• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 "codec_internal.h"
59 #include "internal.h"
60 
61 #include "amrnbdata.h"
62 
63 #define AMR_BLOCK_SIZE              160   ///< samples per frame
64 #define AMR_SAMPLE_BOUND        32768.0   ///< threshold for synthesis overflow
65 
66 /**
67  * Scale from constructed speech to [-1,1]
68  *
69  * AMR is designed to produce 16-bit PCM samples (3GPP TS 26.090 4.2) but
70  * upscales by two (section 6.2.2).
71  *
72  * Fundamentally, this scale is determined by energy_mean through
73  * the fixed vector contribution to the excitation vector.
74  */
75 #define AMR_SAMPLE_SCALE  (2.0 / 32768.0)
76 
77 /** Prediction factor for 12.2kbit/s mode */
78 #define PRED_FAC_MODE_12k2             0.65
79 
80 #define LSF_R_FAC          (8000.0 / 32768.0) ///< LSF residual tables to Hertz
81 #define MIN_LSF_SPACING    (50.0488 / 8000.0) ///< Ensures stability of LPC filter
82 #define PITCH_LAG_MIN_MODE_12k2          18   ///< Lower bound on decoded lag search in 12.2kbit/s mode
83 
84 /** Initial energy in dB. Also used for bad frames (unimplemented). */
85 #define MIN_ENERGY -14.0
86 
87 /** Maximum sharpening factor
88  *
89  * The specification says 0.8, which should be 13107, but the reference C code
90  * uses 13017 instead. (Amusingly the same applies to SHARP_MAX in g729dec.c.)
91  */
92 #define SHARP_MAX 0.79449462890625
93 
94 /** Number of impulse response coefficients used for tilt factor */
95 #define AMR_TILT_RESPONSE   22
96 /** Tilt factor = 1st reflection coefficient * gamma_t */
97 #define AMR_TILT_GAMMA_T   0.8
98 /** Adaptive gain control factor used in post-filter */
99 #define AMR_AGC_ALPHA      0.9
100 
101 typedef struct AMRContext {
102     AMRNBFrame                        frame; ///< decoded AMR parameters (lsf coefficients, codebook indexes, etc)
103     uint8_t             bad_frame_indicator; ///< bad frame ? 1 : 0
104     enum Mode                cur_frame_mode;
105 
106     int16_t     prev_lsf_r[LP_FILTER_ORDER]; ///< residual LSF vector from previous subframe
107     double          lsp[4][LP_FILTER_ORDER]; ///< lsp vectors from current frame
108     double   prev_lsp_sub4[LP_FILTER_ORDER]; ///< lsp vector for the 4th subframe of the previous frame
109 
110     float         lsf_q[4][LP_FILTER_ORDER]; ///< Interpolated LSF vector for fixed gain smoothing
111     float          lsf_avg[LP_FILTER_ORDER]; ///< vector of averaged lsf vector
112 
113     float           lpc[4][LP_FILTER_ORDER]; ///< lpc coefficient vectors for 4 subframes
114 
115     uint8_t                   pitch_lag_int; ///< integer part of pitch lag from current subframe
116 
117     float excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1 + AMR_SUBFRAME_SIZE]; ///< current excitation and all necessary excitation history
118     float                       *excitation; ///< pointer to the current excitation vector in excitation_buf
119 
120     float   pitch_vector[AMR_SUBFRAME_SIZE]; ///< adaptive code book (pitch) vector
121     float   fixed_vector[AMR_SUBFRAME_SIZE]; ///< algebraic codebook (fixed) vector (must be kept zero between frames)
122 
123     float               prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes
124     float                     pitch_gain[5]; ///< quantified pitch gains for the current and previous four subframes
125     float                     fixed_gain[5]; ///< quantified fixed gains for the current and previous four subframes
126 
127     float                              beta; ///< previous pitch_gain, bounded by [0.0,SHARP_MAX]
128     uint8_t                      diff_count; ///< the number of subframes for which diff has been above 0.65
129     uint8_t                      hang_count; ///< the number of subframes since a hangover period started
130 
131     float            prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness processing to determine "onset"
132     uint8_t               prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none
133     uint8_t                 ir_filter_onset; ///< flag for impulse response filter strength
134 
135     float                postfilter_mem[10]; ///< previous intermediate values in the formant filter
136     float                          tilt_mem; ///< previous input to tilt compensation filter
137     float                    postfilter_agc; ///< previous factor used for adaptive gain control
138     float                  high_pass_mem[2]; ///< previous intermediate values in the high-pass filter
139 
140     float samples_in[LP_FILTER_ORDER + AMR_SUBFRAME_SIZE]; ///< floating point samples
141 
142     ACELPFContext                     acelpf_ctx; ///< context for filters for ACELP-based codecs
143     ACELPVContext                     acelpv_ctx; ///< context for vector operations for ACELP-based codecs
144     CELPFContext                       celpf_ctx; ///< context for filters for CELP-based codecs
145     CELPMContext                       celpm_ctx; ///< context for fixed point math operations
146 
147 } AMRContext;
148 
149 typedef struct AMRChannelsContext {
150     AMRContext ch[2];
151 } AMRChannelsContext;
152 
153 /** Double version of ff_weighted_vector_sumf() */
weighted_vector_sumd(double * out,const double * in_a,const double * in_b,double weight_coeff_a,double weight_coeff_b,int length)154 static void weighted_vector_sumd(double *out, const double *in_a,
155                                  const double *in_b, double weight_coeff_a,
156                                  double weight_coeff_b, int length)
157 {
158     int i;
159 
160     for (i = 0; i < length; i++)
161         out[i] = weight_coeff_a * in_a[i]
162                + weight_coeff_b * in_b[i];
163 }
164 
amrnb_decode_init(AVCodecContext * avctx)165 static av_cold int amrnb_decode_init(AVCodecContext *avctx)
166 {
167     AMRChannelsContext *s = avctx->priv_data;
168     int i;
169 
170     if (avctx->ch_layout.nb_channels > 2) {
171         avpriv_report_missing_feature(avctx, ">2 channel AMR");
172         return AVERROR_PATCHWELCOME;
173     }
174 
175     if (!avctx->ch_layout.nb_channels) {
176         av_channel_layout_uninit(&avctx->ch_layout);
177         avctx->ch_layout      = (AVChannelLayout)AV_CHANNEL_LAYOUT_MONO;
178     }
179     if (!avctx->sample_rate)
180         avctx->sample_rate = 8000;
181     avctx->sample_fmt     = AV_SAMPLE_FMT_FLTP;
182 
183     for (int ch = 0; ch < avctx->ch_layout.nb_channels; ch++) {
184         AMRContext *p = &s->ch[ch];
185         // p->excitation always points to the same position in p->excitation_buf
186         p->excitation = &p->excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1];
187 
188         for (i = 0; i < LP_FILTER_ORDER; i++) {
189             p->prev_lsp_sub4[i] =    lsp_sub4_init[i] * 1000 / (float)(1 << 15);
190             p->lsf_avg[i] = p->lsf_q[3][i] = lsp_avg_init[i] / (float)(1 << 15);
191         }
192 
193         for (i = 0; i < 4; i++)
194             p->prediction_error[i] = MIN_ENERGY;
195 
196         ff_acelp_filter_init(&p->acelpf_ctx);
197         ff_acelp_vectors_init(&p->acelpv_ctx);
198         ff_celp_filter_init(&p->celpf_ctx);
199         ff_celp_math_init(&p->celpm_ctx);
200     }
201 
202     return 0;
203 }
204 
205 
206 /**
207  * Unpack an RFC4867 speech frame into the AMR frame mode and parameters.
