blob: 2e1bf18e4e49f40dc501938ee05172beac53f94d
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 "internal.h" |
30 | |
31 | |
32 | #include "g729.h" |
33 | #include "lsp.h" |
34 | #include "celp_math.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 | } G729FormatDescription; |
101 | |
102 | typedef struct { |
103 | AudioDSPContext adsp; |
104 | |
105 | /// past excitation signal buffer |
106 | int16_t exc_base[2*SUBFRAME_SIZE+PITCH_DELAY_MAX+INTERPOL_LEN]; |
107 | |
108 | int16_t* exc; ///< start of past excitation data in buffer |
109 | int pitch_delay_int_prev; ///< integer part of previous subframe's pitch delay (4.1.3) |
110 | |
111 | /// (2.13) LSP quantizer outputs |
112 | int16_t past_quantizer_output_buf[MA_NP + 1][10]; |
113 | int16_t* past_quantizer_outputs[MA_NP + 1]; |
114 | |
115 | int16_t lsfq[10]; ///< (2.13) quantized LSF coefficients from previous frame |
116 | int16_t lsp_buf[2][10]; ///< (0.15) LSP coefficients (previous and current frames) (3.2.5) |
117 | int16_t *lsp[2]; ///< pointers to lsp_buf |
118 | |
119 | int16_t quant_energy[4]; ///< (5.10) past quantized energy |
120 | |
121 | /// previous speech data for LP synthesis filter |
122 | int16_t syn_filter_data[10]; |
123 | |
124 | |
125 | /// residual signal buffer (used in long-term postfilter) |
126 | int16_t residual[SUBFRAME_SIZE + RES_PREV_DATA_SIZE]; |
127 | |
128 | /// previous speech data for residual calculation filter |
129 | int16_t res_filter_data[SUBFRAME_SIZE+10]; |
130 | |
131 | /// previous speech data for short-term postfilter |
132 | int16_t pos_filter_data[SUBFRAME_SIZE+10]; |
133 | |
134 | /// (1.14) pitch gain of current and five previous subframes |
135 | int16_t past_gain_pitch[6]; |
136 | |
137 | /// (14.1) gain code from current and previous subframe |
138 | int16_t past_gain_code[2]; |
139 | |
140 | /// voice decision on previous subframe (0-noise, 1-intermediate, 2-voice), G.729D |
141 | int16_t voice_decision; |
142 | |
143 | int16_t onset; ///< detected onset level (0-2) |
144 | int16_t was_periodic; ///< whether previous frame was declared as periodic or not (4.4) |
145 | int16_t ht_prev_data; ///< previous data for 4.2.3, equation 86 |
146 | int gain_coeff; ///< (1.14) gain coefficient (4.2.4) |
147 | uint16_t rand_value; ///< random number generator value (4.4.4) |
148 | int ma_predictor_prev; ///< switched MA predictor of LSP quantizer from last good frame |
149 | |
150 | /// (14.14) high-pass filter data (past input) |
151 | int hpf_f[2]; |
152 | |
153 | /// high-pass filter data (past output) |
154 | int16_t hpf_z[2]; |
155 | } G729Context; |
156 | |
157 | static const G729FormatDescription format_g729_8k = { |
158 | .ac_index_bits = {8,5}, |
159 | .parity_bit = 1, |
160 | .gc_1st_index_bits = GC_1ST_IDX_BITS_8K, |
161 | .gc_2nd_index_bits = GC_2ND_IDX_BITS_8K, |
162 | .fc_signs_bits = 4, |
163 | .fc_indexes_bits = 13, |
164 | }; |
165 | |
