celp_math: Replace duplicate ff_dot_productf() by ff_scalarproduct_c()
[libav.git] / libavcodec / amrnbdec.c
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 Libav.
7 *
8 * Libav 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 * Libav 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 Libav; 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 "avcodec.h"
47 #include "dsputil.h"
48 #include "libavutil/common.h"
49 #include "celp_filters.h"
50 #include "acelp_filters.h"
51 #include "acelp_vectors.h"
52 #include "acelp_pitch_delay.h"
53 #include "lsp.h"
54 #include "amr.h"
55
56 #include "amrnbdata.h"
57
58 #define AMR_BLOCK_SIZE 160 ///< samples per frame
59 #define AMR_SAMPLE_BOUND 32768.0 ///< threshold for synthesis overflow
60
61 /**
62 * Scale from constructed speech to [-1,1]
63 *
64 * AMR is designed to produce 16-bit PCM samples (3GPP TS 26.090 4.2) but
65 * upscales by two (section 6.2.2).
66 *
67 * Fundamentally, this scale is determined by energy_mean through
68 * the fixed vector contribution to the excitation vector.
69 */
70 #define AMR_SAMPLE_SCALE (2.0 / 32768.0)
71
72 /** Prediction factor for 12.2kbit/s mode */
73 #define PRED_FAC_MODE_12k2 0.65
74
75 #define LSF_R_FAC (8000.0 / 32768.0) ///< LSF residual tables to Hertz
76 #define MIN_LSF_SPACING (50.0488 / 8000.0) ///< Ensures stability of LPC filter
77 #define PITCH_LAG_MIN_MODE_12k2 18 ///< Lower bound on decoded lag search in 12.2kbit/s mode
78
79 /** Initial energy in dB. Also used for bad frames (unimplemented). */
80 #define MIN_ENERGY -14.0
81
82 /** Maximum sharpening factor
83 *
84 * The specification says 0.8, which should be 13107, but the reference C code
85 * uses 13017 instead. (Amusingly the same applies to SHARP_MAX in bitexact G.729.)
86 */
87 #define SHARP_MAX 0.79449462890625
88
89 /** Number of impulse response coefficients used for tilt factor */
90 #define AMR_TILT_RESPONSE 22
91 /** Tilt factor = 1st reflection coefficient * gamma_t */
92 #define AMR_TILT_GAMMA_T 0.8
93 /** Adaptive gain control factor used in post-filter */
94 #define AMR_AGC_ALPHA 0.9
95
96 typedef struct AMRContext {
97 AVFrame avframe; ///< AVFrame for decoded samples
98 AMRNBFrame frame; ///< decoded AMR parameters (lsf coefficients, codebook indexes, etc)
99 uint8_t bad_frame_indicator; ///< bad frame ? 1 : 0
100 enum Mode cur_frame_mode;
101
102 int16_t prev_lsf_r[LP_FILTER_ORDER]; ///< residual LSF vector from previous subframe
103 double lsp[4][LP_FILTER_ORDER]; ///< lsp vectors from current frame
104 double prev_lsp_sub4[LP_FILTER_ORDER]; ///< lsp vector for the 4th subframe of the previous frame
105
106 float lsf_q[4][LP_FILTER_ORDER]; ///< Interpolated LSF vector for fixed gain smoothing
107 float lsf_avg[LP_FILTER_ORDER]; ///< vector of averaged lsf vector
108
109 float lpc[4][LP_FILTER_ORDER]; ///< lpc coefficient vectors for 4 subframes
110
111 uint8_t pitch_lag_int; ///< integer part of pitch lag from current subframe
112
113 float excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1 + AMR_SUBFRAME_SIZE]; ///< current excitation and all necessary excitation history
114 float *excitation; ///< pointer to the current excitation vector in excitation_buf
115
116 float pitch_vector[AMR_SUBFRAME_SIZE]; ///< adaptive code book (pitch) vector
117 float fixed_vector[AMR_SUBFRAME_SIZE]; ///< algebraic codebook (fixed) vector (must be kept zero between frames)
118
119 float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes
120 float pitch_gain[5]; ///< quantified pitch gains for the current and previous four subframes
121 float fixed_gain[5]; ///< quantified fixed gains for the current and previous four subframes
122
123 float beta; ///< previous pitch_gain, bounded by [0.0,SHARP_MAX]
124 uint8_t diff_count; ///< the number of subframes for which diff has been above 0.65
125 uint8_t hang_count; ///< the number of subframes since a hangover period started
126
127 float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness processing to determine "onset"
128 uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none
129 uint8_t ir_filter_onset; ///< flag for impulse response filter strength
130
131 float postfilter_mem[10]; ///< previous intermediate values in the formant filter
132 float tilt_mem; ///< previous input to tilt compensation filter
133 float postfilter_agc; ///< previous factor used for adaptive gain control
134 float high_pass_mem[2]; ///< previous intermediate values in the high-pass filter
135
136 float samples_in[LP_FILTER_ORDER + AMR_SUBFRAME_SIZE]; ///< floating point samples
137
138 } AMRContext;
139
140 /** Double version of ff_weighted_vector_sumf() */
141 static void weighted_vector_sumd(double *out, const double *in_a,
142 const double *in_b, double weight_coeff_a,
143 double weight_coeff_b, int length)
144 {
145 int i;
146
147 for (i = 0; i < length; i++)
148 out[i] = weight_coeff_a * in_a[i]
149 + weight_coeff_b * in_b[i];
150 }
151
152 static av_cold int amrnb_decode_init(AVCodecContext *avctx)
153 {
154 AMRContext *p = avctx->priv_data;
155 int i;
156
157 avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
158
159 // p->excitation always points to the same position in p->excitation_buf
160 p->excitation = &p->excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1];
161
162 for (i = 0; i < LP_FILTER_ORDER; i++) {
163 p->prev_lsp_sub4[i] = lsp_sub4_init[i] * 1000 / (float)(1 << 15);
164 p->lsf_avg[i] = p->lsf_q[3][i] = lsp_avg_init[i] / (float)(1 << 15);
165 }
166
167 for (i = 0; i < 4; i++)
168 p->prediction_error[i] = MIN_ENERGY;
169
170 avcodec_get_frame_defaults(&p->avframe);
171 avctx->coded_frame = &p->avframe;
172
173 return 0;
174 }
175
176
177 /**
178 * Unpack an RFC4867 speech frame into the AMR frame mode and parameters.