208  *
209  * The order of speech bits is specified by 3GPP TS 26.101.
210  *
211  * @param p the context
212  * @param buf               pointer to the input buffer
213  * @param buf_size          size of the input buffer
214  *
215  * @return the frame mode
216  */
unpack_bitstream(AMRContext * p,const uint8_t * buf,int buf_size)217 static enum Mode unpack_bitstream(AMRContext *p, const uint8_t *buf,
218                                   int buf_size)
219 {
220     enum Mode mode;
221 
222     // Decode the first octet.
223     mode = buf[0] >> 3 & 0x0F;                      // frame type
224     p->bad_frame_indicator = (buf[0] & 0x4) != 0x4; // quality bit
225 
226     if (mode >= N_MODES || buf_size < frame_sizes_nb[mode] + 1) {
227         return NO_DATA;
228     }
229 
230     if (mode < MODE_DTX)
231         ff_amr_bit_reorder((uint16_t *) &p->frame, sizeof(AMRNBFrame), buf + 1,
232                            amr_unpacking_bitmaps_per_mode[mode]);
233 
234     return mode;
235 }
236 
237 
238 /// @name AMR pitch LPC coefficient decoding functions
239 /// @{
240 
241 /**
242  * Interpolate the LSF vector (used for fixed gain smoothing).
243  * The interpolation is done over all four subframes even in MODE_12k2.
244  *
245  * @param[in]     ctx       The Context
246  * @param[in,out] lsf_q     LSFs in [0,1] for each subframe
247  * @param[in]     lsf_new   New LSFs in [0,1] for subframe 4
248  */
interpolate_lsf(ACELPVContext * ctx,float lsf_q[4][LP_FILTER_ORDER],float * lsf_new)249 static void interpolate_lsf(ACELPVContext *ctx, float lsf_q[4][LP_FILTER_ORDER], float *lsf_new)
250 {
251     int i;
252 
253     for (i = 0; i < 4; i++)
254         ctx->weighted_vector_sumf(lsf_q[i], lsf_q[3], lsf_new,
255                                 0.25 * (3 - i), 0.25 * (i + 1),
256                                 LP_FILTER_ORDER);
257 }
258 
259 /**
260  * Decode a set of 5 split-matrix quantized lsf indexes into an lsp vector.
261  *
262  * @param p the context
263  * @param lsp output LSP vector
264  * @param lsf_no_r LSF vector without the residual vector added
265  * @param lsf_quantizer pointers to LSF dictionary tables
266  * @param quantizer_offset offset in tables
267  * @param sign for the 3 dictionary table
268  * @param update store data for computing the next frame's LSFs
269  */
lsf2lsp_for_mode12k2(AMRContext * p,double lsp[LP_FILTER_ORDER],const float lsf_no_r[LP_FILTER_ORDER],const int16_t * lsf_quantizer[5],const int quantizer_offset,const int sign,const int update)270 static void lsf2lsp_for_mode12k2(AMRContext *p, double lsp[LP_FILTER_ORDER],
271                                  const float lsf_no_r[LP_FILTER_ORDER],
272                                  const int16_t *lsf_quantizer[5],
273                                  const int quantizer_offset,
274                                  const int sign, const int update)
275 {
276     int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector
277     float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector
278     int i;
279 
280     for (i = 0; i < LP_FILTER_ORDER >> 1; i++)
281         memcpy(&lsf_r[i << 1], &lsf_quantizer[i][quantizer_offset],
282                2 * sizeof(*lsf_r));
283 
284     if (sign) {
285         lsf_r[4] *= -1;
286         lsf_r[5] *= -1;
287     }
288 
289     if (update)
290         memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r));
291 
292     for (i = 0; i < LP_FILTER_ORDER; i++)
293         lsf_q[i] = lsf_r[i] * (LSF_R_FAC / 8000.0) + lsf_no_r[i] * (1.0 / 8000.0);
294 
295     ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
296 
297     if (update)
298         interpolate_lsf(&p->acelpv_ctx, p->lsf_q, lsf_q);
299 
300     ff_acelp_lsf2lspd(lsp, lsf_q, LP_FILTER_ORDER);
301 }
302 
303 /**
304  * Decode a set of 5 split-matrix quantized lsf indexes into 2 lsp vectors.