166 | static const G729FormatDescription format_g729d_6k4 = { |
167 | .ac_index_bits = {8,4}, |
168 | .parity_bit = 0, |
169 | .gc_1st_index_bits = GC_1ST_IDX_BITS_6K4, |
170 | .gc_2nd_index_bits = GC_2ND_IDX_BITS_6K4, |
171 | .fc_signs_bits = 2, |
172 | .fc_indexes_bits = 9, |
173 | }; |
174 | |
175 | /** |
176 | * @brief pseudo random number generator |
177 | */ |
178 | static inline uint16_t g729_prng(uint16_t value) |
179 | { |
180 | return 31821 * value + 13849; |
181 | } |
182 | |
183 | /** |
184 | * Decodes LSF (Line Spectral Frequencies) from L0-L3 (3.2.4). |
185 | * @param[out] lsfq (2.13) quantized LSF coefficients |
186 | * @param[in,out] past_quantizer_outputs (2.13) quantizer outputs from previous frames |
187 | * @param ma_predictor switched MA predictor of LSP quantizer |
188 | * @param vq_1st first stage vector of quantizer |
189 | * @param vq_2nd_low second stage lower vector of LSP quantizer |
190 | * @param vq_2nd_high second stage higher vector of LSP quantizer |
191 | */ |
192 | static void lsf_decode(int16_t* lsfq, int16_t* past_quantizer_outputs[MA_NP + 1], |
193 | int16_t ma_predictor, |
194 | int16_t vq_1st, int16_t vq_2nd_low, int16_t vq_2nd_high) |
195 | { |
196 | int i,j; |
197 | static const uint8_t min_distance[2]={10, 5}; //(2.13) |
198 | int16_t* quantizer_output = past_quantizer_outputs[MA_NP]; |
199 | |
200 | for (i = 0; i < 5; i++) { |
201 | quantizer_output[i] = cb_lsp_1st[vq_1st][i ] + cb_lsp_2nd[vq_2nd_low ][i ]; |
202 | quantizer_output[i + 5] = cb_lsp_1st[vq_1st][i + 5] + cb_lsp_2nd[vq_2nd_high][i + 5]; |
203 | } |
204 | |
205 | for (j = 0; j < 2; j++) { |
206 | for (i = 1; i < 10; i++) { |
207 | int diff = (quantizer_output[i - 1] - quantizer_output[i] + min_distance[j]) >> 1; |
208 | if (diff > 0) { |
209 | quantizer_output[i - 1] -= diff; |
210 | quantizer_output[i ] += diff; |
211 | } |
212 | } |
213 | } |
214 | |
215 | for (i = 0; i < 10; i++) { |
216 | int sum = quantizer_output[i] * cb_ma_predictor_sum[ma_predictor][i]; |
217 | for (j = 0; j < MA_NP; j++) |
218 | sum += past_quantizer_outputs[j][i] * cb_ma_predictor[ma_predictor][j][i]; |
219 | |
220 | lsfq[i] = sum >> 15; |
221 | } |
222 | |
223 | ff_acelp_reorder_lsf(lsfq, LSFQ_DIFF_MIN, LSFQ_MIN, LSFQ_MAX, 10); |
224 | } |
225 | |
226 | /** |
227 | * Restores past LSP quantizer output using LSF from previous frame |
228 | * @param[in,out] lsfq (2.13) quantized LSF coefficients |
229 | * @param[in,out] past_quantizer_outputs (2.13) quantizer outputs from previous frames |
230 | * @param ma_predictor_prev MA predictor from previous frame |
231 | * @param lsfq_prev (2.13) quantized LSF coefficients from previous frame |
232 | */ |
233 | static void lsf_restore_from_previous(int16_t* lsfq, |
234 | int16_t* past_quantizer_outputs[MA_NP + 1], |
235 | int ma_predictor_prev) |
236 | { |
237 | int16_t* quantizer_output = past_quantizer_outputs[MA_NP]; |
238 | int i,k; |
239 | |
