FFmpeg coverage

Directory: ../../../ffmpeg/
File: src/libavcodec/amrnbdec.c
Date: 2022-01-16 20:33:26
Exec Total Coverage
Lines: 369 389 94.9%
Branches: 185 204 90.7%
Line Branch Exec Source
1 /*
2 * AMR narrowband decoder
3 * Copyright (c) 2006-2007 Robert Swain
4 * Copyright (c) 2009 Colin McQuillan
5 *
6 * This file is part of FFmpeg.
7 *
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 */
22
23
24 /**
25 * @file
26 * AMR narrowband decoder
27 *
28 * This decoder uses floats for simplicity and so is not bit-exact. One
29 * difference is that differences in phase can accumulate. The test sequences
30 * in 3GPP TS 26.074 can still be useful.
31 *
32 * - Comparing this file's output to the output of the ref decoder gives a
33 * PSNR of 30 to 80. Plotting the output samples shows a difference in
34 * phase in some areas.
35 *
36 * - Comparing both decoders against their input, this decoder gives a similar
37 * PSNR. If the test sequence homing frames are removed (this decoder does
38 * not detect them), the PSNR is at least as good as the reference on 140
39 * out of 169 tests.
40 */
41
42
43 #include <string.h>
44 #include <math.h>
45
46 #include "libavutil/channel_layout.h"
47 #include "libavutil/float_dsp.h"
48 #include "avcodec.h"
49 #include "libavutil/common.h"
50 #include "libavutil/avassert.h"
51 #include "celp_math.h"
52 #include "celp_filters.h"
53 #include "acelp_filters.h"
54 #include "acelp_vectors.h"
55 #include "acelp_pitch_delay.h"
56 #include "lsp.h"
57 #include "amr.h"
58 #include "internal.h"
59
60 #include "amrnbdata.h"
61
62 #define AMR_BLOCK_SIZE 160 ///< samples per frame
63 #define AMR_SAMPLE_BOUND 32768.0 ///< threshold for synthesis overflow
64
65 /**
66 * Scale from constructed speech to [-1,1]
67 *
68 * AMR is designed to produce 16-bit PCM samples (3GPP TS 26.090 4.2) but
69 * upscales by two (section 6.2.2).
70 *
71 * Fundamentally, this scale is determined by energy_mean through
72 * the fixed vector contribution to the excitation vector.
73 */
74 #define AMR_SAMPLE_SCALE (2.0 / 32768.0)
75
76 /** Prediction factor for 12.2kbit/s mode */
77 #define PRED_FAC_MODE_12k2 0.65
78
79 #define LSF_R_FAC (8000.0 / 32768.0) ///< LSF residual tables to Hertz
80 #define MIN_LSF_SPACING (50.0488 / 8000.0) ///< Ensures stability of LPC filter
81 #define PITCH_LAG_MIN_MODE_12k2 18 ///< Lower bound on decoded lag search in 12.2kbit/s mode
82
83 /** Initial energy in dB. Also used for bad frames (unimplemented). */
84 #define MIN_ENERGY -14.0
85
86 /** Maximum sharpening factor
87 *
88 * The specification says 0.8, which should be 13107, but the reference C code
89 * uses 13017 instead. (Amusingly the same applies to SHARP_MAX in g729dec.c.)
90 */
91 #define SHARP_MAX 0.79449462890625
92
93 /** Number of impulse response coefficients used for tilt factor */
94 #define AMR_TILT_RESPONSE 22
95 /** Tilt factor = 1st reflection coefficient * gamma_t */
96 #define AMR_TILT_GAMMA_T 0.8
97 /** Adaptive gain control factor used in post-filter */
98 #define AMR_AGC_ALPHA 0.9
99
100 typedef struct AMRContext {
101 AMRNBFrame frame; ///< decoded AMR parameters (lsf coefficients, codebook indexes, etc)
102 uint8_t bad_frame_indicator; ///< bad frame ? 1 : 0
103 enum Mode cur_frame_mode;
104
105 int16_t prev_lsf_r[LP_FILTER_ORDER]; ///< residual LSF vector from previous subframe
106 double lsp[4][LP_FILTER_ORDER]; ///< lsp vectors from current frame
107 double prev_lsp_sub4[LP_FILTER_ORDER]; ///< lsp vector for the 4th subframe of the previous frame
108
109 float lsf_q[4][LP_FILTER_ORDER]; ///< Interpolated LSF vector for fixed gain smoothing
110 float lsf_avg[LP_FILTER_ORDER]; ///< vector of averaged lsf vector
111
112 float lpc[4][LP_FILTER_ORDER]; ///< lpc coefficient vectors for 4 subframes
113
114 uint8_t pitch_lag_int; ///< integer part of pitch lag from current subframe
115
116 float excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1 + AMR_SUBFRAME_SIZE]; ///< current excitation and all necessary excitation history
117 float *excitation; ///< pointer to the current excitation vector in excitation_buf
118
119 float pitch_vector[AMR_SUBFRAME_SIZE]; ///< adaptive code book (pitch) vector
120 float fixed_vector[AMR_SUBFRAME_SIZE]; ///< algebraic codebook (fixed) vector (must be kept zero between frames)
121
122 float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes
123 float pitch_gain[5]; ///< quantified pitch gains for the current and previous four subframes
124 float fixed_gain[5]; ///< quantified fixed gains for the current and previous four subframes
125
126 float beta; ///< previous pitch_gain, bounded by [0.0,SHARP_MAX]
127 uint8_t diff_count; ///< the number of subframes for which diff has been above 0.65
128 uint8_t hang_count; ///< the number of subframes since a hangover period started
129
130 float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness processing to determine "onset"
131 uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none
132 uint8_t ir_filter_onset; ///< flag for impulse response filter strength
133
134 float postfilter_mem[10]; ///< previous intermediate values in the formant filter
135 float tilt_mem; ///< previous input to tilt compensation filter
136 float postfilter_agc; ///< previous factor used for adaptive gain control
137 float high_pass_mem[2]; ///< previous intermediate values in the high-pass filter
138
139 float samples_in[LP_FILTER_ORDER + AMR_SUBFRAME_SIZE]; ///< floating point samples
140
141 ACELPFContext acelpf_ctx; ///< context for filters for ACELP-based codecs
142 ACELPVContext acelpv_ctx; ///< context for vector operations for ACELP-based codecs
143 CELPFContext celpf_ctx; ///< context for filters for CELP-based codecs
144 CELPMContext celpm_ctx; ///< context for fixed point math operations
145
146 } AMRContext;
147
148 typedef struct AMRChannelsContext {
149 AMRContext ch[2];
150 } AMRChannelsContext;
151
152 /** Double version of ff_weighted_vector_sumf() */
153 570 static void weighted_vector_sumd(double *out, const double *in_a,
154 const double *in_b, double weight_coeff_a,
155 double weight_coeff_b, int length)
156 {
157 int i;
158
159
2/2
✓ Branch 0 taken 5700 times.
✓ Branch 1 taken 570 times.
 6270 for (i = 0; i < length; i++)
160 5700 out[i] = weight_coeff_a * in_a[i]
161 5700 + weight_coeff_b * in_b[i];
162 570 }
163
164 20 static av_cold int amrnb_decode_init(AVCodecContext *avctx)
165 {
166 20 AMRChannelsContext *s = avctx->priv_data;
167 int i;
168
169
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 20 times.
 20 if (avctx->channels > 2) {
170 ✗ avpriv_report_missing_feature(avctx, ">2 channel AMR");
171 ✗ return AVERROR_PATCHWELCOME;
172 }
173
174
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 20 times.