305  *
306  * @param p                 pointer to the AMRContext
307  */
lsf2lsp_5(AMRContext * p)308 static void lsf2lsp_5(AMRContext *p)
309 {
310     const uint16_t *lsf_param = p->frame.lsf;
311     float lsf_no_r[LP_FILTER_ORDER]; // LSFs without the residual vector
312     const int16_t *lsf_quantizer[5];
313     int i;
314 
315     lsf_quantizer[0] = lsf_5_1[lsf_param[0]];
316     lsf_quantizer[1] = lsf_5_2[lsf_param[1]];
317     lsf_quantizer[2] = lsf_5_3[lsf_param[2] >> 1];
318     lsf_quantizer[3] = lsf_5_4[lsf_param[3]];
319     lsf_quantizer[4] = lsf_5_5[lsf_param[4]];
320 
321     for (i = 0; i < LP_FILTER_ORDER; i++)
322         lsf_no_r[i] = p->prev_lsf_r[i] * LSF_R_FAC * PRED_FAC_MODE_12k2 + lsf_5_mean[i];
323 
324     lsf2lsp_for_mode12k2(p, p->lsp[1], lsf_no_r, lsf_quantizer, 0, lsf_param[2] & 1, 0);
325     lsf2lsp_for_mode12k2(p, p->lsp[3], lsf_no_r, lsf_quantizer, 2, lsf_param[2] & 1, 1);
326 
327     // interpolate LSP vectors at subframes 1 and 3
328     weighted_vector_sumd(p->lsp[0], p->prev_lsp_sub4, p->lsp[1], 0.5, 0.5, LP_FILTER_ORDER);
329     weighted_vector_sumd(p->lsp[2], p->lsp[1]       , p->lsp[3], 0.5, 0.5, LP_FILTER_ORDER);
330 }
331 
332 /**
333  * Decode a set of 3 split-matrix quantized lsf indexes into an lsp vector.
334  *
335  * @param p                 pointer to the AMRContext
336  */
lsf2lsp_3(AMRContext * p)337 static void lsf2lsp_3(AMRContext *p)
338 {
339     const uint16_t *lsf_param = p->frame.lsf;
340     int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector
341     float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector
342     const int16_t *lsf_quantizer;
343     int i, j;
344 
345     lsf_quantizer = (p->cur_frame_mode == MODE_7k95 ? lsf_3_1_MODE_7k95 : lsf_3_1)[lsf_param[0]];
346     memcpy(lsf_r, lsf_quantizer, 3 * sizeof(*lsf_r));
347 
348     lsf_quantizer = lsf_3_2[lsf_param[1] << (p->cur_frame_mode <= MODE_5k15)];
349     memcpy(lsf_r + 3, lsf_quantizer, 3 * sizeof(*lsf_r));
350 
351     lsf_quantizer = (p->cur_frame_mode <= MODE_5k15 ? lsf_3_3_MODE_5k15 : lsf_3_3)[lsf_param[2]];
352     memcpy(lsf_r + 6, lsf_quantizer, 4 * sizeof(*lsf_r));
353 
354     // calculate mean-removed LSF vector and add mean
355     for (i = 0; i < LP_FILTER_ORDER; i++)
356         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);
357 
358     ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
359 
360     // store data for computing the next frame's LSFs
361     interpolate_lsf(&p->acelpv_ctx, p->lsf_q, lsf_q);
362     memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r));
363 
364     ff_acelp_lsf2lspd(p->lsp[3], lsf_q, LP_FILTER_ORDER);
365 
366     // interpolate LSP vectors at subframes 1, 2 and 3
367     for (i = 1; i <= 3; i++)
368         for(j = 0; j < LP_FILTER_ORDER; j++)
369             p->lsp[i-1][j] = p->prev_lsp_sub4[j] +
370                 (p->lsp[3][j] - p->prev_lsp_sub4[j]) * 0.25 * i;
371 }
372 
373 /// @}
374 
375 
376 /// @name AMR pitch vector decoding functions
377 /// @{
378 
379 /**
380  * Like ff_decode_pitch_lag(), but with 1/6 resolution
381  */
decode_pitch_lag_1_6(int * lag_int,int * lag_frac,int pitch_index,const int prev_lag_int,const int subframe)382 static void decode_pitch_lag_1_6(int *lag_int, int *lag_frac, int pitch_index,
383                                  const int prev_lag_int, const int subframe)
384 {
385     if (subframe == 0 || subframe == 2) {
386         if (pitch_index < 463) {
387             *lag_int  = (pitch_index + 107) * 10923 >> 16;
388             *lag_frac = pitch_index - *lag_int * 6 + 105;
389         } else {
390             *lag_int  = pitch_index - 368;
391             *lag_frac = 0;
392         }
393     } else {
394         *lag_int  = ((pitch_index + 5) * 10923 >> 16) - 1;
395         *lag_frac = pitch_index - *lag_int * 6 - 3;
396         *lag_int += av_clip(prev_lag_int - 5, PITCH_LAG_MIN_MODE_12k2,
397                             PITCH_DELAY_MAX - 9);
398     }
399 }
400 
decode_pitch_vector(AMRContext * p,const AMRNBSubframe * amr_subframe,const int subframe)401 static void decode_pitch_vector(AMRContext *p,
402                                 const AMRNBSubframe *amr_subframe,
403                                 const int subframe)
404 {
405     int pitch_lag_int, pitch_lag_frac;
406     enum Mode mode = p->cur_frame_mode;
407 
408     if (p->cur_frame_mode == MODE_12k2) {
409         decode_pitch_lag_1_6(&pitch_lag_int, &pitch_lag_frac,
410                              amr_subframe->p_lag, p->pitch_lag_int,
411                              subframe);
412     } else {
413         ff_decode_pitch_lag(&pitch_lag_int, &pitch_lag_frac,
414                             amr_subframe->p_lag,
415                             p->pitch_lag_int, subframe,
416                             mode != MODE_4k75 && mode != MODE_5k15,
417                             mode <= MODE_6k7 ? 4 : (mode == MODE_7k95 ? 5 : 6));
418         pitch_lag_frac *= 2;
419     }
420 
421     p->pitch_lag_int = pitch_lag_int; // store previous lag in a uint8_t
422 
423     pitch_lag_int += pitch_lag_frac > 0;
424 
425     /* Calculate the pitch vector by interpolating the past excitation at the
426        pitch lag using a b60 hamming windowed sinc function.   */
427     p->acelpf_ctx.acelp_interpolatef(p->excitation,
428                           p->excitation + 1 - pitch_lag_int,
429                           ff_b60_sinc, 6,
430                           pitch_lag_frac + 6 - 6*(pitch_lag_frac > 0),
431                           10, AMR_SUBFRAME_SIZE);
432 
433     memcpy(p->pitch_vector, p->excitation, AMR_SUBFRAME_SIZE * sizeof(float));
434 }
435 
436 /// @}
437 
438 
439 /// @name AMR algebraic code book (fixed) vector decoding functions
440 /// @{
441 
442 /**
443  * Decode a 10-bit algebraic codebook index from a 10.2 kbit/s frame.