240 | for (i = 0; i < 10; i++) { |
241 | int tmp = lsfq[i] << 15; |
242 | |
243 | for (k = 0; k < MA_NP; k++) |
244 | tmp -= past_quantizer_outputs[k][i] * cb_ma_predictor[ma_predictor_prev][k][i]; |
245 | |
246 | quantizer_output[i] = ((tmp >> 15) * cb_ma_predictor_sum_inv[ma_predictor_prev][i]) >> 12; |
247 | } |
248 | } |
249 | |
250 | /** |
251 | * Constructs new excitation signal and applies phase filter to it |
252 | * @param[out] out constructed speech signal |
253 | * @param in original excitation signal |
254 | * @param fc_cur (2.13) original fixed-codebook vector |
255 | * @param gain_code (14.1) gain code |
256 | * @param subframe_size length of the subframe |
257 | */ |
258 | static void g729d_get_new_exc( |
259 | int16_t* out, |
260 | const int16_t* in, |
261 | const int16_t* fc_cur, |
262 | int dstate, |
263 | int gain_code, |
264 | int subframe_size) |
265 | { |
266 | int i; |
267 | int16_t fc_new[SUBFRAME_SIZE]; |
268 | |
269 | ff_celp_convolve_circ(fc_new, fc_cur, phase_filter[dstate], subframe_size); |
270 | |
271 | for(i=0; i<subframe_size; i++) |
272 | { |
273 | out[i] = in[i]; |
274 | out[i] -= (gain_code * fc_cur[i] + 0x2000) >> 14; |
275 | out[i] += (gain_code * fc_new[i] + 0x2000) >> 14; |
276 | } |
277 | } |
278 | |
279 | /** |
280 | * Makes decision about onset in current subframe |
281 | * @param past_onset decision result of previous subframe |
282 | * @param past_gain_code gain code of current and previous subframe |
283 | * |
284 | * @return onset decision result for current subframe |
285 | */ |
286 | static int g729d_onset_decision(int past_onset, const int16_t* past_gain_code) |
287 | { |
288 | if((past_gain_code[0] >> 1) > past_gain_code[1]) |
289 | return 2; |
290 | else |
291 | return FFMAX(past_onset-1, 0); |
292 | } |
293 | |
294 | /** |
295 | * Makes decision about voice presence in current subframe |
296 | * @param onset onset level |
297 | * @param prev_voice_decision voice decision result from previous subframe |
298 | * @param past_gain_pitch pitch gain of current and previous subframes |
299 | * |
300 | * @return voice decision result for current subframe |
301 | */ |
302 | static int16_t g729d_voice_decision(int onset, int prev_voice_decision, const int16_t* past_gain_pitch) |
303 | { |
304 | int i, low_gain_pitch_cnt, voice_decision; |
305 | |
306 | if(past_gain_pitch[0] >= 14745) // 0.9 |
307 | voice_decision = DECISION_VOICE; |
308 | else if (past_gain_pitch[0] <= 9830) // 0.6 |
309 | voice_decision = DECISION_NOISE; |
310 | else |
311 | voice_decision = DECISION_INTERMEDIATE; |
312 | |
313 | for(i=0, low_gain_pitch_cnt=0; i<6; i++) |
314 | if(past_gain_pitch[i] < 9830) |
315 | low_gain_pitch_cnt++; |
316 | |
317 | if(low_gain_pitch_cnt > 2 && !onset) |
318 | voice_decision = DECISION_NOISE; |
319 | |
320 | if(!onset && voice_decision > prev_voice_decision + 1) |
321 | voice_decision--; |
322 | |