179 *
180 * The order of speech bits is specified by 3GPP TS 26.101.
181 *
182 * @param p the context
183 * @param buf pointer to the input buffer
184 * @param buf_size size of the input buffer
185 *
186 * @return the frame mode
187 */
188 static enum Mode unpack_bitstream(AMRContext *p, const uint8_t *buf,
189 int buf_size)
190 {
191 enum Mode mode;
192
193 // Decode the first octet.
194 mode = buf[0] >> 3 & 0x0F; // frame type
195 p->bad_frame_indicator = (buf[0] & 0x4) != 0x4; // quality bit
196
197 if (mode >= N_MODES || buf_size < frame_sizes_nb[mode] + 1) {
198 return NO_DATA;
199 }
200
201 if (mode < MODE_DTX)
202 ff_amr_bit_reorder((uint16_t *) &p->frame, sizeof(AMRNBFrame), buf + 1,
203 amr_unpacking_bitmaps_per_mode[mode]);
204
205 return mode;
206 }
207
208
209 /// @name AMR pitch LPC coefficient decoding functions
210 /// @{
211
212 /**
213 * Interpolate the LSF vector (used for fixed gain smoothing).
214 * The interpolation is done over all four subframes even in MODE_12k2.
215 *
216 * @param[in,out] lsf_q LSFs in [0,1] for each subframe
217 * @param[in] lsf_new New LSFs in [0,1] for subframe 4
218 */
219 static void interpolate_lsf(float lsf_q[4][LP_FILTER_ORDER], float *lsf_new)
220 {
221 int i;
222
223 for (i = 0; i < 4; i++)
224 ff_weighted_vector_sumf(lsf_q[i], lsf_q[3], lsf_new,
225 0.25 * (3 - i), 0.25 * (i + 1),
226 LP_FILTER_ORDER);
227 }
228
229 /**
230 * Decode a set of 5 split-matrix quantized lsf indexes into an lsp vector.
231 *
232 * @param p the context
233 * @param lsp output LSP vector
234 * @param lsf_no_r LSF vector without the residual vector added
235 * @param lsf_quantizer pointers to LSF dictionary tables
236 * @param quantizer_offset offset in tables
237 * @param sign for the 3 dictionary table
238 * @param update store data for computing the next frame's LSFs
239 */
240 static void lsf2lsp_for_mode12k2(AMRContext *p, double lsp[LP_FILTER_ORDER],
241 const float lsf_no_r[LP_FILTER_ORDER],
242 const int16_t *lsf_quantizer[5],
243 const int quantizer_offset,
244 const int sign, const int update)
245 {
246 int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector
247 float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector
248 int i;
249
250 for (i = 0; i < LP_FILTER_ORDER >> 1; i++)
251 memcpy(&lsf_r[i << 1], &lsf_quantizer[i][quantizer_offset],
252 2 * sizeof(*lsf_r));
253
254 if (sign) {
255 lsf_r[4] *= -1;
256 lsf_r[5] *= -1;
257 }
258
259 if (update)
260 memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r));
261
262 for (i = 0; i < LP_FILTER_ORDER; i++)
263 lsf_q[i] = lsf_r[i] * (LSF_R_FAC / 8000.0) + lsf_no_r[i] * (1.0 / 8000.0);
264
265 ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
266
267 if (update)
268 interpolate_lsf(p->lsf_q, lsf_q);
269
270 ff_acelp_lsf2lspd(lsp, lsf_q, LP_FILTER_ORDER);
271 }
272
273 /**
274 * Decode a set of 5 split-matrix quantized lsf indexes into 2 lsp vectors.
275 *
276 * @param p pointer to the AMRContext
277 */
278 static void lsf2lsp_5(AMRContext *p)
279 {
280 const uint16_t *lsf_param = p->frame.lsf;
281 float lsf_no_r[LP_FILTER_ORDER]; // LSFs without the residual vector
282 const int16_t *lsf_quantizer[5];
283 int i;
284
285 lsf_quantizer[0] = lsf_5_1[lsf_param[0]];
286 lsf_quantizer[1] = lsf_5_2[lsf_param[1]];
287 lsf_quantizer[2] = lsf_5_3[lsf_param[2] >> 1];
288 lsf_quantizer[3] = lsf_5_4[lsf_param[3]];
289 lsf_quantizer[4] = lsf_5_5[lsf_param[4]];
290
291 for (i = 0; i < LP_FILTER_ORDER; i++)
292 lsf_no_r[i] = p->prev_lsf_r[i] * LSF_R_FAC * PRED_FAC_MODE_12k2 + lsf_5_mean[i];
293
294 lsf2lsp_for_mode12k2(p, p->lsp[1], lsf_no_r, lsf_quantizer, 0, lsf_param[2] & 1, 0);
295 lsf2lsp_for_mode12k2(p, p->lsp[3], lsf_no_r, lsf_quantizer, 2, lsf_param[2] & 1, 1);
296
297 // interpolate LSP vectors at subframes 1 and 3
298 weighted_vector_sumd(p->lsp[0], p->prev_lsp_sub4, p->lsp[1], 0.5, 0.5, LP_FILTER_ORDER);
299 weighted_vector_sumd(p->lsp[2], p->lsp[1] , p->lsp[3], 0.5, 0.5, LP_FILTER_ORDER);
300 }
301
302 /**
303 * Decode a set of 3 split-matrix quantized lsf indexes into an lsp vector.