 20 if (!avctx->channels) {
175 ✗ avctx->channels = 1;
176 ✗ avctx->channel_layout = AV_CH_LAYOUT_MONO;
177 }
178
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 20 times.
 20 if (!avctx->sample_rate)
179 ✗ avctx->sample_rate = 8000;
180 20 avctx->sample_fmt = AV_SAMPLE_FMT_FLTP;
181
182
2/2
✓ Branch 0 taken 20 times.
✓ Branch 1 taken 20 times.
 40 for (int ch = 0; ch < avctx->channels; ch++) {
183 20 AMRContext *p = &s->ch[ch];
184 // p->excitation always points to the same position in p->excitation_buf
185 20 p->excitation = &p->excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1];
186
187
2/2
✓ Branch 0 taken 200 times.
✓ Branch 1 taken 20 times.
 220 for (i = 0; i < LP_FILTER_ORDER; i++) {
188 200 p->prev_lsp_sub4[i] = lsp_sub4_init[i] * 1000 / (float)(1 << 15);
189 200 p->lsf_avg[i] = p->lsf_q[3][i] = lsp_avg_init[i] / (float)(1 << 15);
190 }
191
192
2/2
✓ Branch 0 taken 80 times.
✓ Branch 1 taken 20 times.
 100 for (i = 0; i < 4; i++)
193 80 p->prediction_error[i] = MIN_ENERGY;
194
195 20 ff_acelp_filter_init(&p->acelpf_ctx);
196 20 ff_acelp_vectors_init(&p->acelpv_ctx);
197 20 ff_celp_filter_init(&p->celpf_ctx);
198 20 ff_celp_math_init(&p->celpm_ctx);
199 }
200
201 20 return 0;
202 }
203
204
205 /**
206 * Unpack an RFC4867 speech frame into the AMR frame mode and parameters.
207 *
208 * The order of speech bits is specified by 3GPP TS 26.101.
209 *
210 * @param p the context
211 * @param buf pointer to the input buffer
212 * @param buf_size size of the input buffer
213 *
214 * @return the frame mode
215 */
216 2284 static enum Mode unpack_bitstream(AMRContext *p, const uint8_t *buf,
217 int buf_size)
218 {
219 enum Mode mode;
220
221 // Decode the first octet.
222 2284 mode = buf[0] >> 3 & 0x0F; // frame type
223 2284 p->bad_frame_indicator = (buf[0] & 0x4) != 0x4; // quality bit
224
225
2/4
✓ Branch 0 taken 2284 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2284 times.
 2284 if (mode >= N_MODES || buf_size < frame_sizes_nb[mode] + 1) {
226 ✗ return NO_DATA;
227 }
228
229
1/2
✓ Branch 0 taken 2284 times.
✗ Branch 1 not taken.
 2284 if (mode < MODE_DTX)
230 2284 ff_amr_bit_reorder((uint16_t *) &p->frame, sizeof(AMRNBFrame), buf + 1,
231 2284 amr_unpacking_bitmaps_per_mode[mode]);
232
233 2284 return mode;
234 }
235
236
237 /// @name AMR pitch LPC coefficient decoding functions
238 /// @{
239
240 /**
241 * Interpolate the LSF vector (used for fixed gain smoothing).
242 * The interpolation is done over all four subframes even in MODE_12k2.
243 *
244 * @param[in] ctx The Context
245 * @param[in,out] lsf_q LSFs in [0,1] for each subframe
246 * @param[in] lsf_new New LSFs in [0,1] for subframe 4
247 */
248 2284 static void interpolate_lsf(ACELPVContext *ctx, float lsf_q[4][LP_FILTER_ORDER], float *lsf_new)
249 {
250 int i;
251
252
2/2
✓ Branch 0 taken 9136 times.
✓ Branch 1 taken 2284 times.
 11420 for (i = 0; i < 4; i++)
253 9136 ctx->weighted_vector_sumf(lsf_q[i], lsf_q[3], lsf_new,
254 9136 0.25 * (3 - i), 0.25 * (i + 1),
255 LP_FILTER_ORDER);
256 2284 }
257
258 /**
259 * Decode a set of 5 split-matrix quantized lsf indexes into an lsp vector.
260 *
261 * @param p the context
262 * @param lsp output LSP vector
263 * @param lsf_no_r LSF vector without the residual vector added
264 * @param lsf_quantizer pointers to LSF dictionary tables
265 * @param quantizer_offset offset in tables
266 * @param sign for the 3 dictionary table
267 * @param update store data for computing the next frame's LSFs
268 */
269 570 static void lsf2lsp_for_mode12k2(AMRContext *p, double lsp[LP_FILTER_ORDER],
270 const float lsf_no_r[LP_FILTER_ORDER],
271 const int16_t *lsf_quantizer[5],
272 const int quantizer_offset,
273 const int sign, const int update)
274 {
275 int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector
276 float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector
277 int i;
278
279
2/2
✓ Branch 0 taken 2850 times.
✓ Branch 1 taken 570 times.
 3420 for (i = 0; i < LP_FILTER_ORDER >> 1; i++)
280 2850 memcpy(&lsf_r[i << 1], &lsf_quantizer[i][quantizer_offset],
281 2 * sizeof(*lsf_r));
282
283
2/2
✓ Branch 0 taken 266 times.
✓ Branch 1 taken 304 times.
 570 if (sign) {
284 266 lsf_r[4] *= -1;
285 266 lsf_r[5] *= -1;
286 }
287
288
2/2
✓ Branch 0 taken 285 times.
✓ Branch 1 taken 285 times.
 570 if (update)
289 285 memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r));
290
291
2/2
✓ Branch 0 taken 5700 times.
✓ Branch 1 taken 570 times.
 6270 for (i = 0; i < LP_FILTER_ORDER; i++)
292 5700 lsf_q[i] = lsf_r[i] * (LSF_R_FAC / 8000.0) + lsf_no_r[i] * (1.0 / 8000.0);
293
294 570 ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
295
296
2/2
✓ Branch 0 taken 285 times.
✓ Branch 1 taken 285 times.
 570 if (update)
297 285 interpolate_lsf(&p->acelpv_ctx, p->lsf_q, lsf_q);
298
299 570 ff_acelp_lsf2lspd(lsp, lsf_q, LP_FILTER_ORDER);
300 570 }
301
302 /**
303 * Decode a set of 5 split-matrix quantized lsf indexes into 2 lsp vectors.
304 *
305 * @param p pointer to the AMRContext
306 */
307 285 static void lsf2lsp_5(AMRContext *p)
308 {
309 285 const uint16_t *lsf_param = p->frame.lsf;
310 float lsf_no_r[LP_FILTER_ORDER]; // LSFs without the residual vector
311 const int16_t *lsf_quantizer[5];
312 int i;
313
314 285 lsf_quantizer[0] = lsf_5_1[lsf_param[0]];
315 285 lsf_quantizer[1] = lsf_5_2[lsf_param[1]];
316 285 lsf_quantizer[2] = lsf_5_3[lsf_param[2] >> 1];
317 285 lsf_quantizer[3] = lsf_5_4[lsf_param[3]];
318 285 lsf_quantizer[4] = lsf_5_5[lsf_param[4]];
319
320
2/2
✓ Branch 0 taken 2850 times.
✓ Branch 1 taken 285 times.
