| 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 | * Get parity bit of bit 2..7 |
| 185 | */ |
| 186 | static inline int get_parity(uint8_t value) |
| 187 | { |
| 188 | return (0x6996966996696996ULL >> (value >> 2)) & 1; |
| 189 | } |
| 190 | |
| 191 | /** |
| 192 | * Decodes LSF (Line Spectral Frequencies) from L0-L3 (3.2.4). |
| 193 | * @param[out] lsfq (2.13) quantized LSF coefficients |
| 194 | * @param[in,out] past_quantizer_outputs (2.13) quantizer outputs from previous frames |
| 195 | * @param ma_predictor switched MA predictor of LSP quantizer |
| 196 | * @param vq_1st first stage vector of quantizer |
| 197 | * @param vq_2nd_low second stage lower vector of LSP quantizer |
| 198 | * @param vq_2nd_high second stage higher vector of LSP quantizer |
| 199 | */ |
| 200 | static void lsf_decode(int16_t* lsfq, int16_t* past_quantizer_outputs[MA_NP + 1], |
| 201 | int16_t ma_predictor, |
| 202 | int16_t vq_1st, int16_t vq_2nd_low, int16_t vq_2nd_high) |
| 203 | { |
| 204 | int i,j; |
| 205 | static const uint8_t min_distance[2]={10, 5}; //(2.13) |
| 206 | int16_t* quantizer_output = past_quantizer_outputs[MA_NP]; |
| 207 | |
| 208 | for (i = 0; i < 5; i++) { |
| 209 | quantizer_output[i] = cb_lsp_1st[vq_1st][i ] + cb_lsp_2nd[vq_2nd_low ][i ]; |
| 210 | quantizer_output[i + 5] = cb_lsp_1st[vq_1st][i + 5] + cb_lsp_2nd[vq_2nd_high][i + 5]; |
| 211 | } |
| 212 | |
| 213 | for (j = 0; j < 2; j++) { |
| 214 | for (i = 1; i < 10; i++) { |
| 215 | int diff = (quantizer_output[i - 1] - quantizer_output[i] + min_distance[j]) >> 1; |
| 216 | if (diff > 0) { |
| 217 | quantizer_output[i - 1] -= diff; |
| 218 | quantizer_output[i ] += diff; |
| 219 | } |
| 220 | } |
| 221 | } |
| 222 | |
| 223 | for (i = 0; i < 10; i++) { |
| 224 | int sum = quantizer_output[i] * cb_ma_predictor_sum[ma_predictor][i]; |
| 225 | for (j = 0; j < MA_NP; j++) |
| 226 | sum += past_quantizer_outputs[j][i] * cb_ma_predictor[ma_predictor][j][i]; |
| 227 | |
| 228 | lsfq[i] = sum >> 15; |
| 229 | } |
| 230 | |
| 231 | ff_acelp_reorder_lsf(lsfq, LSFQ_DIFF_MIN, LSFQ_MIN, LSFQ_MAX, 10); |
| 232 | } |
| 233 | |
| 234 | /** |
| 235 | * Restores past LSP quantizer output using LSF from previous frame |
| 236 | * @param[in,out] lsfq (2.13) quantized LSF coefficients |
| 237 | * @param[in,out] past_quantizer_outputs (2.13) quantizer outputs from previous frames |
| 238 | * @param ma_predictor_prev MA predictor from previous frame |
| 239 | * @param lsfq_prev (2.13) quantized LSF coefficients from previous frame |
| 240 | */ |
| 241 | static void lsf_restore_from_previous(int16_t* lsfq, |
| 242 | int16_t* past_quantizer_outputs[MA_NP + 1], |
| 243 | int ma_predictor_prev) |
| 244 | { |
| 245 | int16_t* quantizer_output = past_quantizer_outputs[MA_NP]; |
| 246 | int i,k; |
| 247 | |
| 248 | for (i = 0; i < 10; i++) { |
| 249 | int tmp = lsfq[i] << 15; |
| 250 | |
| 251 | for (k = 0; k < MA_NP; k++) |
| 252 | tmp -= past_quantizer_outputs[k][i] * cb_ma_predictor[ma_predictor_prev][k][i]; |
| 253 | |
