FFmpeg coverage


Directory: ../../../ffmpeg/
File: src/libavcodec/g729dec.c
Date: 2023-10-02 11:06:47
Exec Total Coverage
Lines: 0 254 0.0%
Functions: 0 10 0.0%
Branches: 0 129 0.0%

Line Branch Exec Source
1 /*
2 * G.729, G729 Annex D decoders
3 * Copyright (c) 2008 Vladimir Voroshilov
4 *
5 * This file is part of FFmpeg.
6 *
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
11 *
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
16 *
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 */
21
22 #include <inttypes.h>
23 #include <string.h>
24
25 #include "avcodec.h"
26 #include "libavutil/avutil.h"
27 #include "get_bits.h"
28 #include "audiodsp.h"
29 #include "codec_internal.h"
30 #include "decode.h"
31
32
33 #include "g729.h"
34 #include "lsp.h"
35 #include "celp_filters.h"
36 #include "acelp_filters.h"
37 #include "acelp_pitch_delay.h"
38 #include "acelp_vectors.h"
39 #include "g729data.h"
40 #include "g729postfilter.h"
41
42 /**
43 * minimum quantized LSF value (3.2.4)
44 * 0.005 in Q13
45 */
46 #define LSFQ_MIN 40
47
48 /**
49 * maximum quantized LSF value (3.2.4)
50 * 3.135 in Q13
51 */
52 #define LSFQ_MAX 25681
53
54 /**
55 * minimum LSF distance (3.2.4)
56 * 0.0391 in Q13
57 */
58 #define LSFQ_DIFF_MIN 321
59
60 /// interpolation filter length
61 #define INTERPOL_LEN 11
62
63 /**
64 * minimum gain pitch value (3.8, Equation 47)
65 * 0.2 in (1.14)
66 */
67 #define SHARP_MIN 3277
68
69 /**
70 * maximum gain pitch value (3.8, Equation 47)
71 * (EE) This does not comply with the specification.
72 * Specification says about 0.8, which should be
73 * 13107 in (1.14), but reference C code uses
74 * 13017 (equals to 0.7945) instead of it.
75 */
76 #define SHARP_MAX 13017
77
78 /**
79 * MR_ENERGY (mean removed energy) = mean_energy + 10 * log10(2^26 * subframe_size) in (7.13)
80 */
81 #define MR_ENERGY 1018156
82
83 #define DECISION_NOISE 0
84 #define DECISION_INTERMEDIATE 1
85 #define DECISION_VOICE 2
86
87 typedef enum {
88 FORMAT_G729_8K = 0,
89 FORMAT_G729D_6K4,
90 FORMAT_COUNT,
91 } G729Formats;
92
93 typedef struct {
94 uint8_t ac_index_bits[2]; ///< adaptive codebook index for second subframe (size in bits)
95 uint8_t parity_bit; ///< parity bit for pitch delay
96 uint8_t gc_1st_index_bits; ///< gain codebook (first stage) index (size in bits)
97 uint8_t gc_2nd_index_bits; ///< gain codebook (second stage) index (size in bits)
98 uint8_t fc_signs_bits; ///< number of pulses in fixed-codebook vector
99 uint8_t fc_indexes_bits; ///< size (in bits) of fixed-codebook index entry
100 uint8_t block_size;
101 } G729FormatDescription;
102
103 typedef struct {
104 /// past excitation signal buffer
105 int16_t exc_base[2*SUBFRAME_SIZE+PITCH_DELAY_MAX+INTERPOL_LEN];
106
107 int16_t* exc; ///< start of past excitation data in buffer
108 int pitch_delay_int_prev; ///< integer part of previous subframe's pitch delay (4.1.3)
109
110 /// (2.13) LSP quantizer outputs
111 int16_t past_quantizer_output_buf[MA_NP + 1][10];
112 int16_t* past_quantizer_outputs[MA_NP + 1];
113
114 int16_t lsfq[10]; ///< (2.13) quantized LSF coefficients from previous frame
115 int16_t lsp_buf[2][10]; ///< (0.15) LSP coefficients (previous and current frames) (3.2.5)
116 int16_t *lsp[2]; ///< pointers to lsp_buf