444  */
decode_10bit_pulse(int code,int pulse_position[8],int i1,int i2,int i3)445 static void decode_10bit_pulse(int code, int pulse_position[8],
446                                int i1, int i2, int i3)
447 {
448     // coded using 7+3 bits with the 3 LSBs being, individually, the LSB of 1 of
449     // the 3 pulses and the upper 7 bits being coded in base 5
450     const uint8_t *positions = base_five_table[code >> 3];
451     pulse_position[i1] = (positions[2] << 1) + ( code       & 1);
452     pulse_position[i2] = (positions[1] << 1) + ((code >> 1) & 1);
453     pulse_position[i3] = (positions[0] << 1) + ((code >> 2) & 1);
454 }
455 
456 /**
457  * Decode the algebraic codebook index to pulse positions and signs and
458  * construct the algebraic codebook vector for MODE_10k2.
459  *
460  * @param fixed_index          positions of the eight pulses
461  * @param fixed_sparse         pointer to the algebraic codebook vector
462  */
decode_8_pulses_31bits(const int16_t * fixed_index,AMRFixed * fixed_sparse)463 static void decode_8_pulses_31bits(const int16_t *fixed_index,
464                                    AMRFixed *fixed_sparse)
465 {
466     int pulse_position[8];
467     int i, temp;
468 
469     decode_10bit_pulse(fixed_index[4], pulse_position, 0, 4, 1);
470     decode_10bit_pulse(fixed_index[5], pulse_position, 2, 6, 5);
471 
472     // coded using 5+2 bits with the 2 LSBs being, individually, the LSB of 1 of
473     // the 2 pulses and the upper 5 bits being coded in base 5
474     temp = ((fixed_index[6] >> 2) * 25 + 12) >> 5;
475     pulse_position[3] = temp % 5;
476     pulse_position[7] = temp / 5;
477     if (pulse_position[7] & 1)
478         pulse_position[3] = 4 - pulse_position[3];
479     pulse_position[3] = (pulse_position[3] << 1) + ( fixed_index[6]       & 1);
480     pulse_position[7] = (pulse_position[7] << 1) + ((fixed_index[6] >> 1) & 1);
481 
482     fixed_sparse->n = 8;
483     for (i = 0; i < 4; i++) {
484         const int pos1   = (pulse_position[i]     << 2) + i;
485         const int pos2   = (pulse_position[i + 4] << 2) + i;
486         const float sign = fixed_index[i] ? -1.0 : 1.0;
487         fixed_sparse->x[i    ] = pos1;
488         fixed_sparse->x[i + 4] = pos2;
489         fixed_sparse->y[i    ] = sign;
490         fixed_sparse->y[i + 4] = pos2 < pos1 ? -sign : sign;
491     }
492 }
493 
494 /**
495  * Decode the algebraic codebook index to pulse positions and signs,
496  * then construct the algebraic codebook vector.
497  *
498  *                              nb of pulses | bits encoding pulses
499  * For MODE_4k75 or MODE_5k15,             2 | 1-3, 4-6, 7
500  *                  MODE_5k9,              2 | 1,   2-4, 5-6, 7-9
501  *                  MODE_6k7,              3 | 1-3, 4,   5-7, 8,  9-11
502  *      MODE_7k4 or MODE_7k95,             4 | 1-3, 4-6, 7-9, 10, 11-13
503  *
504  * @param fixed_sparse pointer to the algebraic codebook vector
505  * @param pulses       algebraic codebook indexes
506  * @param mode         mode of the current frame
507  * @param subframe     current subframe number
508  */
decode_fixed_sparse(AMRFixed * fixed_sparse,const uint16_t * pulses,const enum Mode mode,const int subframe)509 static void decode_fixed_sparse(AMRFixed *fixed_sparse, const uint16_t *pulses,
510                                 const enum Mode mode, const int subframe)
511 {
512     av_assert1(MODE_4k75 <= (signed)mode && mode <= MODE_12k2);
513 
514     if (mode == MODE_12k2) {
515         ff_decode_10_pulses_35bits(pulses, fixed_sparse, gray_decode, 5, 3);
516     } else if (mode == MODE_10k2) {
517         decode_8_pulses_31bits(pulses, fixed_sparse);
518     } else {
519         int *pulse_position = fixed_sparse->x;
520         int i, pulse_subset;
521         const int fixed_index = pulses[0];
522 
523         if (mode <= MODE_5k15) {
524             pulse_subset      = ((fixed_index >> 3) & 8)     + (subframe << 1);
525             pulse_position[0] = ( fixed_index       & 7) * 5 + track_position[pulse_subset];
526             pulse_position[1] = ((fixed_index >> 3) & 7) * 5 + track_position[pulse_subset + 1];
527             fixed_sparse->n = 2;
528         } else if (mode == MODE_5k9) {
529             pulse_subset      = ((fixed_index & 1) << 1) + 1;
530             pulse_position[0] = ((fixed_index >> 1) & 7) * 5 + pulse_subset;
531             pulse_subset      = (fixed_index  >> 4) & 3;
532             pulse_position[1] = ((fixed_index >> 6) & 7) * 5 + pulse_subset + (pulse_subset == 3 ? 1 : 0);
533             fixed_sparse->n = pulse_position[0] == pulse_position[1] ? 1 : 2;
534         } else if (mode == MODE_6k7) {
535             pulse_position[0] = (fixed_index        & 7) * 5;
536             pulse_subset      = (fixed_index  >> 2) & 2;
537             pulse_position[1] = ((fixed_index >> 4) & 7) * 5 + pulse_subset + 1;
538             pulse_subset      = (fixed_index  >> 6) & 2;
539             pulse_position[2] = ((fixed_index >> 8) & 7) * 5 + pulse_subset + 2;
540             fixed_sparse->n = 3;
541         } else { // mode <= MODE_7k95
542             pulse_position[0] = gray_decode[ fixed_index        & 7];
543             pulse_position[1] = gray_decode[(fixed_index >> 3)  & 7] + 1;