323 | if(onset && voice_decision < DECISION_VOICE) |
324 | voice_decision++; |
325 | |
326 | return voice_decision; |
327 | } |
328 | |
329 | static int32_t scalarproduct_int16_c(const int16_t * v1, const int16_t * v2, int order) |
330 | { |
331 | int res = 0; |
332 | |
333 | while (order--) |
334 | res += *v1++ * *v2++; |
335 | |
336 | return res; |
337 | } |
338 | |
339 | static av_cold int decoder_init(AVCodecContext * avctx) |
340 | { |
341 | G729Context* ctx = avctx->priv_data; |
342 | int i,k; |
343 | |
344 | if (avctx->channels != 1) { |
345 | av_log(avctx, AV_LOG_ERROR, "Only mono sound is supported (requested channels: %d).\n", avctx->channels); |
346 | return AVERROR(EINVAL); |
347 | } |
348 | avctx->sample_fmt = AV_SAMPLE_FMT_S16; |
349 | |
350 | /* Both 8kbit/s and 6.4kbit/s modes uses two subframes per frame. */ |
351 | avctx->frame_size = SUBFRAME_SIZE << 1; |
352 | |
353 | ctx->gain_coeff = 16384; // 1.0 in (1.14) |
354 | |
355 | for (k = 0; k < MA_NP + 1; k++) { |
356 | ctx->past_quantizer_outputs[k] = ctx->past_quantizer_output_buf[k]; |
357 | for (i = 1; i < 11; i++) |
358 | ctx->past_quantizer_outputs[k][i - 1] = (18717 * i) >> 3; |
359 | } |
360 | |
361 | ctx->lsp[0] = ctx->lsp_buf[0]; |
362 | ctx->lsp[1] = ctx->lsp_buf[1]; |
363 | memcpy(ctx->lsp[0], lsp_init, 10 * sizeof(int16_t)); |
364 | |
365 | ctx->exc = &ctx->exc_base[PITCH_DELAY_MAX+INTERPOL_LEN]; |
366 | |
367 | ctx->pitch_delay_int_prev = PITCH_DELAY_MIN; |
368 | |
369 | /* random seed initialization */ |
370 | ctx->rand_value = 21845; |
371 | |
372 | /* quantized prediction error */ |
373 | for(i=0; i<4; i++) |
374 | ctx->quant_energy[i] = -14336; // -14 in (5.10) |
375 | |
376 | ff_audiodsp_init(&ctx->adsp); |
377 | ctx->adsp.scalarproduct_int16 = scalarproduct_int16_c; |
378 | |
379 | return 0; |
380 | } |
381 | |
382 | static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame_ptr, |
383 | AVPacket *avpkt) |
384 | { |
385 | const uint8_t *buf = avpkt->data; |
386 | int buf_size = avpkt->size; |
387 | int16_t *out_frame; |
388 | GetBitContext gb; |
389 | const G729FormatDescription *format; |
390 | int frame_erasure = 0; ///< frame erasure detected during decoding |
391 | int bad_pitch = 0; ///< parity check failed |
392 | int i; |
393 | int16_t *tmp; |
394 | G729Formats packet_type; |
395 | G729Context *ctx = avctx->priv_data; |
396 | int16_t lp[2][11]; // (3.12) |
397 | uint8_t ma_predictor; ///< switched MA predictor of LSP quantizer |
398 | uint8_t quantizer_1st; ///< first stage vector of quantizer |
399 | uint8_t quantizer_2nd_lo; ///< second stage lower vector of quantizer (size in bits) |
400 | uint8_t quantizer_2nd_hi; ///< second stage higher vector of quantizer (size in bits) |
401 | |
402 | int pitch_delay_int[2]; // pitch delay, integer part |
403 | int pitch_delay_3x; // pitch delay, multiplied by 3 |
404 | int16_t fc[SUBFRAME_SIZE]; // fixed-codebook vector |