304 *
305 * @param p pointer to the AMRContext
306 */
307 static void lsf2lsp_3(AMRContext *p)
308 {
309 const uint16_t *lsf_param = p->frame.lsf;
310 int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector
311 float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector
312 const int16_t *lsf_quantizer;
313 int i, j;
314
315 lsf_quantizer = (p->cur_frame_mode == MODE_7k95 ? lsf_3_1_MODE_7k95 : lsf_3_1)[lsf_param[0]];
316 memcpy(lsf_r, lsf_quantizer, 3 * sizeof(*lsf_r));
317
318 lsf_quantizer = lsf_3_2[lsf_param[1] << (p->cur_frame_mode <= MODE_5k15)];
319 memcpy(lsf_r + 3, lsf_quantizer, 3 * sizeof(*lsf_r));
320
321 lsf_quantizer = (p->cur_frame_mode <= MODE_5k15 ? lsf_3_3_MODE_5k15 : lsf_3_3)[lsf_param[2]];
322 memcpy(lsf_r + 6, lsf_quantizer, 4 * sizeof(*lsf_r));
323
324 // calculate mean-removed LSF vector and add mean
325 for (i = 0; i < LP_FILTER_ORDER; i++)
326 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);
327
328 ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
329
330 // store data for computing the next frame's LSFs
331 interpolate_lsf(p->lsf_q, lsf_q);
332 memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r));
333
334 ff_acelp_lsf2lspd(p->lsp[3], lsf_q, LP_FILTER_ORDER);
335
336 // interpolate LSP vectors at subframes 1, 2 and 3
337 for (i = 1; i <= 3; i++)
338 for(j = 0; j < LP_FILTER_ORDER; j++)
339 p->lsp[i-1][j] = p->prev_lsp_sub4[j] +
340 (p->lsp[3][j] - p->prev_lsp_sub4[j]) * 0.25 * i;
341 }
342
343 /// @}
344
345
346 /// @name AMR pitch vector decoding functions
347 /// @{
348
349 /**
350 * Like ff_decode_pitch_lag(), but with 1/6 resolution
351 */
352 static void decode_pitch_lag_1_6(int *lag_int, int *lag_frac, int pitch_index,
353 const int prev_lag_int, const int subframe)
354 {
355 if (subframe == 0 || subframe == 2) {
356 if (pitch_index < 463) {
357 *lag_int = (pitch_index + 107) * 10923 >> 16;
358 *lag_frac = pitch_index - *lag_int * 6 + 105;
359 } else {
360 *lag_int = pitch_index - 368;
361 *lag_frac = 0;
362 }
363 } else {
364 *lag_int = ((pitch_index + 5) * 10923 >> 16) - 1;
365 *lag_frac = pitch_index - *lag_int * 6 - 3;
366 *lag_int += av_clip(prev_lag_int - 5, PITCH_LAG_MIN_MODE_12k2,
367 PITCH_DELAY_MAX - 9);
368 }
369 }
370
371 static void decode_pitch_vector(AMRContext *p,
372 const AMRNBSubframe *amr_subframe,
373 const int subframe)
374 {
375 int pitch_lag_int, pitch_lag_frac;
376 enum Mode mode = p->cur_frame_mode;
377
378 if (p->cur_frame_mode == MODE_12k2) {
379 decode_pitch_lag_1_6(&pitch_lag_int, &pitch_lag_frac,
380 amr_subframe->p_lag, p->pitch_lag_int,
381 subframe);
382 } else
383 ff_decode_pitch_lag(&pitch_lag_int, &pitch_lag_frac,
384 amr_subframe->p_lag,
385 p->pitch_lag_int, subframe,
386 mode != MODE_4k75 && mode != MODE_5k15,
387 mode <= MODE_6k7 ? 4 : (mode == MODE_7k95 ? 5 : 6));
388
389 p->pitch_lag_int = pitch_lag_int; // store previous lag in a uint8_t
390
391 pitch_lag_frac <<= (p->cur_frame_mode != MODE_12k2);
392
393 pitch_lag_int += pitch_lag_frac > 0;
394
395 /* Calculate the pitch vector by interpolating the past excitation at the
396 pitch lag using a b60 hamming windowed sinc function. */
397 ff_acelp_interpolatef(p->excitation, p->excitation + 1 - pitch_lag_int,
398 ff_b60_sinc, 6,
399 pitch_lag_frac + 6 - 6*(pitch_lag_frac > 0),
400 10, AMR_SUBFRAME_SIZE);
401
402 memcpy(p->pitch_vector, p->excitation, AMR_SUBFRAME_SIZE * sizeof(float));
403 }
404
405 /// @}
406
407
408 /// @name AMR algebraic code book (fixed) vector decoding functions
409 /// @{
410
411 /**
412 * Decode a 10-bit algebraic codebook index from a 10.2 kbit/s frame.
413 */
414 static void decode_10bit_pulse(int code, int pulse_position[8],
415 int i1, int i2, int i3)
416 {
417 // coded using 7+3 bits with the 3 LSBs being, individually, the LSB of 1 of
418 // the 3 pulses and the upper 7 bits being coded in base 5
419 const uint8_t *positions = base_five_table[code >> 3];
420 pulse_position[i1] = (positions[2] << 1) + ( code & 1);
421 pulse_position[i2] = (positions[1] << 1) + ((code >> 1) & 1);
422 pulse_position[i3] = (positions[0] << 1) + ((code >> 2) & 1);
423 }
424
425 /**
426 * Decode the algebraic codebook index to pulse positions and signs and
427 * construct the algebraic codebook vector for MODE_10k2.