117
118 int16_t quant_energy[4]; ///< (5.10) past quantized energy
119
120 /// previous speech data for LP synthesis filter
121 int16_t syn_filter_data[10];
122
123
124 /// residual signal buffer (used in long-term postfilter)
125 int16_t residual[SUBFRAME_SIZE + RES_PREV_DATA_SIZE];
126
127 /// previous speech data for residual calculation filter
128 int16_t res_filter_data[SUBFRAME_SIZE+10];
129
130 /// previous speech data for short-term postfilter
131 int16_t pos_filter_data[SUBFRAME_SIZE+10];
132
133 /// (1.14) pitch gain of current and five previous subframes
134 int16_t past_gain_pitch[6];
135
136 /// (14.1) gain code from current and previous subframe
137 int16_t past_gain_code[2];
138
139 /// voice decision on previous subframe (0-noise, 1-intermediate, 2-voice), G.729D
140 int16_t voice_decision;
141
142 int16_t onset; ///< detected onset level (0-2)
143 int16_t was_periodic; ///< whether previous frame was declared as periodic or not (4.4)
144 int16_t ht_prev_data; ///< previous data for 4.2.3, equation 86
145 int gain_coeff; ///< (1.14) gain coefficient (4.2.4)
146 uint16_t rand_value; ///< random number generator value (4.4.4)
147 int ma_predictor_prev; ///< switched MA predictor of LSP quantizer from last good frame
148
149 /// (14.14) high-pass filter data (past input)
150 int hpf_f[2];
151
152 /// high-pass filter data (past output)
153 int16_t hpf_z[2];
154 } G729ChannelContext;
155
156 typedef struct {
157 AudioDSPContext adsp;
158
159 G729ChannelContext *channel_context;
160 } G729Context;
161
162 static const G729FormatDescription format_g729_8k = {
163 .ac_index_bits = {8,5},
164 .parity_bit = 1,
165 .gc_1st_index_bits = GC_1ST_IDX_BITS_8K,
166 .gc_2nd_index_bits = GC_2ND_IDX_BITS_8K,
167 .fc_signs_bits = 4,
168 .fc_indexes_bits = 13,
169 .block_size = G729_8K_BLOCK_SIZE,
170 };
171
172 static const G729FormatDescription format_g729d_6k4 = {
173 .ac_index_bits = {8,4},
174 .parity_bit = 0,
175 .gc_1st_index_bits = GC_1ST_IDX_BITS_6K4,
176 .gc_2nd_index_bits = GC_2ND_IDX_BITS_6K4,
177 .fc_signs_bits = 2,
178 .fc_indexes_bits = 9,
179 .block_size = G729D_6K4_BLOCK_SIZE,
180 };
181
182 /**
183 * @brief pseudo random number generator
184 */
185 static inline uint16_t g729_prng(uint16_t value)
186 {
187 return 31821 * value + 13849;
188 }
189
190 /**
191 * Decodes LSF (Line Spectral Frequencies) from L0-L3 (3.2.4).
192 * @param[out] lsfq (2.13) quantized LSF coefficients
193 * @param[in,out] past_quantizer_outputs (2.13) quantizer outputs from previous frames
194 * @param ma_predictor switched MA predictor of LSP quantizer
195 * @param vq_1st first stage vector of quantizer
196 * @param vq_2nd_low second stage lower vector of LSP quantizer
197 * @param vq_2nd_high second stage higher vector of LSP quantizer
198 */
199 static void lsf_decode(int16_t* lsfq, int16_t* past_quantizer_outputs[MA_NP + 1],
200 int16_t ma_predictor,
201 int16_t vq_1st, int16_t vq_2nd_low, int16_t vq_2nd_high)
202 {
203 int i,j;
204 static const uint8_t min_distance[2]={10, 5}; //(2.13)
205 int16_t* quantizer_output = past_quantizer_outputs[MA_NP];
206
207 for (i = 0; i < 5; i++) {
208 quantizer_output[i] = cb_lsp_1st[vq_1st][i ] + cb_lsp_2nd[vq_2nd_low ][i ];