544             pulse_position[2] = gray_decode[(fixed_index >> 6)  & 7] + 2;
545             pulse_subset      = (fixed_index >> 9) & 1;
546             pulse_position[3] = gray_decode[(fixed_index >> 10) & 7] + pulse_subset + 3;
547             fixed_sparse->n = 4;
548         }
549         for (i = 0; i < fixed_sparse->n; i++)
550             fixed_sparse->y[i] = (pulses[1] >> i) & 1 ? 1.0 : -1.0;
551     }
552 }
553 
554 /**
555  * Apply pitch lag to obtain the sharpened fixed vector (section 6.1.2)
556  *
557  * @param p the context
558  * @param subframe unpacked amr subframe
559  * @param mode mode of the current frame
560  * @param fixed_sparse sparse representation of the fixed vector
561  */
pitch_sharpening(AMRContext * p,int subframe,enum Mode mode,AMRFixed * fixed_sparse)562 static void pitch_sharpening(AMRContext *p, int subframe, enum Mode mode,
563                              AMRFixed *fixed_sparse)
564 {
565     // The spec suggests the current pitch gain is always used, but in other
566     // modes the pitch and codebook gains are jointly quantized (sec 5.8.2)
567     // so the codebook gain cannot depend on the quantized pitch gain.
568     if (mode == MODE_12k2)
569         p->beta = FFMIN(p->pitch_gain[4], 1.0);
570 
571     fixed_sparse->pitch_lag  = p->pitch_lag_int;
572     fixed_sparse->pitch_fac  = p->beta;
573 
574     // Save pitch sharpening factor for the next subframe
575     // MODE_4k75 only updates on the 2nd and 4th subframes - this follows from
576     // the fact that the gains for two subframes are jointly quantized.
577     if (mode != MODE_4k75 || subframe & 1)
578         p->beta = av_clipf(p->pitch_gain[4], 0.0, SHARP_MAX);
579 }
580 /// @}
581 
582 
583 /// @name AMR gain decoding functions
584 /// @{
585 
586 /**
587  * fixed gain smoothing
588  * Note that where the spec specifies the "spectrum in the q domain"
589  * in section 6.1.4, in fact frequencies should be used.
590  *
591  * @param p the context
592  * @param lsf LSFs for the current subframe, in the range [0,1]
593  * @param lsf_avg averaged LSFs
594  * @param mode mode of the current frame
595  *
596  * @return fixed gain smoothed
597  */
fixed_gain_smooth(AMRContext * p,const float * lsf,const float * lsf_avg,const enum Mode mode)598 static float fixed_gain_smooth(AMRContext *p , const float *lsf,
599                                const float *lsf_avg, const enum Mode mode)
600 {
601     float diff = 0.0;
602     int i;
603 
604     for (i = 0; i < LP_FILTER_ORDER; i++)
605         diff += fabs(lsf_avg[i] - lsf[i]) / lsf_avg[i];
606 
607     // If diff is large for ten subframes, disable smoothing for a 40-subframe
608     // hangover period.
609     p->diff_count++;
610     if (diff <= 0.65)
611         p->diff_count = 0;
612 
613     if (p->diff_count > 10) {
614         p->hang_count = 0;
615         p->diff_count--; // don't let diff_count overflow
616     }
617 
618     if (p->hang_count < 40) {
619         p->hang_count++;
620     } else if (mode < MODE_7k4 || mode == MODE_10k2) {
621         const float smoothing_factor = av_clipf(4.0 * diff - 1.6, 0.0, 1.0);
622         const float fixed_gain_mean = (p->fixed_gain[0] + p->fixed_gain[1] +
623                                        p->fixed_gain[2] + p->fixed_gain[3] +
624                                        p->fixed_gain[4]) * 0.2;
625         return smoothing_factor * p->fixed_gain[4] +
626                (1.0 - smoothing_factor) * fixed_gain_mean;
627     }
628     return p->fixed_gain[4];
629 }
630 
631 /**
632  * Decode pitch gain and fixed gain factor (part of section 6.1.3).
633  *
634  * @param p the context
635  * @param amr_subframe unpacked amr subframe
636  * @param mode mode of the current frame
637  * @param subframe current subframe number
638  * @param fixed_gain_factor decoded gain correction factor
639  */
decode_gains(AMRContext * p,const AMRNBSubframe * amr_subframe,const enum Mode mode,const int subframe,float * fixed_gain_factor)640 static void decode_gains(AMRContext *p, const AMRNBSubframe *amr_subframe,
641                          const enum Mode mode, const int subframe,
642                          float *fixed_gain_factor)
643 {
644     if (mode == MODE_12k2 || mode == MODE_7k95) {
645         p->pitch_gain[4]   = qua_gain_pit [amr_subframe->p_gain    ]
646             * (1.0 / 16384.0);
647         *fixed_gain_factor = qua_gain_code[amr_subframe->fixed_gain]
648             * (1.0 /  2048.0);
649     } else {
650         const uint16_t *gains;
651 
652         if (mode >= MODE_6k7) {
653             gains = gains_high[amr_subframe->p_gain];
654         } else if (mode >= MODE_5k15) {
655             gains = gains_low [amr_subframe->p_gain];
656         } else {
657             // gain index is only coded in subframes 0,2 for MODE_4k75
658             gains = gains_MODE_4k75[(p->frame.subframe[subframe & 2].p_gain << 1) + (subframe & 1)];
659         }
660 
661         p->pitch_gain[4]   = gains[0] * (1.0 / 16384.0);
662         *fixed_gain_factor = gains[1] * (1.0 /  4096.0);
663     }
664 }
665 
666 /// @}
667 
668 
669 /// @name AMR preprocessing functions
670 /// @{
671 
672 /**
673  * Circularly convolve a sparse fixed vector with a phase dispersion impulse
674  * response filter (D.6.2 of G.729 and 6.1.5 of AMR).