428 *
429 * @param fixed_index positions of the eight pulses
430 * @param fixed_sparse pointer to the algebraic codebook vector
431 */
432 static void decode_8_pulses_31bits(const int16_t *fixed_index,
433 AMRFixed *fixed_sparse)
434 {
435 int pulse_position[8];
436 int i, temp;
437
438 decode_10bit_pulse(fixed_index[4], pulse_position, 0, 4, 1);
439 decode_10bit_pulse(fixed_index[5], pulse_position, 2, 6, 5);
440
441 // coded using 5+2 bits with the 2 LSBs being, individually, the LSB of 1 of
442 // the 2 pulses and the upper 5 bits being coded in base 5
443 temp = ((fixed_index[6] >> 2) * 25 + 12) >> 5;
444 pulse_position[3] = temp % 5;
445 pulse_position[7] = temp / 5;
446 if (pulse_position[7] & 1)
447 pulse_position[3] = 4 - pulse_position[3];
448 pulse_position[3] = (pulse_position[3] << 1) + ( fixed_index[6] & 1);
449 pulse_position[7] = (pulse_position[7] << 1) + ((fixed_index[6] >> 1) & 1);
450
451 fixed_sparse->n = 8;
452 for (i = 0; i < 4; i++) {
453 const int pos1 = (pulse_position[i] << 2) + i;
454 const int pos2 = (pulse_position[i + 4] << 2) + i;
455 const float sign = fixed_index[i] ? -1.0 : 1.0;
456 fixed_sparse->x[i ] = pos1;
457 fixed_sparse->x[i + 4] = pos2;
458 fixed_sparse->y[i ] = sign;
459 fixed_sparse->y[i + 4] = pos2 < pos1 ? -sign : sign;
460 }
461 }
462
463 /**
464 * Decode the algebraic codebook index to pulse positions and signs,
465 * then construct the algebraic codebook vector.
466 *
467 * nb of pulses | bits encoding pulses
468 * For MODE_4k75 or MODE_5k15, 2 | 1-3, 4-6, 7
469 * MODE_5k9, 2 | 1, 2-4, 5-6, 7-9
470 * MODE_6k7, 3 | 1-3, 4, 5-7, 8, 9-11
471 * MODE_7k4 or MODE_7k95, 4 | 1-3, 4-6, 7-9, 10, 11-13
472 *
473 * @param fixed_sparse pointer to the algebraic codebook vector
474 * @param pulses algebraic codebook indexes
475 * @param mode mode of the current frame
476 * @param subframe current subframe number
477 */
478 static void decode_fixed_sparse(AMRFixed *fixed_sparse, const uint16_t *pulses,
479 const enum Mode mode, const int subframe)
480 {
481 assert(MODE_4k75 <= mode && mode <= MODE_12k2);
482
483 if (mode == MODE_12k2) {
484 ff_decode_10_pulses_35bits(pulses, fixed_sparse, gray_decode, 5, 3);
485 } else if (mode == MODE_10k2) {
486 decode_8_pulses_31bits(pulses, fixed_sparse);
487 } else {
488 int *pulse_position = fixed_sparse->x;
489 int i, pulse_subset;
490 const int fixed_index = pulses[0];
491
492 if (mode <= MODE_5k15) {
493 pulse_subset = ((fixed_index >> 3) & 8) + (subframe << 1);
494 pulse_position[0] = ( fixed_index & 7) * 5 + track_position[pulse_subset];
495 pulse_position[1] = ((fixed_index >> 3) & 7) * 5 + track_position[pulse_subset + 1];
496 fixed_sparse->n = 2;
497 } else if (mode == MODE_5k9) {
498 pulse_subset = ((fixed_index & 1) << 1) + 1;
499 pulse_position[0] = ((fixed_index >> 1) & 7) * 5 + pulse_subset;
500 pulse_subset = (fixed_index >> 4) & 3;
501 pulse_position[1] = ((fixed_index >> 6) & 7) * 5 + pulse_subset + (pulse_subset == 3 ? 1 : 0);
502 fixed_sparse->n = pulse_position[0] == pulse_position[1] ? 1 : 2;
503 } else if (mode == MODE_6k7) {
504 pulse_position[0] = (fixed_index & 7) * 5;
505 pulse_subset = (fixed_index >> 2) & 2;
506 pulse_position[1] = ((fixed_index >> 4) & 7) * 5 + pulse_subset + 1;
507 pulse_subset = (fixed_index >> 6) & 2;
508 pulse_position[2] = ((fixed_index >> 8) & 7) * 5 + pulse_subset + 2;
509 fixed_sparse->n = 3;
510 } else { // mode <= MODE_7k95
511 pulse_position[0] = gray_decode[ fixed_index & 7];
512 pulse_position[1] = gray_decode[(fixed_index >> 3) & 7] + 1;
513 pulse_position[2] = gray_decode[(fixed_index >> 6) & 7] + 2;
514 pulse_subset = (fixed_index >> 9) & 1;
515 pulse_position[3] = gray_decode[(fixed_index >> 10) & 7] + pulse_subset + 3;
516 fixed_sparse->n = 4;
517 }
518 for (i = 0; i < fixed_sparse->n; i++)
519 fixed_sparse->y[i] = (pulses[1] >> i) & 1 ? 1.0 : -1.0;
520 }
521 }
522
523 /**
524 * Apply pitch lag to obtain the sharpened fixed vector (section 6.1.2)
525 *
526 * @param p the context
527 * @param subframe unpacked amr subframe
528 * @param mode mode of the current frame
529 * @param fixed_sparse sparse respresentation of the fixed vector
530 */
531 static void pitch_sharpening(AMRContext *p, int subframe, enum Mode mode,
532 AMRFixed *fixed_sparse)
533 {
534 // The spec suggests the current pitch gain is always used, but in other
535 // modes the pitch and codebook gains are joinly quantized (sec 5.8.2)
536 // so the codebook gain cannot depend on the quantized pitch gain.