 3135 for (i = 0; i < LP_FILTER_ORDER; i++)
321 2850 lsf_no_r[i] = p->prev_lsf_r[i] * LSF_R_FAC * PRED_FAC_MODE_12k2 + lsf_5_mean[i];
322
323 285 lsf2lsp_for_mode12k2(p, p->lsp[1], lsf_no_r, lsf_quantizer, 0, lsf_param[2] & 1, 0);
324 285 lsf2lsp_for_mode12k2(p, p->lsp[3], lsf_no_r, lsf_quantizer, 2, lsf_param[2] & 1, 1);
325
326 // interpolate LSP vectors at subframes 1 and 3
327 285 weighted_vector_sumd(p->lsp[0], p->prev_lsp_sub4, p->lsp[1], 0.5, 0.5, LP_FILTER_ORDER);
328 285 weighted_vector_sumd(p->lsp[2], p->lsp[1] , p->lsp[3], 0.5, 0.5, LP_FILTER_ORDER);
329 285 }
330
331 /**
332 * Decode a set of 3 split-matrix quantized lsf indexes into an lsp vector.
333 *
334 * @param p pointer to the AMRContext
335 */
336 1999 static void lsf2lsp_3(AMRContext *p)
337 {
338 1999 const uint16_t *lsf_param = p->frame.lsf;
339 int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector
340 float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector
341 const int16_t *lsf_quantizer;
342 int i, j;
343
344
2/2
✓ Branch 0 taken 285 times.
✓ Branch 1 taken 1714 times.
 1999 lsf_quantizer = (p->cur_frame_mode == MODE_7k95 ? lsf_3_1_MODE_7k95 : lsf_3_1)[lsf_param[0]];
345 1999 memcpy(lsf_r, lsf_quantizer, 3 * sizeof(*lsf_r));
346
347 1999 lsf_quantizer = lsf_3_2[lsf_param[1] << (p->cur_frame_mode <= MODE_5k15)];
348 1999 memcpy(lsf_r + 3, lsf_quantizer, 3 * sizeof(*lsf_r));
349
350
2/2
✓ Branch 0 taken 572 times.
✓ Branch 1 taken 1427 times.
 1999 lsf_quantizer = (p->cur_frame_mode <= MODE_5k15 ? lsf_3_3_MODE_5k15 : lsf_3_3)[lsf_param[2]];
351 1999 memcpy(lsf_r + 6, lsf_quantizer, 4 * sizeof(*lsf_r));
352
353 // calculate mean-removed LSF vector and add mean
354
2/2
✓ Branch 0 taken 19990 times.
✓ Branch 1 taken 1999 times.
 21989 for (i = 0; i < LP_FILTER_ORDER; i++)
355 19990 lsf_q[i] = (lsf_r[i] + p->prev_lsf_r[i] * pred_fac[i]) * (LSF_R_FAC / 8000.0) + lsf_3_mean[i] * (1.0 / 8000.0);
356
357 1999 ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
358
359 // store data for computing the next frame's LSFs
360 1999 interpolate_lsf(&p->acelpv_ctx, p->lsf_q, lsf_q);
361 1999 memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r));
362
363 1999 ff_acelp_lsf2lspd(p->lsp[3], lsf_q, LP_FILTER_ORDER);
364
365 // interpolate LSP vectors at subframes 1, 2 and 3
366
2/2
✓ Branch 0 taken 5997 times.
✓ Branch 1 taken 1999 times.
 7996 for (i = 1; i <= 3; i++)
367
2/2
✓ Branch 0 taken 59970 times.
✓ Branch 1 taken 5997 times.
 65967 for(j = 0; j < LP_FILTER_ORDER; j++)
368 59970 p->lsp[i-1][j] = p->prev_lsp_sub4[j] +
369 59970 (p->lsp[3][j] - p->prev_lsp_sub4[j]) * 0.25 * i;
370 1999 }
371
372 /// @}
373
374
375 /// @name AMR pitch vector decoding functions
376 /// @{
377
378 /**
379 * Like ff_decode_pitch_lag(), but with 1/6 resolution
380 */
381 1140 static void decode_pitch_lag_1_6(int *lag_int, int *lag_frac, int pitch_index,
382 const int prev_lag_int, const int subframe)
383 {
384
4/4
✓ Branch 0 taken 855 times.
✓ Branch 1 taken 285 times.
✓ Branch 2 taken 285 times.
✓ Branch 3 taken 570 times.
 1140 if (subframe == 0 || subframe == 2) {
385
2/2
✓ Branch 0 taken 459 times.
✓ Branch 1 taken 111 times.
 570 if (pitch_index < 463) {
386 459 *lag_int = (pitch_index + 107) * 10923 >> 16;
387 459 *lag_frac = pitch_index - *lag_int * 6 + 105;
388 } else {
389 111 *lag_int = pitch_index - 368;
390 111 *lag_frac = 0;
391 }
392 } else {
393 570 *lag_int = ((pitch_index + 5) * 10923 >> 16) - 1;
394 570 *lag_frac = pitch_index - *lag_int * 6 - 3;
395 570 *lag_int += av_clip(prev_lag_int - 5, PITCH_LAG_MIN_MODE_12k2,
396 PITCH_DELAY_MAX - 9);
397 }
398 1140 }
399
400 9136 static void decode_pitch_vector(AMRContext *p,
401 const AMRNBSubframe *amr_subframe,
402 const int subframe)
403 {
404 int pitch_lag_int, pitch_lag_frac;
405 9136 enum Mode mode = p->cur_frame_mode;
406
407
2/2
✓ Branch 0 taken 1140 times.
✓ Branch 1 taken 7996 times.
 9136 if (p->cur_frame_mode == MODE_12k2) {
408 1140 decode_pitch_lag_1_6(&pitch_lag_int, &pitch_lag_frac,
409 1140 amr_subframe->p_lag, p->pitch_lag_int,
410 subframe);
411 } else {
412
6/6
✓ Branch 0 taken 3428 times.
✓ Branch 1 taken 4568 times.
✓ Branch 2 taken 6848 times.
✓ Branch 3 taken 1148 times.
✓ Branch 4 taken 5708 times.
✓ Branch 5 taken 1140 times.
 11424 ff_decode_pitch_lag(&pitch_lag_int, &pitch_lag_frac,
413 7996 amr_subframe->p_lag,
414 7996 p->pitch_lag_int, subframe,
415 mode != MODE_4k75 && mode != MODE_5k15,
416
2/2
✓ Branch 0 taken 1140 times.
✓ Branch 1 taken 2288 times.
 3428 mode <= MODE_6k7 ? 4 : (mode == MODE_7k95 ? 5 : 6));
417 7996 pitch_lag_frac *= 2;
418 }
419
420 9136 p->pitch_lag_int = pitch_lag_int; // store previous lag in a uint8_t
421
422 9136 pitch_lag_int += pitch_lag_frac > 0;
423
424 /* Calculate the pitch vector by interpolating the past excitation at the
425 pitch lag using a b60 hamming windowed sinc function. */
426 18272 p->acelpf_ctx.acelp_interpolatef(p->excitation,
427 9136 p->excitation + 1 - pitch_lag_int,
428 ff_b60_sinc, 6,
429
2/2
✓ Branch 0 taken 2124 times.
✓ Branch 1 taken 7012 times.
 9136 pitch_lag_frac + 6 - 6*(pitch_lag_frac > 0),
430 10, AMR_SUBFRAME_SIZE);
431
432 9136 memcpy(p->pitch_vector, p->excitation, AMR_SUBFRAME_SIZE * sizeof(float));
433 9136 }
434
435 /// @}
436
437
438 /// @name AMR algebraic code book (fixed) vector decoding functions
439 /// @{
440
441 /**
442 * Decode a 10-bit algebraic codebook index from a 10.2 kbit/s frame.