675  *
676  * @param out vector with filter applied
677  * @param in source vector
678  * @param filter phase filter coefficients
679  *
680  *  out[n] = sum(i,0,len-1){ in[i] * filter[(len + n - i)%len] }
681  */
apply_ir_filter(float * out,const AMRFixed * in,const float * filter)682 static void apply_ir_filter(float *out, const AMRFixed *in,
683                             const float *filter)
684 {
685     float filter1[AMR_SUBFRAME_SIZE],     ///< filters at pitch lag*1 and *2
686           filter2[AMR_SUBFRAME_SIZE];
687     int   lag = in->pitch_lag;
688     float fac = in->pitch_fac;
689     int i;
690 
691     if (lag < AMR_SUBFRAME_SIZE) {
692         ff_celp_circ_addf(filter1, filter, filter, lag, fac,
693                           AMR_SUBFRAME_SIZE);
694 
695         if (lag < AMR_SUBFRAME_SIZE >> 1)
696             ff_celp_circ_addf(filter2, filter, filter1, lag, fac,
697                               AMR_SUBFRAME_SIZE);
698     }
699 
700     memset(out, 0, sizeof(float) * AMR_SUBFRAME_SIZE);
701     for (i = 0; i < in->n; i++) {
702         int   x = in->x[i];
703         float y = in->y[i];
704         const float *filterp;
705 
706         if (x >= AMR_SUBFRAME_SIZE - lag) {
707             filterp = filter;
708         } else if (x >= AMR_SUBFRAME_SIZE - (lag << 1)) {
709             filterp = filter1;
710         } else
711             filterp = filter2;
712 
713         ff_celp_circ_addf(out, out, filterp, x, y, AMR_SUBFRAME_SIZE);
714     }
715 }
716 
717 /**
718  * Reduce fixed vector sparseness by smoothing with one of three IR filters.
719  * Also know as "adaptive phase dispersion".
720  *
721  * This implements 3GPP TS 26.090 section 6.1(5).
722  *
723  * @param p the context
724  * @param fixed_sparse algebraic codebook vector
725  * @param fixed_vector unfiltered fixed vector
726  * @param fixed_gain smoothed gain
727  * @param out space for modified vector if necessary
728  */
anti_sparseness(AMRContext * p,AMRFixed * fixed_sparse,const float * fixed_vector,float fixed_gain,float * out)729 static const float *anti_sparseness(AMRContext *p, AMRFixed *fixed_sparse,
730                                     const float *fixed_vector,
731                                     float fixed_gain, float *out)
732 {
733     int ir_filter_nr;
734 
735     if (p->pitch_gain[4] < 0.6) {
736         ir_filter_nr = 0;      // strong filtering
737     } else if (p->pitch_gain[4] < 0.9) {
738         ir_filter_nr = 1;      // medium filtering
739     } else
740         ir_filter_nr = 2;      // no filtering
741 
742     // detect 'onset'
743     if (fixed_gain > 2.0 * p->prev_sparse_fixed_gain) {
744         p->ir_filter_onset = 2;
745     } else if (p->ir_filter_onset)
746         p->ir_filter_onset--;
747 
748     if (!p->ir_filter_onset) {
749         int i, count = 0;
750 
751         for (i = 0; i < 5; i++)
752             if (p->pitch_gain[i] < 0.6)
753                 count++;
754         if (count > 2)
755             ir_filter_nr = 0;
756 
757         if (ir_filter_nr > p->prev_ir_filter_nr + 1)
758             ir_filter_nr--;
759     } else if (ir_filter_nr < 2)
760         ir_filter_nr++;
761 
762     // Disable filtering for very low level of fixed_gain.
763     // Note this step is not specified in the technical description but is in
764     // the reference source in the function Ph_disp.
765     if (fixed_gain < 5.0)
766         ir_filter_nr = 2;
767 
768     if (p->cur_frame_mode != MODE_7k4 && p->cur_frame_mode < MODE_10k2
769          && ir_filter_nr < 2) {
770         apply_ir_filter(out, fixed_sparse,
771                         (p->cur_frame_mode == MODE_7k95 ?
772                              ir_filters_lookup_MODE_7k95 :
773                              ir_filters_lookup)[ir_filter_nr]);
774         fixed_vector = out;
775     }
776 
777     // update ir filter strength history
778     p->prev_ir_filter_nr       = ir_filter_nr;
779     p->prev_sparse_fixed_gain  = fixed_gain;
780 
781     return fixed_vector;
782 }
783 
784 /// @}
785 
786 
787 /// @name AMR synthesis functions
788 /// @{
789 
790 /**
791  * Conduct 10th order linear predictive coding synthesis.