537 if (mode == MODE_12k2)
538 p->beta = FFMIN(p->pitch_gain[4], 1.0);
539
540 fixed_sparse->pitch_lag = p->pitch_lag_int;
541 fixed_sparse->pitch_fac = p->beta;
542
543 // Save pitch sharpening factor for the next subframe
544 // MODE_4k75 only updates on the 2nd and 4th subframes - this follows from
545 // the fact that the gains for two subframes are jointly quantized.
546 if (mode != MODE_4k75 || subframe & 1)
547 p->beta = av_clipf(p->pitch_gain[4], 0.0, SHARP_MAX);
548 }
549 /// @}
550
551
552 /// @name AMR gain decoding functions
553 /// @{
554
555 /**
556 * fixed gain smoothing
557 * Note that where the spec specifies the "spectrum in the q domain"
558 * in section 6.1.4, in fact frequencies should be used.
559 *
560 * @param p the context
561 * @param lsf LSFs for the current subframe, in the range [0,1]
562 * @param lsf_avg averaged LSFs
563 * @param mode mode of the current frame
564 *
565 * @return fixed gain smoothed
566 */
567 static float fixed_gain_smooth(AMRContext *p , const float *lsf,
568 const float *lsf_avg, const enum Mode mode)
569 {
570 float diff = 0.0;
571 int i;
572
573 for (i = 0; i < LP_FILTER_ORDER; i++)
574 diff += fabs(lsf_avg[i] - lsf[i]) / lsf_avg[i];
575
576 // If diff is large for ten subframes, disable smoothing for a 40-subframe
577 // hangover period.
578 p->diff_count++;
579 if (diff <= 0.65)
580 p->diff_count = 0;
581
582 if (p->diff_count > 10) {
583 p->hang_count = 0;
584 p->diff_count--; // don't let diff_count overflow
585 }
586
587 if (p->hang_count < 40) {
588 p->hang_count++;
589 } else if (mode < MODE_7k4 || mode == MODE_10k2) {
590 const float smoothing_factor = av_clipf(4.0 * diff - 1.6, 0.0, 1.0);
591 const float fixed_gain_mean = (p->fixed_gain[0] + p->fixed_gain[1] +
592 p->fixed_gain[2] + p->fixed_gain[3] +
593 p->fixed_gain[4]) * 0.2;
594 return smoothing_factor * p->fixed_gain[4] +
595 (1.0 - smoothing_factor) * fixed_gain_mean;
596 }
597 return p->fixed_gain[4];
598 }
599
600 /**
601 * Decode pitch gain and fixed gain factor (part of section 6.1.3).
602 *
603 * @param p the context
604 * @param amr_subframe unpacked amr subframe
605 * @param mode mode of the current frame
606 * @param subframe current subframe number
607 * @param fixed_gain_factor decoded gain correction factor
608 */
609 static void decode_gains(AMRContext *p, const AMRNBSubframe *amr_subframe,
610 const enum Mode mode, const int subframe,
611 float *fixed_gain_factor)
612 {
613 if (mode == MODE_12k2 || mode == MODE_7k95) {
614 p->pitch_gain[4] = qua_gain_pit [amr_subframe->p_gain ]
615 * (1.0 / 16384.0);
616 *fixed_gain_factor = qua_gain_code[amr_subframe->fixed_gain]
617 * (1.0 / 2048.0);
618 } else {
619 const uint16_t *gains;
620
621 if (mode >= MODE_6k7) {
622 gains = gains_high[amr_subframe->p_gain];
623 } else if (mode >= MODE_5k15) {
624 gains = gains_low [amr_subframe->p_gain];
625 } else {
626 // gain index is only coded in subframes 0,2 for MODE_4k75
627 gains = gains_MODE_4k75[(p->frame.subframe[subframe & 2].p_gain << 1) + (subframe & 1)];
628 }
629
630 p->pitch_gain[4] = gains[0] * (1.0 / 16384.0);
631 *fixed_gain_factor = gains[1] * (1.0 / 4096.0);
632 }
633 }
634
635 /// @}
636
637
638 /// @name AMR preprocessing functions
639 /// @{
640
641 /**
642 * Circularly convolve a sparse fixed vector with a phase dispersion impulse
643 * response filter (D.6.2 of G.729 and 6.1.5 of AMR).
644 *
645 * @param out vector with filter applied
646 * @param in source vector
647 * @param filter phase filter coefficients
648 *
649 * out[n] = sum(i,0,len-1){ in[i] * filter[(len + n - i)%len] }
650 */
651 static void apply_ir_filter(float *out, const AMRFixed *in,
652 const float *filter)
653 {
654 float filter1[AMR_SUBFRAME_SIZE], ///< filters at pitch lag*1 and *2
655 filter2[AMR_SUBFRAME_SIZE];
656 int lag = in->pitch_lag;
657 float fac = in->pitch_fac;
658 int i;
659
660 if (lag < AMR_SUBFRAME_SIZE) {
661 ff_celp_circ_addf(filter1, filter, filter, lag, fac,
662 AMR_SUBFRAME_SIZE);
663
664 if (lag < AMR_SUBFRAME_SIZE >> 1)
665 ff_celp_circ_addf(filter2, filter, filter1, lag, fac,
666 AMR_SUBFRAME_SIZE);
667 }
668
669 memset(out, 0, sizeof(float) * AMR_SUBFRAME_SIZE);
670 for (i = 0; i < in->n; i++) {
671 int x = in->x[i];
672 float y = in->y[i];
673 const float *filterp;
674
675 if (x >= AMR_SUBFRAME_SIZE - lag) {
676 filterp = filter;
677 } else if (x >= AMR_SUBFRAME_SIZE - (lag << 1)) {
678 filterp = filter1;
679 } else
680 filterp = filter2;
681
682 ff_celp_circ_addf(out, out, filterp, x, y, AMR_SUBFRAME_SIZE);
683 }
684 }
685
686 /**
687 * Reduce fixed vector sparseness by smoothing with one of three IR filters.