443 */
444 2296 static void decode_10bit_pulse(int code, int pulse_position[8],
445 int i1, int i2, int i3)
446 {
447 // coded using 7+3 bits with the 3 LSBs being, individually, the LSB of 1 of
448 // the 3 pulses and the upper 7 bits being coded in base 5
449 2296 const uint8_t *positions = base_five_table[code >> 3];
450 2296 pulse_position[i1] = (positions[2] << 1) + ( code & 1);
451 2296 pulse_position[i2] = (positions[1] << 1) + ((code >> 1) & 1);
452 2296 pulse_position[i3] = (positions[0] << 1) + ((code >> 2) & 1);
453 2296 }
454
455 /**
456 * Decode the algebraic codebook index to pulse positions and signs and
457 * construct the algebraic codebook vector for MODE_10k2.
458 *
459 * @param fixed_index positions of the eight pulses
460 * @param fixed_sparse pointer to the algebraic codebook vector
461 */
462 1148 static void decode_8_pulses_31bits(const int16_t *fixed_index,
463 AMRFixed *fixed_sparse)
464 {
465 int pulse_position[8];
466 int i, temp;
467
468 1148 decode_10bit_pulse(fixed_index[4], pulse_position, 0, 4, 1);
469 1148 decode_10bit_pulse(fixed_index[5], pulse_position, 2, 6, 5);
470
471 // coded using 5+2 bits with the 2 LSBs being, individually, the LSB of 1 of
472 // the 2 pulses and the upper 5 bits being coded in base 5
473 1148 temp = ((fixed_index[6] >> 2) * 25 + 12) >> 5;
474 1148 pulse_position[3] = temp % 5;
475 1148 pulse_position[7] = temp / 5;
476
2/2
✓ Branch 0 taken 480 times.
✓ Branch 1 taken 668 times.
 1148 if (pulse_position[7] & 1)
477 480 pulse_position[3] = 4 - pulse_position[3];
478 1148 pulse_position[3] = (pulse_position[3] << 1) + ( fixed_index[6] & 1);
479 1148 pulse_position[7] = (pulse_position[7] << 1) + ((fixed_index[6] >> 1) & 1);
480
481 1148 fixed_sparse->n = 8;
482
2/2
✓ Branch 0 taken 4592 times.
✓ Branch 1 taken 1148 times.
 5740 for (i = 0; i < 4; i++) {
483 4592 const int pos1 = (pulse_position[i] << 2) + i;
484 4592 const int pos2 = (pulse_position[i + 4] << 2) + i;
485
2/2
✓ Branch 0 taken 2267 times.
✓ Branch 1 taken 2325 times.
 4592 const float sign = fixed_index[i] ? -1.0 : 1.0;
486 4592 fixed_sparse->x[i ] = pos1;
487 4592 fixed_sparse->x[i + 4] = pos2;
488 4592 fixed_sparse->y[i ] = sign;
489
2/2
✓ Branch 0 taken 2154 times.
✓ Branch 1 taken 2438 times.
 4592 fixed_sparse->y[i + 4] = pos2 < pos1 ? -sign : sign;
490 }
491 1148 }
492
493 /**
494 * Decode the algebraic codebook index to pulse positions and signs,
495 * then construct the algebraic codebook vector.
496 *
497 * nb of pulses | bits encoding pulses
498 * For MODE_4k75 or MODE_5k15, 2 | 1-3, 4-6, 7
499 * MODE_5k9, 2 | 1, 2-4, 5-6, 7-9
500 * MODE_6k7, 3 | 1-3, 4, 5-7, 8, 9-11
501 * MODE_7k4 or MODE_7k95, 4 | 1-3, 4-6, 7-9, 10, 11-13
502 *
503 * @param fixed_sparse pointer to the algebraic codebook vector
504 * @param pulses algebraic codebook indexes
505 * @param mode mode of the current frame
506 * @param subframe current subframe number
507 */
508 9136 static void decode_fixed_sparse(AMRFixed *fixed_sparse, const uint16_t *pulses,
509 const enum Mode mode, const int subframe)
510 {
511 av_assert1(MODE_4k75 <= (signed)mode && mode <= MODE_12k2);
512
513
2/2
✓ Branch 0 taken 1140 times.
✓ Branch 1 taken 7996 times.
 9136 if (mode == MODE_12k2) {
514 1140 ff_decode_10_pulses_35bits(pulses, fixed_sparse, gray_decode, 5, 3);
515
2/2
✓ Branch 0 taken 1148 times.
✓ Branch 1 taken 6848 times.
 7996 } else if (mode == MODE_10k2) {
516 1148 decode_8_pulses_31bits(pulses, fixed_sparse);
517 } else {
518 6848 int *pulse_position = fixed_sparse->x;
519 int i, pulse_subset;
520 6848 const int fixed_index = pulses[0];
521
522
2/2
✓ Branch 0 taken 2288 times.
✓ Branch 1 taken 4560 times.
 6848 if (mode <= MODE_5k15) {
523 2288 pulse_subset = ((fixed_index >> 3) & 8) + (subframe << 1);
524 2288 pulse_position[0] = ( fixed_index & 7) * 5 + track_position[pulse_subset];
525 2288 pulse_position[1] = ((fixed_index >> 3) & 7) * 5 + track_position[pulse_subset + 1];
526 2288 fixed_sparse->n = 2;
527
2/2
✓ Branch 0 taken 1140 times.
✓ Branch 1 taken 3420 times.
 4560 } else if (mode == MODE_5k9) {
528 1140 pulse_subset = ((fixed_index & 1) << 1) + 1;
529 1140 pulse_position[0] = ((fixed_index >> 1) & 7) * 5 + pulse_subset;
530 1140 pulse_subset = (fixed_index >> 4) & 3;
531 1140 pulse_position[1] = ((fixed_index >> 6) & 7) * 5 + pulse_subset + (pulse_subset == 3 ? 1 : 0);
532
2/2
✓ Branch 0 taken 1 times.
✓ Branch 1 taken 1139 times.
 1140 fixed_sparse->n = pulse_position[0] == pulse_position[1] ? 1 : 2;
533
2/2
✓ Branch 0 taken 1140 times.
✓ Branch 1 taken 2280 times.
 3420 } else if (mode == MODE_6k7) {
534 1140 pulse_position[0] = (fixed_index & 7) * 5;
535 1140 pulse_subset = (fixed_index >> 2) & 2;
536 1140 pulse_position[1] = ((fixed_index >> 4) & 7) * 5 + pulse_subset + 1;
537 1140 pulse_subset = (fixed_index >> 6) & 2;
538 1140 pulse_position[2] = ((fixed_index >> 8) & 7) * 5 + pulse_subset + 2;
539 1140 fixed_sparse->n = 3;
540 } else { // mode <= MODE_7k95
541 2280 pulse_position[0] = gray_decode[ fixed_index & 7];
542 2280 pulse_position[1] = gray_decode[(fixed_index >> 3) & 7] + 1;
543 2280 pulse_position[2] = gray_decode[(fixed_index >> 6) & 7] + 2;
544 2280 pulse_subset = (fixed_index >> 9) & 1;
545 2280 pulse_position[3] = gray_decode[(fixed_index >> 10) & 7] + pulse_subset + 3;
546 2280 fixed_sparse->n = 4;
547 }
548
2/2
✓ Branch 0 taken 19395 times.
✓ Branch 1 taken 6848 times.
 26243 for (i = 0; i < fixed_sparse->n; i++)
549
2/2
✓ Branch 0 taken 9758 times.
✓ Branch 1 taken 9637 times.