792  *
793  * @param p             pointer to the AMRContext
794  * @param lpc           pointer to the LPC coefficients
795  * @param fixed_gain    fixed codebook gain for synthesis
796  * @param fixed_vector  algebraic codebook vector
797  * @param samples       pointer to the output speech samples
798  * @param overflow      16-bit overflow flag
799  */
synthesis(AMRContext * p,float * lpc,float fixed_gain,const float * fixed_vector,float * samples,uint8_t overflow)800 static int synthesis(AMRContext *p, float *lpc,
801                      float fixed_gain, const float *fixed_vector,
802                      float *samples, uint8_t overflow)
803 {
804     int i;
805     float excitation[AMR_SUBFRAME_SIZE];
806 
807     // if an overflow has been detected, the pitch vector is scaled down by a
808     // factor of 4
809     if (overflow)
810         for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
811             p->pitch_vector[i] *= 0.25;
812 
813     p->acelpv_ctx.weighted_vector_sumf(excitation, p->pitch_vector, fixed_vector,
814                             p->pitch_gain[4], fixed_gain, AMR_SUBFRAME_SIZE);
815 
816     // emphasize pitch vector contribution
817     if (p->pitch_gain[4] > 0.5 && !overflow) {
818         float energy = p->celpm_ctx.dot_productf(excitation, excitation,
819                                                     AMR_SUBFRAME_SIZE);
820         float pitch_factor =
821             p->pitch_gain[4] *
822             (p->cur_frame_mode == MODE_12k2 ?
823                 0.25 * FFMIN(p->pitch_gain[4], 1.0) :
824                 0.5  * FFMIN(p->pitch_gain[4], SHARP_MAX));
825 
826         for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
827             excitation[i] += pitch_factor * p->pitch_vector[i];
828 
829         ff_scale_vector_to_given_sum_of_squares(excitation, excitation, energy,
830                                                 AMR_SUBFRAME_SIZE);
831     }
832 
833     p->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation,
834                                  AMR_SUBFRAME_SIZE,
835                                  LP_FILTER_ORDER);
836 
837     // detect overflow
838     for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
839         if (fabsf(samples[i]) > AMR_SAMPLE_BOUND) {
840             return 1;
841         }
842 
843     return 0;
844 }
845 
846 /// @}
847 
848 
849 /// @name AMR update functions
850 /// @{
851 
852 /**
853  * Update buffers and history at the end of decoding a subframe.
854  *
855  * @param p             pointer to the AMRContext
856  */
update_state(AMRContext * p)857 static void update_state(AMRContext *p)
858 {
859     memcpy(p->prev_lsp_sub4, p->lsp[3], LP_FILTER_ORDER * sizeof(p->lsp[3][0]));
860 
861     memmove(&p->excitation_buf[0], &p->excitation_buf[AMR_SUBFRAME_SIZE],
862             (PITCH_DELAY_MAX + LP_FILTER_ORDER + 1) * sizeof(float));
863 
864     memmove(&p->pitch_gain[0], &p->pitch_gain[1], 4 * sizeof(float));
865     memmove(&p->fixed_gain[0], &p->fixed_gain[1], 4 * sizeof(float));
866 
867     memmove(&p->samples_in[0], &p->samples_in[AMR_SUBFRAME_SIZE],
868             LP_FILTER_ORDER * sizeof(float));
869 }
870 
871 /// @}
872 
873 
874 /// @name AMR Postprocessing functions
875 /// @{
876 
877 /**
878  * Get the tilt factor of a formant filter from its transfer function
879  *
880  * @param p     The Context
881  * @param lpc_n LP_FILTER_ORDER coefficients of the numerator
882  * @param lpc_d LP_FILTER_ORDER coefficients of the denominator
883  */
tilt_factor(AMRContext * p,float * lpc_n,float * lpc_d)884 static float tilt_factor(AMRContext *p, float *lpc_n, float *lpc_d)
885 {
886     float rh0, rh1; // autocorrelation at lag 0 and 1
887 
888     // LP_FILTER_ORDER prior zeros are needed for ff_celp_lp_synthesis_filterf
889     float impulse_buffer[LP_FILTER_ORDER + AMR_TILT_RESPONSE] = { 0 };
890     float *hf = impulse_buffer + LP_FILTER_ORDER; // start of impulse response
891 
892     hf[0] = 1.0;
893     memcpy(hf + 1, lpc_n, sizeof(float) * LP_FILTER_ORDER);
894     p->celpf_ctx.celp_lp_synthesis_filterf(hf, lpc_d, hf,
895                                  AMR_TILT_RESPONSE,
896                                  LP_FILTER_ORDER);
897 
898     rh0 = p->celpm_ctx.dot_productf(hf, hf,     AMR_TILT_RESPONSE);
899     rh1 = p->celpm_ctx.dot_productf(hf, hf + 1, AMR_TILT_RESPONSE - 1);
900 
901     // The spec only specifies this check for 12.2 and 10.2 kbit/s
902     // modes. But in the ref source the tilt is always non-negative.
903     return rh1 >= 0.0 ? rh1 / rh0 * AMR_TILT_GAMMA_T : 0.0;
904 }
905 
906 /**
907  * Perform adaptive post-filtering to enhance the quality of the speech.
908  * See section 6.2.1.