| 254 | quantizer_output[i] = ((tmp >> 15) * cb_ma_predictor_sum_inv[ma_predictor_prev][i]) >> 12; |
| 255 | } |
| 256 | } |
| 257 | |
| 258 | /** |
| 259 | * Constructs new excitation signal and applies phase filter to it |
| 260 | * @param[out] out constructed speech signal |
| 261 | * @param in original excitation signal |
| 262 | * @param fc_cur (2.13) original fixed-codebook vector |
| 263 | * @param gain_code (14.1) gain code |
| 264 | * @param subframe_size length of the subframe |
| 265 | */ |
| 266 | static void g729d_get_new_exc( |
| 267 | int16_t* out, |
| 268 | const int16_t* in, |
| 269 | const int16_t* fc_cur, |
| 270 | int dstate, |
| 271 | int gain_code, |
| 272 | int subframe_size) |
| 273 | { |
| 274 | int i; |
| 275 | int16_t fc_new[SUBFRAME_SIZE]; |
| 276 | |
| 277 | ff_celp_convolve_circ(fc_new, fc_cur, phase_filter[dstate], subframe_size); |
| 278 | |
| 279 | for(i=0; i<subframe_size; i++) |
| 280 | { |
| 281 | out[i] = in[i]; |
| 282 | out[i] -= (gain_code * fc_cur[i] + 0x2000) >> 14; |
| 283 | out[i] += (gain_code * fc_new[i] + 0x2000) >> 14; |
| 284 | } |
| 285 | } |
| 286 | |
| 287 | /** |
| 288 | * Makes decision about onset in current subframe |
| 289 | * @param past_onset decision result of previous subframe |
| 290 | * @param past_gain_code gain code of current and previous subframe |
| 291 | * |
| 292 | * @return onset decision result for current subframe |
| 293 | */ |
| 294 | static int g729d_onset_decision(int past_onset, const int16_t* past_gain_code) |
| 295 | { |
| 296 | if((past_gain_code[0] >> 1) > past_gain_code[1]) |
| 297 | return 2; |
| 298 | else |
| 299 | return FFMAX(past_onset-1, 0); |
| 300 | } |
| 301 | |
| 302 | /** |
| 303 | * Makes decision about voice presence in current subframe |
| 304 | * @param onset onset level |
| 305 | * @param prev_voice_decision voice decision result from previous subframe |
| 306 | * @param past_gain_pitch pitch gain of current and previous subframes |
| 307 | * |
| 308 | * @return voice decision result for current subframe |
| 309 | */ |
| 310 | static int16_t g729d_voice_decision(int onset, int prev_voice_decision, const int16_t* past_gain_pitch) |
| 311 | { |
| 312 | int i, low_gain_pitch_cnt, voice_decision; |
| 313 | |
| 314 | if(past_gain_pitch[0] >= 14745) // 0.9 |
| 315 | voice_decision = DECISION_VOICE; |
| 316 | else if (past_gain_pitch[0] <= 9830) // 0.6 |
| 317 | voice_decision = DECISION_NOISE; |
| 318 | else |
| 319 | voice_decision = DECISION_INTERMEDIATE; |
| 320 | |
| 321 | for(i=0, low_gain_pitch_cnt=0; i<6; i++) |
| 322 | if(past_gain_pitch[i] < 9830) |
| 323 | low_gain_pitch_cnt++; |
| 324 | |
| 325 | if(low_gain_pitch_cnt > 2 && !onset) |
| 326 | voice_decision = DECISION_NOISE; |
| 327 | |
| 328 | if(!onset && voice_decision > prev_voice_decision + 1) |
| 329 | voice_decision--; |
| 330 | |
| 331 | if(onset && voice_decision < DECISION_VOICE) |
| 332 | voice_decision++; |
| 333 | |
| 334 | return voice_decision; |
| 335 | } |
| 336 | |
| 337 | static int32_t scalarproduct_int16_c(const int16_t * v1, const int16_t * v2, int order) |