405 | int16_t synth[SUBFRAME_SIZE+10]; // fixed-codebook vector |
406 | int j, ret; |
407 | int gain_before, gain_after; |
408 | int is_periodic = 0; // whether one of the subframes is declared as periodic or not |
409 | AVFrame *frame = data; |
410 | |
411 | frame->nb_samples = SUBFRAME_SIZE<<1; |
412 | if ((ret = ff_get_buffer(avctx, frame, 0)) < 0) |
413 | return ret; |
414 | out_frame = (int16_t*) frame->data[0]; |
415 | |
416 | if (buf_size % 10 == 0) { |
417 | packet_type = FORMAT_G729_8K; |
418 | format = &format_g729_8k; |
419 | //Reset voice decision |
420 | ctx->onset = 0; |
421 | ctx->voice_decision = DECISION_VOICE; |
422 | av_log(avctx, AV_LOG_DEBUG, "Packet type: %s\n", "G.729 @ 8kbit/s"); |
423 | } else if (buf_size == 8) { |
424 | packet_type = FORMAT_G729D_6K4; |
425 | format = &format_g729d_6k4; |
426 | av_log(avctx, AV_LOG_DEBUG, "Packet type: %s\n", "G.729D @ 6.4kbit/s"); |
427 | } else { |
428 | av_log(avctx, AV_LOG_ERROR, "Packet size %d is unknown.\n", buf_size); |
429 | return AVERROR_INVALIDDATA; |
430 | } |
431 | |
432 | for (i=0; i < buf_size; i++) |
433 | frame_erasure |= buf[i]; |
434 | frame_erasure = !frame_erasure; |
435 | |
436 | init_get_bits(&gb, buf, 8*buf_size); |
437 | |
438 | ma_predictor = get_bits(&gb, 1); |
439 | quantizer_1st = get_bits(&gb, VQ_1ST_BITS); |
440 | quantizer_2nd_lo = get_bits(&gb, VQ_2ND_BITS); |
441 | quantizer_2nd_hi = get_bits(&gb, VQ_2ND_BITS); |
442 | |
443 | if(frame_erasure) |
444 | lsf_restore_from_previous(ctx->lsfq, ctx->past_quantizer_outputs, |
445 | ctx->ma_predictor_prev); |
446 | else { |
447 | lsf_decode(ctx->lsfq, ctx->past_quantizer_outputs, |
448 | ma_predictor, |
449 | quantizer_1st, quantizer_2nd_lo, quantizer_2nd_hi); |
450 | ctx->ma_predictor_prev = ma_predictor; |
451 | } |
452 | |
453 | tmp = ctx->past_quantizer_outputs[MA_NP]; |
454 | memmove(ctx->past_quantizer_outputs + 1, ctx->past_quantizer_outputs, |
455 | MA_NP * sizeof(int16_t*)); |
456 | ctx->past_quantizer_outputs[0] = tmp; |
457 | |
458 | ff_acelp_lsf2lsp(ctx->lsp[1], ctx->lsfq, 10); |
459 | |
460 | ff_acelp_lp_decode(&lp[0][0], &lp[1][0], ctx->lsp[1], ctx->lsp[0], 10); |
461 | |
462 | FFSWAP(int16_t*, ctx->lsp[1], ctx->lsp[0]); |
463 | |
464 | for (i = 0; i < 2; i++) { |
465 | int gain_corr_factor; |
466 | |
467 | uint8_t ac_index; ///< adaptive codebook index |
468 | uint8_t pulses_signs; ///< fixed-codebook vector pulse signs |
469 | int fc_indexes; ///< fixed-codebook indexes |
470 | uint8_t gc_1st_index; ///< gain codebook (first stage) index |
471 | uint8_t gc_2nd_index; ///< gain codebook (second stage) index |
472 | |
473 | ac_index = get_bits(&gb, format->ac_index_bits[i]); |
474 | if(!i && format->parity_bit) |
475 | bad_pitch = av_parity(ac_index >> 2) == get_bits1(&gb); |
476 | fc_indexes = get_bits(&gb, format->fc_indexes_bits); |
477 | pulses_signs = get_bits(&gb, format->fc_signs_bits); |
478 | gc_1st_index = get_bits(&gb, format->gc_1st_index_bits); |
