267 25072 hb_fir_filter(hb_samples, bpf_6_7_coef, ctx->bpf_6_7_mem,
1268 25072 &ctx->samples_hb[LP_ORDER_16k]);
1269
1270
2/2
✓ Branch 0 taken 8656 times.
✓ Branch 1 taken 16416 times.
25072 if (ctx->fr_cur_mode == MODE_23k85)
1271 8656 hb_fir_filter(hb_samples, lpf_7_coef, ctx->lpf_7_mem,
1272 hb_samples);
1273
1274 /* Add the low and high frequency bands */
1275
2/2
✓ Branch 0 taken 2005760 times.
✓ Branch 1 taken 25072 times.
2030832 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
1276 2005760 sub_buf[i] = (sub_buf[i] + hb_samples[i]) * (1.0f / (1 << 15));
1277
1278 /* Update buffers and history */
1279 25072 update_sub_state(ctx);
1280 }
1281
1282 /* update state for next frame */
1283 6268 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(ctx->isp[3][0]));
1284 6268 memcpy(ctx->isf_past_final, ctx->isf_cur, LP_ORDER * sizeof(float));
1285
1286 6268 buf += expected_fr_size;
1287 6268 buf_size -= expected_fr_size;
1288 }
1289
1290 6268 *got_frame_ptr = 1;
1291
1292 6268 return avpkt->size;
1293 }
1294
1295 const AVCodec ff_amrwb_decoder = {
1296 .name = "amrwb",
1297 .long_name = NULL_IF_CONFIG_SMALL("AMR-WB (Adaptive Multi-Rate WideBand)"),
1298 .type = AVMEDIA_TYPE_AUDIO,
1299 .id = AV_CODEC_ID_AMR_WB,
1300 .priv_data_size = sizeof(AMRWBChannelsContext),
1301 .init = amrwb_decode_init,
1302 .decode = amrwb_decode_frame,
1303 .capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF,
1304 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT,
1305 AV_SAMPLE_FMT_NONE },
1306 };
1307