 19395 fixed_sparse->y[i] = (pulses[1] >> i) & 1 ? 1.0 : -1.0;
550 }
551 9136 }
552
553 /**
554 * Apply pitch lag to obtain the sharpened fixed vector (section 6.1.2)
555 *
556 * @param p the context
557 * @param subframe unpacked amr subframe
558 * @param mode mode of the current frame
559 * @param fixed_sparse sparse representation of the fixed vector
560 */
561 9136 static void pitch_sharpening(AMRContext *p, int subframe, enum Mode mode,
562 AMRFixed *fixed_sparse)
563 {
564 // The spec suggests the current pitch gain is always used, but in other
565 // modes the pitch and codebook gains are jointly quantized (sec 5.8.2)
566 // so the codebook gain cannot depend on the quantized pitch gain.
567
2/2
✓ Branch 0 taken 1140 times.
✓ Branch 1 taken 7996 times.
 9136 if (mode == MODE_12k2)
568
2/2
✓ Branch 0 taken 106 times.
✓ Branch 1 taken 1034 times.
 1140 p->beta = FFMIN(p->pitch_gain[4], 1.0);
569
570 9136 fixed_sparse->pitch_lag = p->pitch_lag_int;
571 9136 fixed_sparse->pitch_fac = p->beta;
572
573 // Save pitch sharpening factor for the next subframe
574 // MODE_4k75 only updates on the 2nd and 4th subframes - this follows from
575 // the fact that the gains for two subframes are jointly quantized.
576
4/4
✓ Branch 0 taken 1148 times.
✓ Branch 1 taken 7988 times.
✓ Branch 2 taken 574 times.
✓ Branch 3 taken 574 times.
 9136 if (mode != MODE_4k75 || subframe & 1)
577 8562 p->beta = av_clipf(p->pitch_gain[4], 0.0, SHARP_MAX);
578 9136 }
579 /// @}
580
581
582 /// @name AMR gain decoding functions
583 /// @{
584
585 /**
586 * fixed gain smoothing
587 * Note that where the spec specifies the "spectrum in the q domain"
588 * in section 6.1.4, in fact frequencies should be used.
589 *
590 * @param p the context
591 * @param lsf LSFs for the current subframe, in the range [0,1]
592 * @param lsf_avg averaged LSFs
593 * @param mode mode of the current frame
594 *
595 * @return fixed gain smoothed
596 */
597 9136 static float fixed_gain_smooth(AMRContext *p , const float *lsf,
598 const float *lsf_avg, const enum Mode mode)
599 {
600 9136 float diff = 0.0;
601 int i;
602
603
2/2
✓ Branch 0 taken 91360 times.
✓ Branch 1 taken 9136 times.
 100496 for (i = 0; i < LP_FILTER_ORDER; i++)
604 91360 diff += fabs(lsf_avg[i] - lsf[i]) / lsf_avg[i];
605
606 // If diff is large for ten subframes, disable smoothing for a 40-subframe
607 // hangover period.
608 9136 p->diff_count++;
609
2/2
✓ Branch 0 taken 6733 times.
✓ Branch 1 taken 2403 times.
 9136 if (diff <= 0.65)
610 6733 p->diff_count = 0;
611
612
2/2
✓ Branch 0 taken 687 times.
✓ Branch 1 taken 8449 times.
 9136 if (p->diff_count > 10) {
613 687 p->hang_count = 0;
614 687 p->diff_count--; // don't let diff_count overflow
615 }
616
617
2/2
✓ Branch 0 taken 3851 times.
✓ Branch 1 taken 5285 times.
 9136 if (p->hang_count < 40) {
618 3851 p->hang_count++;
619
4/4
✓ Branch 0 taken 2539 times.
✓ Branch 1 taken 2746 times.
✓ Branch 2 taken 667 times.
✓ Branch 3 taken 1872 times.
 5285 } else if (mode < MODE_7k4 || mode == MODE_10k2) {
620 3413 const float smoothing_factor = av_clipf(4.0 * diff - 1.6, 0.0, 1.0);
621 3413 const float fixed_gain_mean = (p->fixed_gain[0] + p->fixed_gain[1] +
622 3413 p->fixed_gain[2] + p->fixed_gain[3] +
623 3413 p->fixed_gain[4]) * 0.2;
624 3413 return smoothing_factor * p->fixed_gain[4] +
625 3413 (1.0 - smoothing_factor) * fixed_gain_mean;
626 }
627 5723 return p->fixed_gain[4];
628 }
629
630 /**
631 * Decode pitch gain and fixed gain factor (part of section 6.1.3).
632 *
633 * @param p the context
634 * @param amr_subframe unpacked amr subframe
635 * @param mode mode of the current frame
636 * @param subframe current subframe number
637 * @param fixed_gain_factor decoded gain correction factor
638 */
639 9136 static void decode_gains(AMRContext *p, const AMRNBSubframe *amr_subframe,
640 const enum Mode mode, const int subframe,
641 float *fixed_gain_factor)
642 {
643
4/4
✓ Branch 0 taken 7996 times.
✓ Branch 1 taken 1140 times.
✓ Branch 2 taken 1140 times.
✓ Branch 3 taken 6856 times.
 9136 if (mode == MODE_12k2 || mode == MODE_7k95) {
644 2280 p->pitch_gain[4] = qua_gain_pit [amr_subframe->p_gain ]
645 2280 * (1.0 / 16384.0);
646 2280 *fixed_gain_factor = qua_gain_code[amr_subframe->fixed_gain]
647 2280 * (1.0 / 2048.0);
648 } else {
649 const uint16_t *gains;
650
651
2/2
✓ Branch 0 taken 3428 times.
✓ Branch 1 taken 3428 times.
 6856 if (mode >= MODE_6k7) {
652 3428 gains = gains_high[amr_subframe->p_gain];
653
2/2
✓ Branch 0 taken 2280 times.
✓ Branch 1 taken 1148 times.
 3428 } else if (mode >= MODE_5k15) {
654 2280 gains = gains_low [amr_subframe->p_gain];
655 } else {
656 // gain index is only coded in subframes 0,2 for MODE_4k75
657 1148 gains = gains_MODE_4k75[(p->frame.subframe[subframe & 2].p_gain << 1) + (subframe & 1)];
658 }
659
660 6856 p->pitch_gain[4] = gains[0] * (1.0 / 16384.0);
661 6856 *fixed_gain_factor = gains[1] * (1.0 / 4096.0);
662 }
663 9136 }
664
665 /// @}
666
667
668 /// @name AMR preprocessing functions
669 /// @{
670
671 /**
672 * Circularly convolve a sparse fixed vector with a phase dispersion impulse
673 * response filter (D.6.2 of G.729 and 6.1.5 of AMR).
674 *
675 * @param out vector with filter applied
676 * @param in source vector
677 * @param filter phase filter coefficients
678 *
679 * out[n] = sum(i,0,len-1){ in[i] * filter[(len + n - i)%len] }
680 */
681 4869 static void apply_ir_filter(float *out, const AMRFixed *in,
682 const float *filter)
683 {
684 float filter1[AMR_SUBFRAME_SIZE], ///< filters at pitch lag*1 and *2
685 filter2[AMR_SUBFRAME_SIZE];
686 4869 int lag = in->pitch_lag;
687 4869 float fac = in->pitch_fac;
688 int i;
689
690
2/2
✓ Branch 0 taken 741 times.
✓ Branch 1 taken 4128 times.
 4869 if (lag < AMR_SUBFRAME_SIZE) {
691 741 ff_celp_circ_addf(filter1, filter, filter, lag, fac,
692 AMR_SUBFRAME_SIZE);
693
694
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 741 times.
 741 if (lag < AMR_SUBFRAME_SIZE >> 1)
695 ✗ ff_celp_circ_addf(filter2, filter, filter1, lag, fac,
696 AMR_SUBFRAME_SIZE);
697 }
698
699 4869 memset(out, 0, sizeof(float) * AMR_SUBFRAME_SIZE);
700
2/2
✓ Branch 0 taken 12807 times.