| 338 | { |
| 339 | int res = 0; |
| 340 | |
| 341 | while (order--) |
| 342 | res += *v1++ * *v2++; |
| 343 | |
| 344 | return res; |
| 345 | } |
| 346 | |
| 347 | static av_cold int decoder_init(AVCodecContext * avctx) |
| 348 | { |
| 349 | G729Context* ctx = avctx->priv_data; |
| 350 | int i,k; |
| 351 | |
| 352 | if (avctx->channels != 1) { |
| 353 | av_log(avctx, AV_LOG_ERROR, "Only mono sound is supported (requested channels: %d).\n", avctx->channels); |
| 354 | return AVERROR(EINVAL); |
| 355 | } |
| 356 | avctx->sample_fmt = AV_SAMPLE_FMT_S16; |
| 357 | |
| 358 | /* Both 8kbit/s and 6.4kbit/s modes uses two subframes per frame. */ |
| 359 | avctx->frame_size = SUBFRAME_SIZE << 1; |
| 360 | |
| 361 | ctx->gain_coeff = 16384; // 1.0 in (1.14) |
| 362 | |
| 363 | for (k = 0; k < MA_NP + 1; k++) { |
| 364 | ctx->past_quantizer_outputs[k] = ctx->past_quantizer_output_buf[k]; |
| 365 | for (i = 1; i < 11; i++) |
| 366 | ctx->past_quantizer_outputs[k][i - 1] = (18717 * i) >> 3; |
| 367 | } |
| 368 | |
| 369 | ctx->lsp[0] = ctx->lsp_buf[0]; |
| 370 | ctx->lsp[1] = ctx->lsp_buf[1]; |
| 371 | memcpy(ctx->lsp[0], lsp_init, 10 * sizeof(int16_t)); |
| 372 | |
| 373 | ctx->exc = &ctx->exc_base[PITCH_DELAY_MAX+INTERPOL_LEN]; |
| 374 | |
| 375 | ctx->pitch_delay_int_prev = PITCH_DELAY_MIN; |
| 376 | |
| 377 | /* random seed initialization */ |
| 378 | ctx->rand_value = 21845; |
| 379 | |
| 380 | /* quantized prediction error */ |
| 381 | for(i=0; i<4; i++) |
| 382 | ctx->quant_energy[i] = -14336; // -14 in (5.10) |
| 383 | |
| 384 | ff_audiodsp_init(&ctx->adsp); |
| 385 | ctx->adsp.scalarproduct_int16 = scalarproduct_int16_c; |
| 386 | |
| 387 | return 0; |
| 388 | } |
| 389 | |
| 390 | static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame_ptr, |
| 391 | AVPacket *avpkt) |
| 392 | { |
| 393 | const uint8_t *buf = avpkt->data; |
| 394 | int buf_size = avpkt->size; |
| 395 | int16_t *out_frame; |
| 396 | GetBitContext gb; |
| 397 | const G729FormatDescription *format; |
| 398 | int frame_erasure = 0; ///< frame erasure detected during decoding |
| 399 | int bad_pitch = 0; ///< parity check failed |
| 400 | int i; |
| 401 | int16_t *tmp; |
| 402 | G729Formats packet_type; |
| 403 | G729Context *ctx = avctx->priv_data; |
| 404 | int16_t lp[2][11]; // (3.12) |
| 405 | uint8_t ma_predictor; ///< switched MA predictor of LSP quantizer |
| 406 | uint8_t quantizer_1st; ///< first stage vector of quantizer |
| 407 | uint8_t quantizer_2nd_lo; ///< second stage lower vector of quantizer (size in bits) |
| 408 | uint8_t quantizer_2nd_hi; ///< second stage higher vector of quantizer (size in bits) |
| 409 | |
| 410 | int pitch_delay_int[2]; // pitch delay, integer part |
| 411 | int pitch_delay_3x; // pitch delay, multiplied by 3 |
| 412 | int16_t fc[SUBFRAME_SIZE]; // fixed-codebook vector |
| 413 | int16_t synth[SUBFRAME_SIZE+10]; // fixed-codebook vector |
| 414 | int j, ret; |
| 415 | int gain_before, gain_after; |