209 quantizer_output[i + 5] = cb_lsp_1st[vq_1st][i + 5] + cb_lsp_2nd[vq_2nd_high][i + 5];
210 }
211
212 for (j = 0; j < 2; j++) {
213 for (i = 1; i < 10; i++) {
214 int diff = (quantizer_output[i - 1] - quantizer_output[i] + min_distance[j]) >> 1;
215 if (diff > 0) {
216 quantizer_output[i - 1] -= diff;
217 quantizer_output[i ] += diff;
218 }
219 }
220 }
221
222 for (i = 0; i < 10; i++) {
223 int sum = quantizer_output[i] * cb_ma_predictor_sum[ma_predictor][i];
224 for (j = 0; j < MA_NP; j++)
225 sum += past_quantizer_outputs[j][i] * cb_ma_predictor[ma_predictor][j][i];
226
227 lsfq[i] = sum >> 15;
228 }
229
230 ff_acelp_reorder_lsf(lsfq, LSFQ_DIFF_MIN, LSFQ_MIN, LSFQ_MAX, 10);
231 }
232
233 /**
234 * Restores past LSP quantizer output using LSF from previous frame
235 * @param[in,out] lsfq (2.13) quantized LSF coefficients
236 * @param[in,out] past_quantizer_outputs (2.13) quantizer outputs from previous frames
237 * @param ma_predictor_prev MA predictor from previous frame
238 * @param lsfq_prev (2.13) quantized LSF coefficients from previous frame
239 */
240 static void lsf_restore_from_previous(int16_t* lsfq,
241 int16_t* past_quantizer_outputs[MA_NP + 1],
242 int ma_predictor_prev)
243 {
244 int16_t* quantizer_output = past_quantizer_outputs[MA_NP];
245 int i,k;
246
247 for (i = 0; i < 10; i++) {
248 int tmp = lsfq[i] << 15;
249
250 for (k = 0; k < MA_NP; k++)
251 tmp -= past_quantizer_outputs[k][i] * cb_ma_predictor[ma_predictor_prev][k][i];
252
253 quantizer_output[i] = ((tmp >> 15) * cb_ma_predictor_sum_inv[ma_predictor_prev][i]) >> 12;
254 }
255 }
256
257 /**
258 * Constructs new excitation signal and applies phase filter to it
259 * @param[out] out constructed speech signal
260 * @param in original excitation signal
261 * @param fc_cur (2.13) original fixed-codebook vector
262 * @param gain_code (14.1) gain code
263 * @param subframe_size length of the subframe
264 */
265 static void g729d_get_new_exc(
266 int16_t* out,
267 const int16_t* in,
268 const int16_t* fc_cur,
269 int dstate,
270 int gain_code,
271 int subframe_size)
272 {
273 int i;
274 int16_t fc_new[SUBFRAME_SIZE];
275
276 ff_celp_convolve_circ(fc_new, fc_cur, phase_filter[dstate], subframe_size);
277
278 for (i = 0; i < subframe_size; i++) {
279 out[i] = in[i];
280 out[i] -= (gain_code * fc_cur[i] + 0x2000) >> 14;
281 out[i] += (gain_code * fc_new[i] + 0x2000) >> 14;
282 }
283 }
284
285 /**
286 * Makes decision about onset in current subframe
287 * @param past_onset decision result of previous subframe
288 * @param past_gain_code gain code of current and previous subframe
289 *
290 * @return onset decision result for current subframe
291 */
292 static int g729d_onset_decision(int past_onset, const int16_t* past_gain_code)
293 {
294 if ((past_gain_code[0] >> 1) > past_gain_code[1])
295 return 2;
296
297 return FFMAX(past_onset-1, 0);
298 }
299
300 /**
301 * Makes decision about voice presence in current subframe
302 * @param onset onset level
303 * @param prev_voice_decision voice decision result from previous subframe
304 * @param past_gain_pitch pitch gain of current and previous subframes
305 *
306 * @return voice decision result for current subframe
307 */
308 static int16_t g729d_voice_decision(int onset, int prev_voice_decision, const int16_t* past_gain_pitch)