479 | gc_2nd_index = get_bits(&gb, format->gc_2nd_index_bits); |
480 | |
481 | if (frame_erasure) |
482 | pitch_delay_3x = 3 * ctx->pitch_delay_int_prev; |
483 | else if(!i) { |
484 | if (bad_pitch) |
485 | pitch_delay_3x = 3 * ctx->pitch_delay_int_prev; |
486 | else |
487 | pitch_delay_3x = ff_acelp_decode_8bit_to_1st_delay3(ac_index); |
488 | } else { |
489 | int pitch_delay_min = av_clip(ctx->pitch_delay_int_prev - 5, |
490 | PITCH_DELAY_MIN, PITCH_DELAY_MAX - 9); |
491 | |
492 | if(packet_type == FORMAT_G729D_6K4) |
493 | pitch_delay_3x = ff_acelp_decode_4bit_to_2nd_delay3(ac_index, pitch_delay_min); |
494 | else |
495 | pitch_delay_3x = ff_acelp_decode_5_6_bit_to_2nd_delay3(ac_index, pitch_delay_min); |
496 | } |
497 | |
498 | /* Round pitch delay to nearest (used everywhere except ff_acelp_interpolate). */ |
499 | pitch_delay_int[i] = (pitch_delay_3x + 1) / 3; |
500 | if (pitch_delay_int[i] > PITCH_DELAY_MAX) { |
501 | av_log(avctx, AV_LOG_WARNING, "pitch_delay_int %d is too large\n", pitch_delay_int[i]); |
502 | pitch_delay_int[i] = PITCH_DELAY_MAX; |
503 | } |
504 | |
505 | if (frame_erasure) { |
506 | ctx->rand_value = g729_prng(ctx->rand_value); |
507 | fc_indexes = av_mod_uintp2(ctx->rand_value, format->fc_indexes_bits); |
508 | |
509 | ctx->rand_value = g729_prng(ctx->rand_value); |
510 | pulses_signs = ctx->rand_value; |
511 | } |
512 | |
513 | |
514 | memset(fc, 0, sizeof(int16_t) * SUBFRAME_SIZE); |
515 | switch (packet_type) { |
516 | case FORMAT_G729_8K: |
517 | ff_acelp_fc_pulse_per_track(fc, ff_fc_4pulses_8bits_tracks_13, |
518 | ff_fc_4pulses_8bits_track_4, |
519 | fc_indexes, pulses_signs, 3, 3); |
520 | break; |
521 | case FORMAT_G729D_6K4: |
522 | ff_acelp_fc_pulse_per_track(fc, ff_fc_2pulses_9bits_track1_gray, |
523 | ff_fc_2pulses_9bits_track2_gray, |
524 | fc_indexes, pulses_signs, 1, 4); |
525 | break; |
526 | } |
527 | |
528 | /* |
529 | This filter enhances harmonic components of the fixed-codebook vector to |
530 | improve the quality of the reconstructed speech. |
531 | |
532 | / fc_v[i], i < pitch_delay |
533 | fc_v[i] = < |
534 | \ fc_v[i] + gain_pitch * fc_v[i-pitch_delay], i >= pitch_delay |
535 | */ |
536 | ff_acelp_weighted_vector_sum(fc + pitch_delay_int[i], |
537 | fc + pitch_delay_int[i], |
538 | fc, 1 << 14, |
539 | av_clip(ctx->past_gain_pitch[0], SHARP_MIN, SHARP_MAX), |
540 | 0, 14, |
541 | SUBFRAME_SIZE - pitch_delay_int[i]); |
542 | |
543 | memmove(ctx->past_gain_pitch+1, ctx->past_gain_pitch, 5 * sizeof(int16_t)); |
544 | ctx->past_gain_code[1] = ctx->past_gain_code[0]; |
545 | |
546 | if (frame_erasure) { |
547 | ctx->past_gain_pitch[0] = (29491 * ctx->past_gain_pitch[0]) >> 15; // 0.90 (0.15) |
548 | ctx->past_gain_code[0] = ( 2007 * ctx->past_gain_code[0] ) >> 11; // 0.98 (0.11) |
549 | |
550 | gain_corr_factor = 0; |
551 | } else { |
552 | if (packet_type == FORMAT_G729D_6K4) { |