✓ Branch 1 taken 4869 times.
 17676 for (i = 0; i < in->n; i++) {
701 12807 int x = in->x[i];
702 12807 float y = in->y[i];
703 const float *filterp;
704
705
2/2
✓ Branch 0 taken 12481 times.
✓ Branch 1 taken 326 times.
 12807 if (x >= AMR_SUBFRAME_SIZE - lag) {
706 12481 filterp = filter;
707
1/2
✓ Branch 0 taken 326 times.
✗ Branch 1 not taken.
 326 } else if (x >= AMR_SUBFRAME_SIZE - (lag << 1)) {
708 326 filterp = filter1;
709 } else
710 ✗ filterp = filter2;
711
712 12807 ff_celp_circ_addf(out, out, filterp, x, y, AMR_SUBFRAME_SIZE);
713 }
714 4869 }
715
716 /**
717 * Reduce fixed vector sparseness by smoothing with one of three IR filters.
718 * Also know as "adaptive phase dispersion".
719 *
720 * This implements 3GPP TS 26.090 section 6.1(5).
721 *
722 * @param p the context
723 * @param fixed_sparse algebraic codebook vector
724 * @param fixed_vector unfiltered fixed vector
725 * @param fixed_gain smoothed gain
726 * @param out space for modified vector if necessary
727 */
728 9136 static const float *anti_sparseness(AMRContext *p, AMRFixed *fixed_sparse,
729 const float *fixed_vector,
730 float fixed_gain, float *out)
731 {
732 int ir_filter_nr;
733
734
2/2
✓ Branch 0 taken 4165 times.
✓ Branch 1 taken 4971 times.
 9136 if (p->pitch_gain[4] < 0.6) {
735 4165 ir_filter_nr = 0; // strong filtering
736
2/2
✓ Branch 0 taken 2995 times.
✓ Branch 1 taken 1976 times.
 4971 } else if (p->pitch_gain[4] < 0.9) {
737 2995 ir_filter_nr = 1; // medium filtering
738 } else
739 1976 ir_filter_nr = 2; // no filtering
740
741 // detect 'onset'
742
2/2
✓ Branch 0 taken 232 times.
✓ Branch 1 taken 8904 times.
 9136 if (fixed_gain > 2.0 * p->prev_sparse_fixed_gain) {
743 232 p->ir_filter_onset = 2;
744
2/2
✓ Branch 0 taken 336 times.
✓ Branch 1 taken 8568 times.
 8904 } else if (p->ir_filter_onset)
745 336 p->ir_filter_onset--;
746
747
2/2
✓ Branch 0 taken 8732 times.
✓ Branch 1 taken 404 times.
 9136 if (!p->ir_filter_onset) {
748 8732 int i, count = 0;
749
750
2/2
✓ Branch 0 taken 43660 times.
✓ Branch 1 taken 8732 times.
 52392 for (i = 0; i < 5; i++)
751
2/2
✓ Branch 0 taken 19971 times.
✓ Branch 1 taken 23689 times.
 43660 if (p->pitch_gain[i] < 0.6)
752 19971 count++;
753
2/2
✓ Branch 0 taken 3912 times.
✓ Branch 1 taken 4820 times.
 8732 if (count > 2)
754 3912 ir_filter_nr = 0;
755
756
2/2
✓ Branch 0 taken 489 times.
✓ Branch 1 taken 8243 times.
 8732 if (ir_filter_nr > p->prev_ir_filter_nr + 1)
757 489 ir_filter_nr--;
758
2/2
✓ Branch 0 taken 262 times.
✓ Branch 1 taken 142 times.
 404 } else if (ir_filter_nr < 2)
759 262 ir_filter_nr++;
760
761 // Disable filtering for very low level of fixed_gain.
762 // Note this step is not specified in the technical description but is in
763 // the reference source in the function Ph_disp.
764
2/2
✓ Branch 0 taken 16 times.
✓ Branch 1 taken 9120 times.
 9136 if (fixed_gain < 5.0)
765 16 ir_filter_nr = 2;
766
767
4/4
✓ Branch 0 taken 7996 times.
✓ Branch 1 taken 1140 times.
✓ Branch 2 taken 5708 times.
✓ Branch 3 taken 2288 times.
 9136 if (p->cur_frame_mode != MODE_7k4 && p->cur_frame_mode < MODE_10k2
768
2/2
✓ Branch 0 taken 4869 times.
✓ Branch 1 taken 839 times.
 5708 && ir_filter_nr < 2) {
769 4869 apply_ir_filter(out, fixed_sparse,
770 4869 (p->cur_frame_mode == MODE_7k95 ?
771
2/2
✓ Branch 0 taken 1038 times.
✓ Branch 1 taken 3831 times.
 4869 ir_filters_lookup_MODE_7k95 :
772 4869 ir_filters_lookup)[ir_filter_nr]);
773 4869 fixed_vector = out;
774 }
775
776 // update ir filter strength history
777 9136 p->prev_ir_filter_nr = ir_filter_nr;
778 9136 p->prev_sparse_fixed_gain = fixed_gain;
779
780 9136 return fixed_vector;
781 }
782
783 /// @}
784
785
786 /// @name AMR synthesis functions
787 /// @{
788
789 /**
790 * Conduct 10th order linear predictive coding synthesis.
791 *
792 * @param p pointer to the AMRContext
793 * @param lpc pointer to the LPC coefficients
794 * @param fixed_gain fixed codebook gain for synthesis
795 * @param fixed_vector algebraic codebook vector
796 * @param samples pointer to the output speech samples
797 * @param overflow 16-bit overflow flag
798 */
799 9136 static int synthesis(AMRContext *p, float *lpc,
800 float fixed_gain, const float *fixed_vector,
801 float *samples, uint8_t overflow)
802 {
803 int i;
804 float excitation[AMR_SUBFRAME_SIZE];
805
806 // if an overflow has been detected, the pitch vector is scaled down by a
807 // factor of 4
808
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 9136 times.
 9136 if (overflow)
809 ✗ for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
810 ✗ p->pitch_vector[i] *= 0.25;
811
812 9136 p->acelpv_ctx.weighted_vector_sumf(excitation, p->pitch_vector, fixed_vector,
813 p->pitch_gain[4], fixed_gain, AMR_SUBFRAME_SIZE);
814
815 // emphasize pitch vector contribution
816
3/4
✓ Branch 0 taken 6236 times.
✓ Branch 1 taken 2900 times.
✓ Branch 2 taken 6236 times.
✗ Branch 3 not taken.
 9136 if (p->pitch_gain[4] > 0.5 && !overflow) {
817 6236 float energy = p->celpm_ctx.dot_productf(excitation, excitation,
818 AMR_SUBFRAME_SIZE);
819 6236 float pitch_factor =
820 12472 p->pitch_gain[4] *
821 6236 (p->cur_frame_mode == MODE_12k2 ?
822
4/4
✓ Branch 0 taken 704 times.
✓ Branch 1 taken 5532 times.
✓ Branch 2 taken 106 times.
✓ Branch 3 taken 598 times.
 11768 0.25 * FFMIN(p->pitch_gain[4], 1.0) :
823
2/2
✓ Branch 0 taken 2748 times.
✓ Branch 1 taken 2784 times.
 5532 0.5 * FFMIN(p->pitch_gain[4], SHARP_MAX));
824
825
2/2
✓ Branch 0 taken 249440 times.
✓ Branch 1 taken 6236 times.