| 416 | int is_periodic = 0; // whether one of the subframes is declared as periodic or not |
| 417 | AVFrame *frame = data; |
| 418 | |
| 419 | frame->nb_samples = SUBFRAME_SIZE<<1; |
| 420 | if ((ret = ff_get_buffer(avctx, frame, 0)) < 0) |
| 421 | return ret; |
| 422 | out_frame = (int16_t*) frame->data[0]; |
| 423 | |
| 424 | if (buf_size == 10) { |
| 425 | packet_type = FORMAT_G729_8K; |
| 426 | format = &format_g729_8k; |
| 427 | //Reset voice decision |
| 428 | ctx->onset = 0; |
| 429 | ctx->voice_decision = DECISION_VOICE; |
| 430 | av_log(avctx, AV_LOG_DEBUG, "Packet type: %s\n", "G.729 @ 8kbit/s"); |
| 431 | } else if (buf_size == 8) { |
| 432 | packet_type = FORMAT_G729D_6K4; |
| 433 | format = &format_g729d_6k4; |
| 434 | av_log(avctx, AV_LOG_DEBUG, "Packet type: %s\n", "G.729D @ 6.4kbit/s"); |
| 435 | } else { |
| 436 | av_log(avctx, AV_LOG_ERROR, "Packet size %d is unknown.\n", buf_size); |
| 437 | return AVERROR_INVALIDDATA; |
| 438 | } |
| 439 | |
| 440 | for (i=0; i < buf_size; i++) |
| 441 | frame_erasure |= buf[i]; |
| 442 | frame_erasure = !frame_erasure; |
| 443 | |
| 444 | init_get_bits(&gb, buf, 8*buf_size); |
| 445 | |
| 446 | ma_predictor = get_bits(&gb, 1); |
| 447 | quantizer_1st = get_bits(&gb, VQ_1ST_BITS); |
| 448 | quantizer_2nd_lo = get_bits(&gb, VQ_2ND_BITS); |
| 449 | quantizer_2nd_hi = get_bits(&gb, VQ_2ND_BITS); |
| 450 | |
| 451 | if(frame_erasure) |
| 452 | lsf_restore_from_previous(ctx->lsfq, ctx->past_quantizer_outputs, |
| 453 | ctx->ma_predictor_prev); |
| 454 | else { |
| 455 | lsf_decode(ctx->lsfq, ctx->past_quantizer_outputs, |
| 456 | ma_predictor, |
| 457 | quantizer_1st, quantizer_2nd_lo, quantizer_2nd_hi); |
| 458 | ctx->ma_predictor_prev = ma_predictor; |
| 459 | } |
| 460 | |
| 461 | tmp = ctx->past_quantizer_outputs[MA_NP]; |
| 462 | memmove(ctx->past_quantizer_outputs + 1, ctx->past_quantizer_outputs, |
| 463 | MA_NP * sizeof(int16_t*)); |
| 464 | ctx->past_quantizer_outputs[0] = tmp; |
| 465 | |
| 466 | ff_acelp_lsf2lsp(ctx->lsp[1], ctx->lsfq, 10); |
| 467 | |
| 468 | ff_acelp_lp_decode(&lp[0][0], &lp[1][0], ctx->lsp[1], ctx->lsp[0], 10); |
| 469 | |
| 470 | FFSWAP(int16_t*, ctx->lsp[1], ctx->lsp[0]); |
| 471 | |
| 472 | for (i = 0; i < 2; i++) { |
| 473 | int gain_corr_factor; |
| 474 | |
| 475 | uint8_t ac_index; ///< adaptive codebook index |
| 476 | uint8_t pulses_signs; ///< fixed-codebook vector pulse signs |
| 477 | int fc_indexes; ///< fixed-codebook indexes |
| 478 | uint8_t gc_1st_index; ///< gain codebook (first stage) index |
| 479 | uint8_t gc_2nd_index; ///< gain codebook (second stage) index |
| 480 | |
| 481 | ac_index = get_bits(&gb, format->ac_index_bits[i]); |
| 482 | if(!i && format->parity_bit) |
| 483 | bad_pitch = get_parity(ac_index) == get_bits1(&gb); |
| 484 | fc_indexes = get_bits(&gb, format->fc_indexes_bits); |
| 485 | pulses_signs = get_bits(&gb, format->fc_signs_bits); |
| 486 | gc_1st_index = get_bits(&gb, format->gc_1st_index_bits); |
| 487 | gc_2nd_index = get_bits(&gb, format->gc_2nd_index_bits); |
| 488 | |