553 | ctx->past_gain_pitch[0] = cb_gain_1st_6k4[gc_1st_index][0] + |
554 | cb_gain_2nd_6k4[gc_2nd_index][0]; |
555 | gain_corr_factor = cb_gain_1st_6k4[gc_1st_index][1] + |
556 | cb_gain_2nd_6k4[gc_2nd_index][1]; |
557 | |
558 | /* Without check below overflow can occur in ff_acelp_update_past_gain. |
559 | It is not issue for G.729, because gain_corr_factor in it's case is always |
560 | greater than 1024, while in G.729D it can be even zero. */ |
561 | gain_corr_factor = FFMAX(gain_corr_factor, 1024); |
562 | #ifndef G729_BITEXACT |
563 | gain_corr_factor >>= 1; |
564 | #endif |
565 | } else { |
566 | ctx->past_gain_pitch[0] = cb_gain_1st_8k[gc_1st_index][0] + |
567 | cb_gain_2nd_8k[gc_2nd_index][0]; |
568 | gain_corr_factor = cb_gain_1st_8k[gc_1st_index][1] + |
569 | cb_gain_2nd_8k[gc_2nd_index][1]; |
570 | } |
571 | |
572 | /* Decode the fixed-codebook gain. */ |
573 | ctx->past_gain_code[0] = ff_acelp_decode_gain_code(&ctx->adsp, gain_corr_factor, |
574 | fc, MR_ENERGY, |
575 | ctx->quant_energy, |
576 | ma_prediction_coeff, |
577 | SUBFRAME_SIZE, 4); |
578 | #ifdef G729_BITEXACT |
579 | /* |
580 | This correction required to get bit-exact result with |
581 | reference code, because gain_corr_factor in G.729D is |
582 | two times larger than in original G.729. |
583 | |
584 | If bit-exact result is not issue then gain_corr_factor |
585 | can be simpler divided by 2 before call to g729_get_gain_code |
586 | instead of using correction below. |
587 | */ |
588 | if (packet_type == FORMAT_G729D_6K4) { |
589 | gain_corr_factor >>= 1; |
590 | ctx->past_gain_code[0] >>= 1; |
591 | } |
592 | #endif |
593 | } |
594 | ff_acelp_update_past_gain(ctx->quant_energy, gain_corr_factor, 2, frame_erasure); |
595 | |
596 | /* Routine requires rounding to lowest. */ |
597 | ff_acelp_interpolate(ctx->exc + i * SUBFRAME_SIZE, |
598 | ctx->exc + i * SUBFRAME_SIZE - pitch_delay_3x / 3, |
599 | ff_acelp_interp_filter, 6, |
600 | (pitch_delay_3x % 3) << 1, |
601 | 10, SUBFRAME_SIZE); |
602 | |
603 | ff_acelp_weighted_vector_sum(ctx->exc + i * SUBFRAME_SIZE, |
604 | ctx->exc + i * SUBFRAME_SIZE, fc, |
605 | (!ctx->was_periodic && frame_erasure) ? 0 : ctx->past_gain_pitch[0], |
606 | ( ctx->was_periodic && frame_erasure) ? 0 : ctx->past_gain_code[0], |
607 | 1 << 13, 14, SUBFRAME_SIZE); |
608 | |
609 | memcpy(synth, ctx->syn_filter_data, 10 * sizeof(int16_t)); |
610 | |
611 | if (ff_celp_lp_synthesis_filter( |
612 | synth+10, |
613 | &lp[i][1], |
614 | ctx->exc + i * SUBFRAME_SIZE, |
615 | SUBFRAME_SIZE, |
616 | 10, |
617 | 1, |
618 | 0, |
619 | 0x800)) |
620 | /* Overflow occurred, downscale excitation signal... */ |
621 | for (j = 0; j < 2 * SUBFRAME_SIZE + PITCH_DELAY_MAX + INTERPOL_LEN; j++) |
622 | ctx->exc_base[j] >>= 2; |
623 | |
624 | /* ... and make synthesis again. */ |
625 | if (packet_type == FORMAT_G729D_6K4) { |
626 | int16_t exc_new[SUBFRAME_SIZE]; |
627 | |