 255676 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
826 249440 excitation[i] += pitch_factor * p->pitch_vector[i];
827
828 6236 ff_scale_vector_to_given_sum_of_squares(excitation, excitation, energy,
829 AMR_SUBFRAME_SIZE);
830 }
831
832 9136 p->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation,
833 AMR_SUBFRAME_SIZE,
834 LP_FILTER_ORDER);
835
836 // detect overflow
837
2/2
✓ Branch 0 taken 365440 times.
✓ Branch 1 taken 9136 times.
 374576 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
838
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 365440 times.
 365440 if (fabsf(samples[i]) > AMR_SAMPLE_BOUND) {
839 ✗ return 1;
840 }
841
842 9136 return 0;
843 }
844
845 /// @}
846
847
848 /// @name AMR update functions
849 /// @{
850
851 /**
852 * Update buffers and history at the end of decoding a subframe.
853 *
854 * @param p pointer to the AMRContext
855 */
856 9136 static void update_state(AMRContext *p)
857 {
858 9136 memcpy(p->prev_lsp_sub4, p->lsp[3], LP_FILTER_ORDER * sizeof(p->lsp[3][0]));
859
860 9136 memmove(&p->excitation_buf[0], &p->excitation_buf[AMR_SUBFRAME_SIZE],
861 (PITCH_DELAY_MAX + LP_FILTER_ORDER + 1) * sizeof(float));
862
863 9136 memmove(&p->pitch_gain[0], &p->pitch_gain[1], 4 * sizeof(float));
864 9136 memmove(&p->fixed_gain[0], &p->fixed_gain[1], 4 * sizeof(float));
865
866 9136 memmove(&p->samples_in[0], &p->samples_in[AMR_SUBFRAME_SIZE],
867 LP_FILTER_ORDER * sizeof(float));
868 9136 }
869
870 /// @}
871
872
873 /// @name AMR Postprocessing functions
874 /// @{
875
876 /**
877 * Get the tilt factor of a formant filter from its transfer function
878 *
879 * @param p The Context
880 * @param lpc_n LP_FILTER_ORDER coefficients of the numerator
881 * @param lpc_d LP_FILTER_ORDER coefficients of the denominator
882 */
883 9136 static float tilt_factor(AMRContext *p, float *lpc_n, float *lpc_d)
884 {
885 float rh0, rh1; // autocorrelation at lag 0 and 1
886
887 // LP_FILTER_ORDER prior zeros are needed for ff_celp_lp_synthesis_filterf
888 9136 float impulse_buffer[LP_FILTER_ORDER + AMR_TILT_RESPONSE] = { 0 };
889 9136 float *hf = impulse_buffer + LP_FILTER_ORDER; // start of impulse response
890
891 9136 hf[0] = 1.0;
892 9136 memcpy(hf + 1, lpc_n, sizeof(float) * LP_FILTER_ORDER);
893 9136 p->celpf_ctx.celp_lp_synthesis_filterf(hf, lpc_d, hf,
894 AMR_TILT_RESPONSE,
895 LP_FILTER_ORDER);
896
897 9136 rh0 = p->celpm_ctx.dot_productf(hf, hf, AMR_TILT_RESPONSE);
898 9136 rh1 = p->celpm_ctx.dot_productf(hf, hf + 1, AMR_TILT_RESPONSE - 1);
899
900 // The spec only specifies this check for 12.2 and 10.2 kbit/s
901 // modes. But in the ref source the tilt is always non-negative.
902
1/2
✓ Branch 0 taken 9136 times.
✗ Branch 1 not taken.
 9136 return rh1 >= 0.0 ? rh1 / rh0 * AMR_TILT_GAMMA_T : 0.0;
903 }
904
905 /**
906 * Perform adaptive post-filtering to enhance the quality of the speech.
907 * See section 6.2.1.
908 *
909 * @param p pointer to the AMRContext
910 * @param lpc interpolated LP coefficients for this subframe
911 * @param buf_out output of the filter
912 */
913 9136 static void postfilter(AMRContext *p, float *lpc, float *buf_out)
914 {
915 int i;
916 9136 float *samples = p->samples_in + LP_FILTER_ORDER; // Start of input
917
918 9136 float speech_gain = p->celpm_ctx.dot_productf(samples, samples,
919 AMR_SUBFRAME_SIZE);
920
921 float pole_out[AMR_SUBFRAME_SIZE + LP_FILTER_ORDER]; // Output of pole filter
922 const float *gamma_n, *gamma_d; // Formant filter factor table
923 float lpc_n[LP_FILTER_ORDER], lpc_d[LP_FILTER_ORDER]; // Transfer function coefficients
924
925
4/4
✓ Branch 0 taken 7996 times.
✓ Branch 1 taken 1140 times.
✓ Branch 2 taken 1148 times.
✓ Branch 3 taken 6848 times.
 9136 if (p->cur_frame_mode == MODE_12k2 || p->cur_frame_mode == MODE_10k2) {
926 2288 gamma_n = ff_pow_0_7;
927 2288 gamma_d = ff_pow_0_75;
928 } else {
929 6848 gamma_n = ff_pow_0_55;
930 6848 gamma_d = ff_pow_0_7;
931 }
932
933
2/2
✓ Branch 0 taken 91360 times.
✓ Branch 1 taken 9136 times.
 100496 for (i = 0; i < LP_FILTER_ORDER; i++) {
934 91360 lpc_n[i] = lpc[i] * gamma_n[i];
935 91360 lpc_d[i] = lpc[i] * gamma_d[i];
936 }
937
938 9136 memcpy(pole_out, p->postfilter_mem, sizeof(float) * LP_FILTER_ORDER);
939 9136 p->celpf_ctx.celp_lp_synthesis_filterf(pole_out + LP_FILTER_ORDER, lpc_d, samples,
940 AMR_SUBFRAME_SIZE, LP_FILTER_ORDER);
941 9136 memcpy(p->postfilter_mem, pole_out + AMR_SUBFRAME_SIZE,
942 sizeof(float) * LP_FILTER_ORDER);
943
944 9136 p->celpf_ctx.celp_lp_zero_synthesis_filterf(buf_out, lpc_n,
945 pole_out + LP_FILTER_ORDER,
946 AMR_SUBFRAME_SIZE, LP_FILTER_ORDER);
947
948 9136 ff_tilt_compensation(&p->tilt_mem, tilt_factor(p, lpc_n, lpc_d), buf_out,
949 AMR_SUBFRAME_SIZE);
950
951 9136 ff_adaptive_gain_control(buf_out, buf_out, speech_gain, AMR_SUBFRAME_SIZE,
952 AMR_AGC_ALPHA, &p->postfilter_agc);
953 9136 }
954
955 /// @}
956
957 2284 static int amrnb_decode_frame(AVCodecContext *avctx, void *data,
958 int *got_frame_ptr, AVPacket *avpkt)
959 {
960
961 2284 AMRChannelsContext *s = avctx->priv_data; // pointer to private data
962 2284 AVFrame *frame = data;
963 2284 const uint8_t *buf = avpkt->data;
964 2284 int buf_size = avpkt->size;
965 int ret;
966
967 /* get output buffer */
968 2284 frame->nb_samples = AMR_BLOCK_SIZE;
969
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 2284 times.
 2284 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
970 ✗ return ret;
971
972
2/2
✓ Branch 0 taken 2284 times.
✓ Branch 1 taken 2284 times.
 4568 for (int ch = 0; ch < avctx->channels; ch++) {
973 2284 AMRContext *p = &s->ch[ch];
974 float fixed_gain_factor;
975 2284 AMRFixed fixed_sparse = {0}; // fixed vector up to anti-sparseness processing
976 float spare_vector[AMR_SUBFRAME_SIZE]; // extra stack space to hold result from anti-sparseness processing
977 float synth_fixed_gain; // the fixed gain that synthesis should use
978 const float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
979 2284 float *buf_out = (float *)frame->extended_data[ch];
980 int channel_size;
981 int i, subframe;
982
983 2284 p->cur_frame_mode = unpack_bitstream(p, buf, buf_size);
984
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2284 times.