688 * Also know as "adaptive phase dispersion".
689 *
690 * This implements 3GPP TS 26.090 section 6.1(5).
691 *
692 * @param p the context
693 * @param fixed_sparse algebraic codebook vector
694 * @param fixed_vector unfiltered fixed vector
695 * @param fixed_gain smoothed gain
696 * @param out space for modified vector if necessary
697 */
698 static const float *anti_sparseness(AMRContext *p, AMRFixed *fixed_sparse,
699 const float *fixed_vector,
700 float fixed_gain, float *out)
701 {
702 int ir_filter_nr;
703
704 if (p->pitch_gain[4] < 0.6) {
705 ir_filter_nr = 0; // strong filtering
706 } else if (p->pitch_gain[4] < 0.9) {
707 ir_filter_nr = 1; // medium filtering
708 } else
709 ir_filter_nr = 2; // no filtering
710
711 // detect 'onset'
712 if (fixed_gain > 2.0 * p->prev_sparse_fixed_gain) {
713 p->ir_filter_onset = 2;
714 } else if (p->ir_filter_onset)
715 p->ir_filter_onset--;
716
717 if (!p->ir_filter_onset) {
718 int i, count = 0;
719
720 for (i = 0; i < 5; i++)
721 if (p->pitch_gain[i] < 0.6)
722 count++;
723 if (count > 2)
724 ir_filter_nr = 0;
725
726 if (ir_filter_nr > p->prev_ir_filter_nr + 1)
727 ir_filter_nr--;
728 } else if (ir_filter_nr < 2)
729 ir_filter_nr++;
730
731 // Disable filtering for very low level of fixed_gain.
732 // Note this step is not specified in the technical description but is in
733 // the reference source in the function Ph_disp.
734 if (fixed_gain < 5.0)
735 ir_filter_nr = 2;
736
737 if (p->cur_frame_mode != MODE_7k4 && p->cur_frame_mode < MODE_10k2
738 && ir_filter_nr < 2) {
739 apply_ir_filter(out, fixed_sparse,
740 (p->cur_frame_mode == MODE_7k95 ?
741 ir_filters_lookup_MODE_7k95 :
742 ir_filters_lookup)[ir_filter_nr]);
743 fixed_vector = out;
744 }
745
746 // update ir filter strength history
747 p->prev_ir_filter_nr = ir_filter_nr;
748 p->prev_sparse_fixed_gain = fixed_gain;
749
750 return fixed_vector;
751 }
752
753 /// @}
754
755
756 /// @name AMR synthesis functions
757 /// @{
758
759 /**
760 * Conduct 10th order linear predictive coding synthesis.
761 *
762 * @param p pointer to the AMRContext
763 * @param lpc pointer to the LPC coefficients
764 * @param fixed_gain fixed codebook gain for synthesis
765 * @param fixed_vector algebraic codebook vector
766 * @param samples pointer to the output speech samples
767 * @param overflow 16-bit overflow flag
768 */
769 static int synthesis(AMRContext *p, float *lpc,
770 float fixed_gain, const float *fixed_vector,
771 float *samples, uint8_t overflow)
772 {
773 int i;
774 float excitation[AMR_SUBFRAME_SIZE];
775
776 // if an overflow has been detected, the pitch vector is scaled down by a
777 // factor of 4
778 if (overflow)
779 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
780 p->pitch_vector[i] *= 0.25;
781
782 ff_weighted_vector_sumf(excitation, p->pitch_vector, fixed_vector,
783 p->pitch_gain[4], fixed_gain, AMR_SUBFRAME_SIZE);
784
785 // emphasize pitch vector contribution
786 if (p->pitch_gain[4] > 0.5 && !overflow) {
787 float energy = ff_scalarproduct_float_c(excitation, excitation,
788 AMR_SUBFRAME_SIZE);
789 float pitch_factor =
790 p->pitch_gain[4] *
791 (p->cur_frame_mode == MODE_12k2 ?
792 0.25 * FFMIN(p->pitch_gain[4], 1.0) :
793 0.5 * FFMIN(p->pitch_gain[4], SHARP_MAX));
794
795 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
796 excitation[i] += pitch_factor * p->pitch_vector[i];
797
798 ff_scale_vector_to_given_sum_of_squares(excitation, excitation, energy,
799 AMR_SUBFRAME_SIZE);
800 }
801
802 ff_celp_lp_synthesis_filterf(samples, lpc, excitation, AMR_SUBFRAME_SIZE,
803 LP_FILTER_ORDER);
804
805 // detect overflow
806 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
807 if (fabsf(samples[i]) > AMR_SAMPLE_BOUND) {
808 return 1;
809 }
810
811 return 0;
812 }
813
814 /// @}
815
816
817 /// @name AMR update functions
818 /// @{
819
820 /**
821 * Update buffers and history at the end of decoding a subframe.