628 | ctx->onset = g729d_onset_decision(ctx->onset, ctx->past_gain_code); |
629 | ctx->voice_decision = g729d_voice_decision(ctx->onset, ctx->voice_decision, ctx->past_gain_pitch); |
630 | |
631 | g729d_get_new_exc(exc_new, ctx->exc + i * SUBFRAME_SIZE, fc, ctx->voice_decision, ctx->past_gain_code[0], SUBFRAME_SIZE); |
632 | |
633 | ff_celp_lp_synthesis_filter( |
634 | synth+10, |
635 | &lp[i][1], |
636 | exc_new, |
637 | SUBFRAME_SIZE, |
638 | 10, |
639 | 0, |
640 | 0, |
641 | 0x800); |
642 | } else { |
643 | ff_celp_lp_synthesis_filter( |
644 | synth+10, |
645 | &lp[i][1], |
646 | ctx->exc + i * SUBFRAME_SIZE, |
647 | SUBFRAME_SIZE, |
648 | 10, |
649 | 0, |
650 | 0, |
651 | 0x800); |
652 | } |
653 | /* Save data (without postfilter) for use in next subframe. */ |
654 | memcpy(ctx->syn_filter_data, synth+SUBFRAME_SIZE, 10 * sizeof(int16_t)); |
655 | |
656 | /* Calculate gain of unfiltered signal for use in AGC. */ |
657 | gain_before = 0; |
658 | for (j = 0; j < SUBFRAME_SIZE; j++) |
659 | gain_before += FFABS(synth[j+10]); |
660 | |
661 | /* Call postfilter and also update voicing decision for use in next frame. */ |
662 | ff_g729_postfilter( |
663 | &ctx->adsp, |
664 | &ctx->ht_prev_data, |
665 | &is_periodic, |
666 | &lp[i][0], |
667 | pitch_delay_int[0], |
668 | ctx->residual, |
669 | ctx->res_filter_data, |
670 | ctx->pos_filter_data, |
671 | synth+10, |
672 | SUBFRAME_SIZE); |
673 | |
674 | /* Calculate gain of filtered signal for use in AGC. */ |
675 | gain_after = 0; |
676 | for(j=0; j<SUBFRAME_SIZE; j++) |
677 | gain_after += FFABS(synth[j+10]); |
678 | |
679 | ctx->gain_coeff = ff_g729_adaptive_gain_control( |
680 | gain_before, |
681 | gain_after, |
682 | synth+10, |
683 | SUBFRAME_SIZE, |
684 | ctx->gain_coeff); |
685 | |
686 | if (frame_erasure) |
687 | ctx->pitch_delay_int_prev = FFMIN(ctx->pitch_delay_int_prev + 1, PITCH_DELAY_MAX); |
688 | else |
689 | ctx->pitch_delay_int_prev = pitch_delay_int[i]; |
690 | |
691 | memcpy(synth+8, ctx->hpf_z, 2*sizeof(int16_t)); |
692 | ff_acelp_high_pass_filter( |
693 | out_frame + i*SUBFRAME_SIZE, |
694 | ctx->hpf_f, |
695 | synth+10, |
696 | SUBFRAME_SIZE); |
697 | memcpy(ctx->hpf_z, synth+8+SUBFRAME_SIZE, 2*sizeof(int16_t)); |
698 | } |
699 | |
700 | ctx->was_periodic = is_periodic; |
701 | |
702 | /* Save signal for use in next frame. */ |
703 | memmove(ctx->exc_base, ctx->exc_base + 2 * SUBFRAME_SIZE, (PITCH_DELAY_MAX+INTERPOL_LEN)*sizeof(int16_t)); |
704 | |
705 | *got_frame_ptr = 1; |
706 | return packet_type == FORMAT_G729_8K ? 10 : 8; |
707 | } |
708 | |
709 | AVCodec ff_g729_decoder = { |
710 | .name = "g729", |
711 | .long_name = NULL_IF_CONFIG_SMALL("G.729"), |
712 | .type = AVMEDIA_TYPE_AUDIO, |
713 | .id = AV_CODEC_ID_G729, |
714 | .priv_data_size = sizeof(G729Context), |
715 | .init = decoder_init, |
716 | .decode = decode_frame, |
717 | .capabilities = AV_CODEC_CAP_SUBFRAMES | AV_CODEC_CAP_DR1, |
718 | }; |
719 |