 2284 if (p->cur_frame_mode == NO_DATA) {
985 ✗ av_log(avctx, AV_LOG_ERROR, "Corrupt bitstream\n");
986 ✗ return AVERROR_INVALIDDATA;
987 }
988
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2284 times.
 2284 if (p->cur_frame_mode == MODE_DTX) {
989 ✗ avpriv_report_missing_feature(avctx, "dtx mode");
990 ✗ av_log(avctx, AV_LOG_INFO, "Note: libopencore_amrnb supports dtx\n");
991 ✗ return AVERROR_PATCHWELCOME;
992 }
993
994 2284 channel_size = frame_sizes_nb[p->cur_frame_mode] + 1; // +7 for rounding and +8 for TOC
995
2/2
✓ Branch 0 taken 285 times.
✓ Branch 1 taken 1999 times.
 2284 if (p->cur_frame_mode == MODE_12k2) {
996 285 lsf2lsp_5(p);
997 } else
998 1999 lsf2lsp_3(p);
999
1000
2/2
✓ Branch 0 taken 9136 times.
✓ Branch 1 taken 2284 times.
 11420 for (i = 0; i < 4; i++)
1001 9136 ff_acelp_lspd2lpc(p->lsp[i], p->lpc[i], 5);
1002
1003
2/2
✓ Branch 0 taken 9136 times.
✓ Branch 1 taken 2284 times.
 11420 for (subframe = 0; subframe < 4; subframe++) {
1004 9136 const AMRNBSubframe *amr_subframe = &p->frame.subframe[subframe];
1005
1006 9136 decode_pitch_vector(p, amr_subframe, subframe);
1007
1008 9136 decode_fixed_sparse(&fixed_sparse, amr_subframe->pulses,
1009 p->cur_frame_mode, subframe);
1010
1011 // The fixed gain (section 6.1.3) depends on the fixed vector
1012 // (section 6.1.2), but the fixed vector calculation uses
1013 // pitch sharpening based on the on the pitch gain (section 6.1.3).
1014 // So the correct order is: pitch gain, pitch sharpening, fixed gain.
1015 9136 decode_gains(p, amr_subframe, p->cur_frame_mode, subframe,
1016 &fixed_gain_factor);
1017
1018 9136 pitch_sharpening(p, subframe, p->cur_frame_mode, &fixed_sparse);
1019
1020
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 9136 times.
 9136 if (fixed_sparse.pitch_lag == 0) {
1021 ✗ av_log(avctx, AV_LOG_ERROR, "The file is corrupted, pitch_lag = 0 is not allowed\n");
1022 ✗ return AVERROR_INVALIDDATA;
1023 }
1024 9136 ff_set_fixed_vector(p->fixed_vector, &fixed_sparse, 1.0,
1025 AMR_SUBFRAME_SIZE);
1026
1027 9136 p->fixed_gain[4] =
1028 9136 ff_amr_set_fixed_gain(fixed_gain_factor,
1029 9136 p->celpm_ctx.dot_productf(p->fixed_vector,
1030 9136 p->fixed_vector,
1031 AMR_SUBFRAME_SIZE) /
1032 AMR_SUBFRAME_SIZE,
1033 9136 p->prediction_error,
1034 9136 energy_mean[p->cur_frame_mode], energy_pred_fac);
1035
1036 // The excitation feedback is calculated without any processing such
1037 // as fixed gain smoothing. This isn't mentioned in the specification.
1038
2/2
✓ Branch 0 taken 365440 times.
✓ Branch 1 taken 9136 times.
 374576 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
1039 365440 p->excitation[i] *= p->pitch_gain[4];
1040 9136 ff_set_fixed_vector(p->excitation, &fixed_sparse, p->fixed_gain[4],
1041 AMR_SUBFRAME_SIZE);
1042
1043 // In the ref decoder, excitation is stored with no fractional bits.
1044 // This step prevents buzz in silent periods. The ref encoder can
1045 // emit long sequences with pitch factor greater than one. This
1046 // creates unwanted feedback if the excitation vector is nonzero.
1047 // (e.g. test sequence T19_795.COD in 3GPP TS 26.074)
1048
2/2
✓ Branch 0 taken 365440 times.
✓ Branch 1 taken 9136 times.
 374576 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
1049 365440 p->excitation[i] = truncf(p->excitation[i]);
1050
1051 // Smooth fixed gain.
1052 // The specification is ambiguous, but in the reference source, the
1053 // smoothed value is NOT fed back into later fixed gain smoothing.
1054 9136 synth_fixed_gain = fixed_gain_smooth(p, p->lsf_q[subframe],
1055 9136 p->lsf_avg, p->cur_frame_mode);
1056
1057 9136 synth_fixed_vector = anti_sparseness(p, &fixed_sparse, p->fixed_vector,
1058 synth_fixed_gain, spare_vector);
1059
1060
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 9136 times.
 9136 if (synthesis(p, p->lpc[subframe], synth_fixed_gain,
1061 synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 0))
1062 // overflow detected -> rerun synthesis scaling pitch vector down
1063 // by a factor of 4, skipping pitch vector contribution emphasis
1064 // and adaptive gain control
1065 ✗ synthesis(p, p->lpc[subframe], synth_fixed_gain,
1066 synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 1);
1067
1068 9136 postfilter(p, p->lpc[subframe], buf_out + subframe * AMR_SUBFRAME_SIZE);
1069
1070 // update buffers and history
1071 9136 ff_clear_fixed_vector(p->fixed_vector, &fixed_sparse, AMR_SUBFRAME_SIZE);
1072 9136 update_state(p);
1073 }
1074
1075 2284 p->acelpf_ctx.acelp_apply_order_2_transfer_function(buf_out,
1076 buf_out, highpass_zeros,
1077 highpass_poles,
1078 2284 highpass_gain * AMR_SAMPLE_SCALE,
1079 2284 p->high_pass_mem, AMR_BLOCK_SIZE);
1080
1081 /* Update averaged lsf vector (used for fixed gain smoothing).
1082 *
1083 * Note that lsf_avg should not incorporate the current frame's LSFs
1084 * for fixed_gain_smooth.
1085 * The specification has an incorrect formula: the reference decoder uses
1086 * qbar(n-1) rather than qbar(n) in section 6.1(4) equation 71. */
1087 2284 p->acelpv_ctx.weighted_vector_sumf(p->lsf_avg, p->lsf_avg, p->lsf_q[3],
1088 0.84, 0.16, LP_FILTER_ORDER);
1089 2284 buf += channel_size;
1090 2284 buf_size -= channel_size;
1091 }
1092
1093 2284 *got_frame_ptr = 1;
1094
1095 2284 return avpkt->size;
1096 }
1097
1098
1099 const AVCodec ff_amrnb_decoder = {
1100 .name = "amrnb",
1101 .long_name = NULL_IF_CONFIG_SMALL("AMR-NB (Adaptive Multi-Rate NarrowBand)"),
1102 .type = AVMEDIA_TYPE_AUDIO,
1103 .id = AV_CODEC_ID_AMR_NB,
1104 .priv_data_size = sizeof(AMRChannelsContext),
1105 .init = amrnb_decode_init,
1106 .decode = amrnb_decode_frame,
1107 .capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF,
1108 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLTP,
1109 AV_SAMPLE_FMT_NONE },
1110 };
1111