309 {
310 int i, low_gain_pitch_cnt, voice_decision;
311
312 if (past_gain_pitch[0] >= 14745) { // 0.9
313 voice_decision = DECISION_VOICE;
314 } else if (past_gain_pitch[0] <= 9830) { // 0.6
315 voice_decision = DECISION_NOISE;
316 } else {
317 voice_decision = DECISION_INTERMEDIATE;
318 }
319
320 for (i = 0, low_gain_pitch_cnt = 0; i < 6; i++)
321 if (past_gain_pitch[i] < 9830)
322 low_gain_pitch_cnt++;
323
324 if (low_gain_pitch_cnt > 2 && !onset)
325 voice_decision = DECISION_NOISE;
326
327 if (!onset && voice_decision > prev_voice_decision + 1)
328 voice_decision--;
329
330 if (onset && voice_decision < DECISION_VOICE)
331 voice_decision++;
332
333 return voice_decision;
334 }
335
336 static int32_t scalarproduct_int16_c(const int16_t * v1, const int16_t * v2, int order)
337 {
338 int64_t res = 0;
339
340 while (order--)
341 res += *v1++ * *v2++;
342
343 if (res > INT32_MAX) return INT32_MAX;
344 else if (res < INT32_MIN) return INT32_MIN;
345
346 return res;
347 }
348
349 static av_cold int decoder_init(AVCodecContext * avctx)
350 {
351 G729Context *s = avctx->priv_data;
352 G729ChannelContext *ctx;
353 int channels = avctx->ch_layout.nb_channels;
354 int c,i,k;
355
356 if (channels < 1 || channels > 2) {
357 av_log(avctx, AV_LOG_ERROR, "Only mono and stereo are supported (requested channels: %d).\n", channels);
358 return AVERROR(EINVAL);
359 }
360 avctx->sample_fmt = AV_SAMPLE_FMT_S16P;
361
362 /* Both 8kbit/s and 6.4kbit/s modes uses two subframes per frame. */
363 avctx->frame_size = SUBFRAME_SIZE << 1;
364
365 ctx =
366 s->channel_context = av_mallocz(sizeof(G729ChannelContext) * channels);
367 if (!ctx)
368 return AVERROR(ENOMEM);
369
370 for (c = 0; c < channels; c++) {
371 ctx->gain_coeff = 16384; // 1.0 in (1.14)
372
373 for (k = 0; k < MA_NP + 1; k++) {
374 ctx->past_quantizer_outputs[k] = ctx->past_quantizer_output_buf[k];
375 for (i = 1; i < 11; i++)
376 ctx->past_quantizer_outputs[k][i - 1] = (18717 * i) >> 3;
377 }
378
379 ctx->lsp[0] = ctx->lsp_buf[0];
380 ctx->lsp[1] = ctx->lsp_buf[1];
381 memcpy(ctx->lsp[0], lsp_init, 10 * sizeof(int16_t));
382
383 ctx->exc = &ctx->exc_base[PITCH_DELAY_MAX+INTERPOL_LEN];
384
385 ctx->pitch_delay_int_prev = PITCH_DELAY_MIN;
386
387 /* random seed initialization */
388 ctx->rand_value = 21845;
389
390 /* quantized prediction error */
391 for (i = 0; i < 4; i++)
392 ctx->quant_energy[i] = -14336; // -14 in (5.10)
393
394 ctx++;
395 }
396
397 ff_audiodsp_init(&s->adsp);
398 s->adsp.scalarproduct_int16 = scalarproduct_int16_c;
399
400 return 0;
401 }
402
403 static int decode_frame(AVCodecContext *avctx, AVFrame *frame,
404 int *got_frame_ptr, AVPacket *avpkt)
405 {
406 const uint8_t *buf = avpkt->data;
407 int buf_size = avpkt->size;
408 int16_t *out_frame;
409 GetBitContext gb;
410 const G729FormatDescription *format;
411 int c, i;
412 int16_t *tmp;
413 G729Formats packet_type;
414 G729Context *s = avctx->priv_data;
415 G729ChannelContext *ctx = s->channel_context;
416 int channels = avctx->ch_layout.nb_channels;
417 int16_t lp[2][11]; // (3.12)
418 uint8_t ma_predictor; ///< switched MA predictor of LSP quantizer
419 uint8_t quantizer_1st; ///< first stage vector of quantizer