909  *
910  * @param p             pointer to the AMRContext
911  * @param lpc           interpolated LP coefficients for this subframe
912  * @param buf_out       output of the filter
913  */
postfilter(AMRContext * p,float * lpc,float * buf_out)914 static void postfilter(AMRContext *p, float *lpc, float *buf_out)
915 {
916     int i;
917     float *samples          = p->samples_in + LP_FILTER_ORDER; // Start of input
918 
919     float speech_gain       = p->celpm_ctx.dot_productf(samples, samples,
920                                                            AMR_SUBFRAME_SIZE);
921 
922     float pole_out[AMR_SUBFRAME_SIZE + LP_FILTER_ORDER];  // Output of pole filter
923     const float *gamma_n, *gamma_d;                       // Formant filter factor table
924     float lpc_n[LP_FILTER_ORDER], lpc_d[LP_FILTER_ORDER]; // Transfer function coefficients
925 
926     if (p->cur_frame_mode == MODE_12k2 || p->cur_frame_mode == MODE_10k2) {
927         gamma_n = ff_pow_0_7;
928         gamma_d = ff_pow_0_75;
929     } else {
930         gamma_n = ff_pow_0_55;
931         gamma_d = ff_pow_0_7;
932     }
933 
934     for (i = 0; i < LP_FILTER_ORDER; i++) {
935          lpc_n[i] = lpc[i] * gamma_n[i];
936          lpc_d[i] = lpc[i] * gamma_d[i];
937     }
938 
939     memcpy(pole_out, p->postfilter_mem, sizeof(float) * LP_FILTER_ORDER);
940     p->celpf_ctx.celp_lp_synthesis_filterf(pole_out + LP_FILTER_ORDER, lpc_d, samples,
941                                  AMR_SUBFRAME_SIZE, LP_FILTER_ORDER);
942     memcpy(p->postfilter_mem, pole_out + AMR_SUBFRAME_SIZE,
943            sizeof(float) * LP_FILTER_ORDER);
944 
945     p->celpf_ctx.celp_lp_zero_synthesis_filterf(buf_out, lpc_n,
946                                       pole_out + LP_FILTER_ORDER,
947                                       AMR_SUBFRAME_SIZE, LP_FILTER_ORDER);
948 
949     ff_tilt_compensation(&p->tilt_mem, tilt_factor(p, lpc_n, lpc_d), buf_out,
950                          AMR_SUBFRAME_SIZE);
951 
952     ff_adaptive_gain_control(buf_out, buf_out, speech_gain, AMR_SUBFRAME_SIZE,
953                              AMR_AGC_ALPHA, &p->postfilter_agc);
954 }
955 
956 /// @}
957 
amrnb_decode_frame(AVCodecContext * avctx,AVFrame * frame,int * got_frame_ptr,AVPacket * avpkt)958 static int amrnb_decode_frame(AVCodecContext *avctx, AVFrame *frame,
959                               int *got_frame_ptr, AVPacket *avpkt)
960 {
961 
962     AMRChannelsContext *s = avctx->priv_data;        // pointer to private data
963     const uint8_t *buf = avpkt->data;
964     int buf_size       = avpkt->size;
965     int ret;
966 
967     /* get output buffer */
968     frame->nb_samples = AMR_BLOCK_SIZE;
969     if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
970         return ret;
971 
972     for (int ch = 0; ch < avctx->ch_layout.nb_channels; ch++) {
973         AMRContext *p = &s->ch[ch];
974         float fixed_gain_factor;
975         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         float *buf_out = (float *)frame->extended_data[ch];
980         int channel_size;
981         int i, subframe;
982 
983         p->cur_frame_mode = unpack_bitstream(p, buf, buf_size);
984         if (p->cur_frame_mode == NO_DATA) {
985             av_log(avctx, AV_LOG_ERROR, "Corrupt bitstream\n");
986             return AVERROR_INVALIDDATA;
987         }
988         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         channel_size = frame_sizes_nb[p->cur_frame_mode] + 1; // +7 for rounding and +8 for TOC
995         if (p->cur_frame_mode == MODE_12k2) {
996             lsf2lsp_5(p);
997         } else
998             lsf2lsp_3(p);
999 
1000         for (i = 0; i < 4; i++)
1001             ff_acelp_lspd2lpc(p->lsp[i], p->lpc[i], 5);
1002 
1003         for (subframe = 0; subframe < 4; subframe++) {
1004             const AMRNBSubframe *amr_subframe = &p->frame.subframe[subframe];
1005 
1006             decode_pitch_vector(p, amr_subframe, subframe);
1007 
1008             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             decode_gains(p, amr_subframe, p->cur_frame_mode, subframe,
1016                          &fixed_gain_factor);
1017 
1018             pitch_sharpening(p, subframe, p->cur_frame_mode, &fixed_sparse);
1019 
1020             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             ff_set_fixed_vector(p->fixed_vector, &fixed_sparse, 1.0,
1025                                 AMR_SUBFRAME_SIZE);
1026 
1027             p->fixed_gain[4] =
1028                 ff_amr_set_fixed_gain(fixed_gain_factor,
1029                                       p->celpm_ctx.dot_productf(p->fixed_vector,
1030                                                                 p->fixed_vector,
1031                                                                 AMR_SUBFRAME_SIZE) /
1032                                       AMR_SUBFRAME_SIZE,
1033                                       p->prediction_error,
1034                                       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             for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
1039                 p->excitation[i] *= p->pitch_gain[4];
1040             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             for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
1049                 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             synth_fixed_gain = fixed_gain_smooth(p, p->lsf_q[subframe],
1055                                                  p->lsf_avg, p->cur_frame_mode);
1056 
1057             synth_fixed_vector = anti_sparseness(p, &fixed_sparse, p->fixed_vector,
1058                                                  synth_fixed_gain, spare_vector);
1059 
1060             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             postfilter(p, p->lpc[subframe], buf_out + subframe * AMR_SUBFRAME_SIZE);
1069 
1070             // update buffers and history
1071             ff_clear_fixed_vector(p->fixed_vector, &fixed_sparse, AMR_SUBFRAME_SIZE);
1072             update_state(p);
1073         }
1074 
1075         p->acelpf_ctx.acelp_apply_order_2_transfer_function(buf_out,
1076                                                             buf_out, highpass_zeros,
1077                                                             highpass_poles,
1078                                                             highpass_gain * AMR_SAMPLE_SCALE,
1079                                                             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         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         buf += channel_size;
1090         buf_size -= channel_size;
1091     }
1092 
1093     *got_frame_ptr = 1;
1094 
1095     return avpkt->size;
1096 }
1097 
1098 
1099 const FFCodec ff_amrnb_decoder = {
1100     .p.name         = "amrnb",
1101     .p.long_name    = NULL_IF_CONFIG_SMALL("AMR-NB (Adaptive Multi-Rate NarrowBand)"),
1102     .p.type         = AVMEDIA_TYPE_AUDIO,
1103     .p.id           = AV_CODEC_ID_AMR_NB,
1104     .priv_data_size = sizeof(AMRChannelsContext),
1105     .init           = amrnb_decode_init,
1106     FF_CODEC_DECODE_CB(amrnb_decode_frame),
1107     .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF,
1108     .p.sample_fmts  = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLTP,
1109                                                      AV_SAMPLE_FMT_NONE },
1110     .caps_internal  = FF_CODEC_CAP_INIT_THREADSAFE,
1111 };
1112