822 *
823 * @param p pointer to the AMRContext
824 */
825 static void update_state(AMRContext *p)
826 {
827 memcpy(p->prev_lsp_sub4, p->lsp[3], LP_FILTER_ORDER * sizeof(p->lsp[3][0]));
828
829 memmove(&p->excitation_buf[0], &p->excitation_buf[AMR_SUBFRAME_SIZE],
830 (PITCH_DELAY_MAX + LP_FILTER_ORDER + 1) * sizeof(float));
831
832 memmove(&p->pitch_gain[0], &p->pitch_gain[1], 4 * sizeof(float));
833 memmove(&p->fixed_gain[0], &p->fixed_gain[1], 4 * sizeof(float));
834
835 memmove(&p->samples_in[0], &p->samples_in[AMR_SUBFRAME_SIZE],
836 LP_FILTER_ORDER * sizeof(float));
837 }
838
839 /// @}
840
841
842 /// @name AMR Postprocessing functions
843 /// @{
844
845 /**
846 * Get the tilt factor of a formant filter from its transfer function
847 *
848 * @param lpc_n LP_FILTER_ORDER coefficients of the numerator
849 * @param lpc_d LP_FILTER_ORDER coefficients of the denominator
850 */
851 static float tilt_factor(float *lpc_n, float *lpc_d)
852 {
853 float rh0, rh1; // autocorrelation at lag 0 and 1
854
855 // LP_FILTER_ORDER prior zeros are needed for ff_celp_lp_synthesis_filterf
856 float impulse_buffer[LP_FILTER_ORDER + AMR_TILT_RESPONSE] = { 0 };
857 float *hf = impulse_buffer + LP_FILTER_ORDER; // start of impulse response
858
859 hf[0] = 1.0;
860 memcpy(hf + 1, lpc_n, sizeof(float) * LP_FILTER_ORDER);
861 ff_celp_lp_synthesis_filterf(hf, lpc_d, hf, AMR_TILT_RESPONSE,
862 LP_FILTER_ORDER);
863
864 rh0 = ff_scalarproduct_float_c(hf, hf, AMR_TILT_RESPONSE);
865 rh1 = ff_scalarproduct_float_c(hf, hf + 1, AMR_TILT_RESPONSE - 1);
866
867 // The spec only specifies this check for 12.2 and 10.2 kbit/s
868 // modes. But in the ref source the tilt is always non-negative.
869 return rh1 >= 0.0 ? rh1 / rh0 * AMR_TILT_GAMMA_T : 0.0;
870 }
871
872 /**
873 * Perform adaptive post-filtering to enhance the quality of the speech.
874 * See section 6.2.1.
875 *
876 * @param p pointer to the AMRContext
877 * @param lpc interpolated LP coefficients for this subframe
878 * @param buf_out output of the filter
879 */
880 static void postfilter(AMRContext *p, float *lpc, float *buf_out)
881 {
882 int i;
883 float *samples = p->samples_in + LP_FILTER_ORDER; // Start of input
884
885 float speech_gain = ff_scalarproduct_float_c(samples, samples,
886 AMR_SUBFRAME_SIZE);
887
888 float pole_out[AMR_SUBFRAME_SIZE + LP_FILTER_ORDER]; // Output of pole filter
889 const float *gamma_n, *gamma_d; // Formant filter factor table
890 float lpc_n[LP_FILTER_ORDER], lpc_d[LP_FILTER_ORDER]; // Transfer function coefficients
891
892 if (p->cur_frame_mode == MODE_12k2 || p->cur_frame_mode == MODE_10k2) {
893 gamma_n = ff_pow_0_7;
894 gamma_d = ff_pow_0_75;
895 } else {
896 gamma_n = ff_pow_0_55;
897 gamma_d = ff_pow_0_7;
898 }
899
900 for (i = 0; i < LP_FILTER_ORDER; i++) {
901 lpc_n[i] = lpc[i] * gamma_n[i];
902 lpc_d[i] = lpc[i] * gamma_d[i];
903 }
904
905 memcpy(pole_out, p->postfilter_mem, sizeof(float) * LP_FILTER_ORDER);
906 ff_celp_lp_synthesis_filterf(pole_out + LP_FILTER_ORDER, lpc_d, samples,
907 AMR_SUBFRAME_SIZE, LP_FILTER_ORDER);
908 memcpy(p->postfilter_mem, pole_out + AMR_SUBFRAME_SIZE,
909 sizeof(float) * LP_FILTER_ORDER);
910
911 ff_celp_lp_zero_synthesis_filterf(buf_out, lpc_n,
912 pole_out + LP_FILTER_ORDER,
913 AMR_SUBFRAME_SIZE, LP_FILTER_ORDER);
914
915 ff_tilt_compensation(&p->tilt_mem, tilt_factor(lpc_n, lpc_d), buf_out,
916 AMR_SUBFRAME_SIZE);
917
918 ff_adaptive_gain_control(buf_out, buf_out, speech_gain, AMR_SUBFRAME_SIZE,
919 AMR_AGC_ALPHA, &p->postfilter_agc);
920 }
921
922 /// @}
923
924 static int amrnb_decode_frame(AVCodecContext *avctx, void *data,
925 int *got_frame_ptr, AVPacket *avpkt)
926 {
927
928 AMRContext *p = avctx->priv_data; // pointer to private data
929 const uint8_t *buf = avpkt->data;
930 int buf_size = avpkt->size;
931 float *buf_out; // pointer to the output data buffer
932 int i, subframe, ret;
933 float fixed_gain_factor;
934 AMRFixed fixed_sparse = {0}; // fixed vector up to anti-sparseness processing
935 float spare_vector[AMR_SUBFRAME_SIZE]; // extra stack space to hold result from anti-sparseness processing
936 float synth_fixed_gain; // the fixed gain that synthesis should use
937 const float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
938
939 /* get output buffer */
940 p->avframe.nb_samples = AMR_BLOCK_SIZE;
941 if ((ret = avctx->get_buffer(avctx, &p->avframe)) < 0) {
942 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
943 return ret;
944 }
945 buf_out = (float *)p->avframe.data[0];
946
947 p->cur_frame_mode = unpack_bitstream(p, buf, buf_size);
948 if (p->cur_frame_mode == NO_DATA) {
949 av_log(avctx, AV_LOG_ERROR, "Corrupt bitstream\n");
950 return AVERROR_INVALIDDATA;
951 }
952 if (p->cur_frame_mode == MODE_DTX) {
953 av_log_missing_feature(avctx, "dtx mode", 1);
954 return -1;
955 }
956
957 if (p->cur_frame_mode == MODE_12k2) {
958 lsf2lsp_5(p);
959 } else
960 lsf2lsp_3(p);
961
962 for (i = 0; i < 4; i++)
963 ff_acelp_lspd2lpc(p->lsp[i], p->lpc[i], 5);
964
965 for (subframe = 0; subframe < 4; subframe++) {
966 const AMRNBSubframe *amr_subframe = &p->frame.subframe[subframe];
967
968 decode_pitch_vector(p, amr_subframe, subframe);
969
970 decode_fixed_sparse(&fixed_sparse, amr_subframe->pulses,
971 p->cur_frame_mode, subframe);
972
973 // The fixed gain (section 6.1.3) depends on the fixed vector
974 // (section 6.1.2), but the fixed vector calculation uses
975 // pitch sharpening based on the on the pitch gain (section 6.1.3).