420 uint8_t quantizer_2nd_lo; ///< second stage lower vector of quantizer (size in bits)
421 uint8_t quantizer_2nd_hi; ///< second stage higher vector of quantizer (size in bits)
422
423 int pitch_delay_int[2]; // pitch delay, integer part
424 int pitch_delay_3x; // pitch delay, multiplied by 3
425 int16_t fc[SUBFRAME_SIZE]; // fixed-codebook vector
426 int16_t synth[SUBFRAME_SIZE+10]; // fixed-codebook vector
427 int j, ret;
428 int gain_before, gain_after;
429
430 frame->nb_samples = SUBFRAME_SIZE<<1;
431 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
432 return ret;
433
434 if (buf_size && buf_size % ((G729_8K_BLOCK_SIZE + (avctx->codec_id == AV_CODEC_ID_ACELP_KELVIN)) * channels) == 0) {
435 packet_type = FORMAT_G729_8K;
436 format = &format_g729_8k;
437 //Reset voice decision
438 ctx->onset = 0;
439 ctx->voice_decision = DECISION_VOICE;
440 av_log(avctx, AV_LOG_DEBUG, "Packet type: %s\n", "G.729 @ 8kbit/s");
441 } else if (buf_size == G729D_6K4_BLOCK_SIZE * channels && avctx->codec_id != AV_CODEC_ID_ACELP_KELVIN) {
442 packet_type = FORMAT_G729D_6K4;
443 format = &format_g729d_6k4;
444 av_log(avctx, AV_LOG_DEBUG, "Packet type: %s\n", "G.729D @ 6.4kbit/s");
445 } else {
446 av_log(avctx, AV_LOG_ERROR, "Packet size %d is unknown.\n", buf_size);
447 return AVERROR_INVALIDDATA;
448 }
449
450 for (c = 0; c < channels; c++) {
451 int frame_erasure = 0; ///< frame erasure detected during decoding
452 int bad_pitch = 0; ///< parity check failed
453 int is_periodic = 0; ///< whether one of the subframes is declared as periodic or not
454 out_frame = (int16_t*)frame->data[c];
455 if (avctx->codec_id == AV_CODEC_ID_ACELP_KELVIN) {
456 if (*buf != ((avctx->ch_layout.nb_channels - 1 - c) * 0x80 | 2))
457 avpriv_request_sample(avctx, "First byte value %x for channel %d", *buf, c);
458 buf++;
459 }
460
461 for (i = 0; i < format->block_size; i++)
462 frame_erasure |= buf[i];
463 frame_erasure = !frame_erasure;
464
465 init_get_bits8(&gb, buf, format->block_size);
466
467 ma_predictor = get_bits(&gb, 1);
468 quantizer_1st = get_bits(&gb, VQ_1ST_BITS);
469 quantizer_2nd_lo = get_bits(&gb, VQ_2ND_BITS);
470 quantizer_2nd_hi = get_bits(&gb, VQ_2ND_BITS);
471
472 if (frame_erasure) {
473 lsf_restore_from_previous(ctx->lsfq, ctx->past_quantizer_outputs,
474 ctx->ma_predictor_prev);
475 } else {
476 lsf_decode(ctx->lsfq, ctx->past_quantizer_outputs,
477 ma_predictor,
478 quantizer_1st, quantizer_2nd_lo, quantizer_2nd_hi);
479 ctx->ma_predictor_prev = ma_predictor;
480 }
481
482 tmp = ctx->past_quantizer_outputs[MA_NP];
483 memmove(ctx->past_quantizer_outputs + 1, ctx->past_quantizer_outputs,
484 MA_NP * sizeof(int16_t*));
485 ctx->past_quantizer_outputs[0] = tmp;
486
487 ff_acelp_lsf2lsp(ctx->lsp[1], ctx->lsfq, 10);
488
489 ff_acelp_lp_decode(&lp[0][0], &lp[1][0], ctx->lsp[1], ctx->lsp[0], 10);
490
491 FFSWAP(int16_t*, ctx->lsp[1], ctx->lsp[0]);
492
493 for (i = 0; i < 2; i++) {
494 int gain_corr_factor;
495
496 uint8_t ac_index; ///< adaptive codebook index
497 uint8_t pulses_signs; ///< fixed-codebook vector pulse signs
498 int fc_indexes; ///< fixed-codebook indexes
499 uint8_t gc_1st_index; ///< gain codebook (first stage) index
500 uint8_t gc_2nd_index; ///< gain codebook (second stage) index
501
502 ac_index = get_bits(&gb, format->ac_index_bits[i]);
503 if (!i && format->parity_bit)
504 bad_pitch = av_parity(ac_index >> 2) == get_bits1(&gb);
505 fc_indexes = get_bits(&gb, format->fc_indexes_bits);
506 pulses_signs = get_bits(&gb, format->fc_signs_bits);
507 gc_1st_index = get_bits(&gb, format->gc_1st_index_bits);
508 gc_2nd_index = get_bits(&gb, format->gc_2nd_index_bits);
509
510 if (frame_erasure) {
511 pitch_delay_3x = 3 * ctx->pitch_delay_int_prev;
512 } else if (!i) {
513 if (bad_pitch) {
514 pitch_delay_3x = 3 * ctx->pitch_delay_int_prev;
515 } else {
516 pitch_delay_3x = ff_acelp_decode_8bit_to_1st_delay3(ac_index);
517 }
518 } else {
519 int pitch_delay_min = av_clip(ctx->pitch_delay_int_prev - 5,
520 PITCH_DELAY_MIN, PITCH_DELAY_MAX - 9);
521
522 if (packet_type == FORMAT_G729D_6K4) {
523 pitch_delay_3x = ff_acelp_decode_4bit_to_2nd_delay3(ac_index, pitch_delay_min);
524 } else {
525 pitch_delay_3x = ff_acelp_decode_5_6_bit_to_2nd_delay3(ac_index, pitch_delay_min);
526 }
527 }
528
529 /* Round pitch delay to nearest (used everywhere except ff_acelp_interpolate). */
530 pitch_delay_int[i] = (pitch_delay_3x + 1) / 3;
531 if (pitch_delay_int[i] > PITCH_DELAY_MAX) {
532 av_log(avctx, AV_LOG_WARNING, "pitch_delay_int %d is too large\n", pitch_delay_int[i]);
533 pitch_delay_int[i] = PITCH_DELAY_MAX;
534 }
535
536 if (frame_erasure) {
537 ctx->rand_value = g729_prng(ctx->rand_value);
538 fc_indexes = av_mod_uintp2(ctx->rand_value, format->fc_indexes_bits);
539
540 ctx->rand_value = g729_prng(ctx->rand_value);
541 pulses_signs = ctx->rand_value;
542 }
543
544
545 memset(fc, 0, sizeof(int16_t) * SUBFRAME_SIZE);
546 switch (packet_type) {
547 case FORMAT_G729_8K:
548 ff_acelp_fc_pulse_per_track(fc, ff_fc_4pulses_8bits_tracks_13,
549 ff_fc_4pulses_8bits_track_4,
550 fc_indexes, pulses_signs, 3, 3);
551 break;
552 case FORMAT_G729D_6K4:
553 ff_acelp_fc_pulse_per_track(fc, ff_fc_2pulses_9bits_track1_gray,
554 ff_fc_2pulses_9bits_track2_gray,
555 fc_indexes, pulses_signs, 1, 4);
556 break;
557 }
558
559 /*
560 This filter enhances harmonic components of the fixed-codebook vector to
561 improve the quality of the reconstructed speech.
562
563 / fc_v[i], i < pitch_delay
564 fc_v[i] = <
565 \ fc_v[i] + gain_pitch * fc_v[i-pitch_delay], i >= pitch_delay
566 */
567 if (SUBFRAME_SIZE > pitch_delay_int[i])
568 ff_acelp_weighted_vector_sum(fc + pitch_delay_int[i],
569 fc + pitch_delay_int[i],
570 fc, 1 << 14,
571 av_clip(ctx->past_gain_pitch[0], SHARP_MIN, SHARP_MAX),
572 0, 14,
573 SUBFRAME_SIZE - pitch_delay_int[i]);
574
575 memmove(ctx->past_gain_pitch+1, ctx->past_gain_pitch, 5 * sizeof(int16_t));
576 ctx->past_gain_code[1] = ctx->past_gain_code[0];
577
578 if (frame_erasure) {
579 ctx->past_gain_pitch[0] = (29491 * ctx->past_gain_pitch[0]) >> 15; // 0.90 (0.15)
580 ctx->past_gain_code[0] = ( 2007 * ctx->past_gain_code[0] ) >> 11; // 0.98 (0.11)
581
582 gain_corr_factor = 0;
583 } else {
584 if (packet_type == FORMAT_G729D_6K4) {
585 ctx->past_gain_pitch[0] = cb_gain_1st_6k4[gc_1st_index][0] +
586 cb_gain_2nd_6k4[gc_2nd_index][0];
587 gain_corr_factor = cb_gain_1st_6k4[gc_1st_index][1] +
588 cb_gain_2nd_6k4[gc_2nd_index][1];
589
590 /* Without check below overflow can occur in ff_acelp_update_past_gain.