976 // So the correct order is: pitch gain, pitch sharpening, fixed gain.
977 decode_gains(p, amr_subframe, p->cur_frame_mode, subframe,
978 &fixed_gain_factor);
979
980 pitch_sharpening(p, subframe, p->cur_frame_mode, &fixed_sparse);
981
982 if (fixed_sparse.pitch_lag == 0) {
983 av_log(avctx, AV_LOG_ERROR, "The file is corrupted, pitch_lag = 0 is not allowed\n");
984 return AVERROR_INVALIDDATA;
985 }
986 ff_set_fixed_vector(p->fixed_vector, &fixed_sparse, 1.0,
987 AMR_SUBFRAME_SIZE);
988
989 p->fixed_gain[4] =
990 ff_amr_set_fixed_gain(fixed_gain_factor,
991 ff_scalarproduct_float_c(p->fixed_vector,
992 p->fixed_vector,
993 AMR_SUBFRAME_SIZE) /
994 AMR_SUBFRAME_SIZE,
995 p->prediction_error,
996 energy_mean[p->cur_frame_mode], energy_pred_fac);
997
998 // The excitation feedback is calculated without any processing such
999 // as fixed gain smoothing. This isn't mentioned in the specification.
1000 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
1001 p->excitation[i] *= p->pitch_gain[4];
1002 ff_set_fixed_vector(p->excitation, &fixed_sparse, p->fixed_gain[4],
1003 AMR_SUBFRAME_SIZE);
1004
1005 // In the ref decoder, excitation is stored with no fractional bits.
1006 // This step prevents buzz in silent periods. The ref encoder can
1007 // emit long sequences with pitch factor greater than one. This
1008 // creates unwanted feedback if the excitation vector is nonzero.
1009 // (e.g. test sequence T19_795.COD in 3GPP TS 26.074)
1010 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
1011 p->excitation[i] = truncf(p->excitation[i]);
1012
1013 // Smooth fixed gain.
1014 // The specification is ambiguous, but in the reference source, the
1015 // smoothed value is NOT fed back into later fixed gain smoothing.
1016 synth_fixed_gain = fixed_gain_smooth(p, p->lsf_q[subframe],
1017 p->lsf_avg, p->cur_frame_mode);
1018
1019 synth_fixed_vector = anti_sparseness(p, &fixed_sparse, p->fixed_vector,
1020 synth_fixed_gain, spare_vector);
1021
1022 if (synthesis(p, p->lpc[subframe], synth_fixed_gain,
1023 synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 0))
1024 // overflow detected -> rerun synthesis scaling pitch vector down
1025 // by a factor of 4, skipping pitch vector contribution emphasis
1026 // and adaptive gain control
1027 synthesis(p, p->lpc[subframe], synth_fixed_gain,
1028 synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 1);
1029
1030 postfilter(p, p->lpc[subframe], buf_out + subframe * AMR_SUBFRAME_SIZE);
1031
1032 // update buffers and history
1033 ff_clear_fixed_vector(p->fixed_vector, &fixed_sparse, AMR_SUBFRAME_SIZE);
1034 update_state(p);
1035 }
1036
1037 ff_acelp_apply_order_2_transfer_function(buf_out, buf_out, highpass_zeros,
1038 highpass_poles,
1039 highpass_gain * AMR_SAMPLE_SCALE,
1040 p->high_pass_mem, AMR_BLOCK_SIZE);
1041
1042 /* Update averaged lsf vector (used for fixed gain smoothing).
1043 *
1044 * Note that lsf_avg should not incorporate the current frame's LSFs
1045 * for fixed_gain_smooth.
1046 * The specification has an incorrect formula: the reference decoder uses
1047 * qbar(n-1) rather than qbar(n) in section 6.1(4) equation 71. */
1048 ff_weighted_vector_sumf(p->lsf_avg, p->lsf_avg, p->lsf_q[3],
1049 0.84, 0.16, LP_FILTER_ORDER);
1050
1051 *got_frame_ptr = 1;
1052 *(AVFrame *)data = p->avframe;
1053
1054 /* return the amount of bytes consumed if everything was OK */
1055 return frame_sizes_nb[p->cur_frame_mode] + 1; // +7 for rounding and +8 for TOC
1056 }
1057
1058
1059 AVCodec ff_amrnb_decoder = {
1060 .name = "amrnb",
1061 .type = AVMEDIA_TYPE_AUDIO,
1062 .id = AV_CODEC_ID_AMR_NB,
1063 .priv_data_size = sizeof(AMRContext),
1064 .init = amrnb_decode_init,
1065 .decode = amrnb_decode_frame,
1066 .capabilities = CODEC_CAP_DR1,
1067 .long_name = NULL_IF_CONFIG_SMALL("AMR-NB (Adaptive Multi-Rate NarrowBand)"),
1068 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT,
1069 AV_SAMPLE_FMT_NONE },
1070 };