591 It is not issue for G.729, because gain_corr_factor in it's case is always
592 greater than 1024, while in G.729D it can be even zero. */
593 gain_corr_factor = FFMAX(gain_corr_factor, 1024);
594 #ifndef G729_BITEXACT
595 gain_corr_factor >>= 1;
596 #endif
597 } else {
598 ctx->past_gain_pitch[0] = cb_gain_1st_8k[gc_1st_index][0] +
599 cb_gain_2nd_8k[gc_2nd_index][0];
600 gain_corr_factor = cb_gain_1st_8k[gc_1st_index][1] +
601 cb_gain_2nd_8k[gc_2nd_index][1];
602 }
603
604 /* Decode the fixed-codebook gain. */
605 ctx->past_gain_code[0] = ff_acelp_decode_gain_code(&s->adsp, gain_corr_factor,
606 fc, MR_ENERGY,
607 ctx->quant_energy,
608 ma_prediction_coeff,
609 SUBFRAME_SIZE, 4);
610 #ifdef G729_BITEXACT
611 /*
612 This correction required to get bit-exact result with
613 reference code, because gain_corr_factor in G.729D is
614 two times larger than in original G.729.
615
616 If bit-exact result is not issue then gain_corr_factor
617 can be simpler divided by 2 before call to g729_get_gain_code
618 instead of using correction below.
619 */
620 if (packet_type == FORMAT_G729D_6K4) {
621 gain_corr_factor >>= 1;
622 ctx->past_gain_code[0] >>= 1;
623 }
624 #endif
625 }
626 ff_acelp_update_past_gain(ctx->quant_energy, gain_corr_factor, 2, frame_erasure);
627
628 /* Routine requires rounding to lowest. */
629 ff_acelp_interpolate(ctx->exc + i * SUBFRAME_SIZE,
630 ctx->exc + i * SUBFRAME_SIZE - pitch_delay_3x / 3,
631 ff_acelp_interp_filter, 6,
632 (pitch_delay_3x % 3) << 1,
633 10, SUBFRAME_SIZE);
634
635 ff_acelp_weighted_vector_sum(ctx->exc + i * SUBFRAME_SIZE,
636 ctx->exc + i * SUBFRAME_SIZE, fc,
637 (!ctx->was_periodic && frame_erasure) ? 0 : ctx->past_gain_pitch[0],
638 ( ctx->was_periodic && frame_erasure) ? 0 : ctx->past_gain_code[0],
639 1 << 13, 14, SUBFRAME_SIZE);
640
641 memcpy(synth, ctx->syn_filter_data, 10 * sizeof(int16_t));
642
643 if (ff_celp_lp_synthesis_filter(
644 synth+10,
645 &lp[i][1],
646 ctx->exc + i * SUBFRAME_SIZE,
647 SUBFRAME_SIZE,
648 10,
649 1,
650 0,
651 0x800))
652 /* Overflow occurred, downscale excitation signal... */
653 for (j = 0; j < 2 * SUBFRAME_SIZE + PITCH_DELAY_MAX + INTERPOL_LEN; j++)
654 ctx->exc_base[j] >>= 2;
655
656 /* ... and make synthesis again. */
657 if (packet_type == FORMAT_G729D_6K4) {
658 int16_t exc_new[SUBFRAME_SIZE];
659
660 ctx->onset = g729d_onset_decision(ctx->onset, ctx->past_gain_code);
661 ctx->voice_decision = g729d_voice_decision(ctx->onset, ctx->voice_decision, ctx->past_gain_pitch);
662
663 g729d_get_new_exc(exc_new, ctx->exc + i * SUBFRAME_SIZE, fc, ctx->voice_decision, ctx->past_gain_code[0], SUBFRAME_SIZE);
664
665 ff_celp_lp_synthesis_filter(
666 synth+10,
667 &lp[i][1],
668 exc_new,
669 SUBFRAME_SIZE,
670 10,
671 0,
672 0,
673 0x800);
674 } else {
675 ff_celp_lp_synthesis_filter(
676 synth+10,
677 &lp[i][1],
678 ctx->exc + i * SUBFRAME_SIZE,
679 SUBFRAME_SIZE,
680 10,
681 0,
682 0,
683 0x800);
684 }
685 /* Save data (without postfilter) for use in next subframe. */
686 memcpy(ctx->syn_filter_data, synth+SUBFRAME_SIZE, 10 * sizeof(int16_t));
687
688 /* Calculate gain of unfiltered signal for use in AGC. */
689 gain_before = 0;
690 for (j = 0; j < SUBFRAME_SIZE; j++)
691 gain_before += FFABS(synth[j+10]);
692
693 /* Call postfilter and also update voicing decision for use in next frame. */
694 ff_g729_postfilter(
695 &s->adsp,
696 &ctx->ht_prev_data,
697 &is_periodic,
698 &lp[i][0],
699 pitch_delay_int[0],
700 ctx->residual,
701 ctx->res_filter_data,
702 ctx->pos_filter_data,
703 synth+10,
704 SUBFRAME_SIZE);
705
706 /* Calculate gain of filtered signal for use in AGC. */
707 gain_after = 0;
708 for (j = 0; j < SUBFRAME_SIZE; j++)
709 gain_after += FFABS(synth[j+10]);
710
711 ctx->gain_coeff = ff_g729_adaptive_gain_control(
712 gain_before,
713 gain_after,
714 synth+10,
715 SUBFRAME_SIZE,
716 ctx->gain_coeff);
717
718 if (frame_erasure) {
719 ctx->pitch_delay_int_prev = FFMIN(ctx->pitch_delay_int_prev + 1, PITCH_DELAY_MAX);
720 } else {
721 ctx->pitch_delay_int_prev = pitch_delay_int[i];
722 }
723
724 memcpy(synth+8, ctx->hpf_z, 2*sizeof(int16_t));
725 ff_acelp_high_pass_filter(
726 out_frame + i*SUBFRAME_SIZE,
727 ctx->hpf_f,
728 synth+10,
729 SUBFRAME_SIZE);
730 memcpy(ctx->hpf_z, synth+8+SUBFRAME_SIZE, 2*sizeof(int16_t));
731 }
732
733 ctx->was_periodic = is_periodic;
734
735 /* Save signal for use in next frame. */
736 memmove(ctx->exc_base, ctx->exc_base + 2 * SUBFRAME_SIZE, (PITCH_DELAY_MAX+INTERPOL_LEN)*sizeof(int16_t));
737
738 buf += format->block_size;
739 ctx++;
740 }
741
742 *got_frame_ptr = 1;
743 return (format->block_size + (avctx->codec_id == AV_CODEC_ID_ACELP_KELVIN)) * channels;
744 }
745
746 static av_cold int decode_close(AVCodecContext *avctx)
747 {
748 G729Context *s = avctx->priv_data;
749 av_freep(&s->channel_context);
750
751 return 0;
752 }
753
754 const FFCodec ff_g729_decoder = {
755 .p.name = "g729",
756 CODEC_LONG_NAME("G.729"),
757 .p.type = AVMEDIA_TYPE_AUDIO,
758 .p.id = AV_CODEC_ID_G729,
759 .priv_data_size = sizeof(G729Context),
760 .init = decoder_init,
761 FF_CODEC_DECODE_CB(decode_frame),
762 .close = decode_close,
763 .p.capabilities =
764 #if FF_API_SUBFRAMES
765 AV_CODEC_CAP_SUBFRAMES |
766 #endif
767 AV_CODEC_CAP_DR1,
768 };
769
770 const FFCodec ff_acelp_kelvin_decoder = {
771 .p.name = "acelp.kelvin",
772 CODEC_LONG_NAME("Sipro ACELP.KELVIN"),
773 .p.type = AVMEDIA_TYPE_AUDIO,
774 .p.id = AV_CODEC_ID_ACELP_KELVIN,
775 .priv_data_size = sizeof(G729Context),
776 .init = decoder_init,
777 FF_CODEC_DECODE_CB(decode_frame),
778 .close = decode_close,
779 .p.capabilities =
780 #if FF_API_SUBFRAMES
781 AV_CODEC_CAP_SUBFRAMES |
782 #endif
783 AV_CODEC_CAP_DR1,
784 };
785