| 489 | if (frame_erasure) |
| 490 | pitch_delay_3x = 3 * ctx->pitch_delay_int_prev; |
| 491 | else if(!i) { |
| 492 | if (bad_pitch) |
| 493 | pitch_delay_3x = 3 * ctx->pitch_delay_int_prev; |
| 494 | else |
| 495 | pitch_delay_3x = ff_acelp_decode_8bit_to_1st_delay3(ac_index); |
| 496 | } else { |
| 497 | int pitch_delay_min = av_clip(ctx->pitch_delay_int_prev - 5, |
| 498 | PITCH_DELAY_MIN, PITCH_DELAY_MAX - 9); |
| 499 | |
| 500 | if(packet_type == FORMAT_G729D_6K4) |
| 501 | pitch_delay_3x = ff_acelp_decode_4bit_to_2nd_delay3(ac_index, pitch_delay_min); |
| 502 | else |
| 503 | pitch_delay_3x = ff_acelp_decode_5_6_bit_to_2nd_delay3(ac_index, pitch_delay_min); |
| 504 | } |
| 505 | |
| 506 | /* Round pitch delay to nearest (used everywhere except ff_acelp_interpolate). */ |
| 507 | pitch_delay_int[i] = (pitch_delay_3x + 1) / 3; |
| 508 | if (pitch_delay_int[i] > PITCH_DELAY_MAX) { |
| 509 | av_log(avctx, AV_LOG_WARNING, "pitch_delay_int %d is too large\n", pitch_delay_int[i]); |
| 510 | pitch_delay_int[i] = PITCH_DELAY_MAX; |
| 511 | } |
| 512 | |
| 513 | if (frame_erasure) { |
| 514 | ctx->rand_value = g729_prng(ctx->rand_value); |
| 515 | fc_indexes = ctx->rand_value & ((1 << format->fc_indexes_bits) - 1); |
| 516 | |
| 517 | ctx->rand_value = g729_prng(ctx->rand_value); |
| 518 | pulses_signs = ctx->rand_value; |
| 519 | } |
| 520 | |
| 521 | |
| 522 | memset(fc, 0, sizeof(int16_t) * SUBFRAME_SIZE); |
| 523 | switch (packet_type) { |
| 524 | case FORMAT_G729_8K: |
| 525 | ff_acelp_fc_pulse_per_track(fc, ff_fc_4pulses_8bits_tracks_13, |
| 526 | ff_fc_4pulses_8bits_track_4, |
| 527 | fc_indexes, pulses_signs, 3, 3); |
| 528 | break; |
| 529 | case FORMAT_G729D_6K4: |
| 530 | ff_acelp_fc_pulse_per_track(fc, ff_fc_2pulses_9bits_track1_gray, |
| 531 | ff_fc_2pulses_9bits_track2_gray, |
| 532 | fc_indexes, pulses_signs, 1, 4); |
| 533 | break; |
| 534 | } |
| 535 | |
| 536 | /* |
| 537 | This filter enhances harmonic components of the fixed-codebook vector to |
| 538 | improve the quality of the reconstructed speech. |
| 539 | |
| 540 | / fc_v[i], i < pitch_delay |
| 541 | fc_v[i] = < |
| 542 | \ fc_v[i] + gain_pitch * fc_v[i-pitch_delay], i >= pitch_delay |
| 543 | */ |
| 544 | ff_acelp_weighted_vector_sum(fc + pitch_delay_int[i], |
| 545 | fc + pitch_delay_int[i], |
| 546 | fc, 1 << 14, |
| 547 | av_clip(ctx->past_gain_pitch[0], SHARP_MIN, SHARP_MAX), |
| 548 | 0, 14, |
| 549 | SUBFRAME_SIZE - pitch_delay_int[i]); |
| 550 | |
| 551 | memmove(ctx->past_gain_pitch+1, ctx->past_gain_pitch, 5 * sizeof(int16_t)); |
| 552 | ctx->past_gain_code[1] = ctx->past_gain_code[0]; |
| 553 | |
| 554 | if (frame_erasure) { |
| 555 | ctx->past_gain_pitch[0] = (29491 * ctx->past_gain_pitch[0]) >> 15; // 0.90 (0.15) |
| 556 | ctx->past_gain_code[0] = ( 2007 * ctx->past_gain_code[0] ) >> 11; // 0.98 (0.11) |
| 557 | |
| 558 | gain_corr_factor = 0; |
| 559 | } else { |
| 560 | if (packet_type == FORMAT_G729D_6K4) { |
| 561 | ctx->past_gain_pitch[0] = cb_gain_1st_6k4[gc_1st_index][0] + |
| 562 | cb_gain_2nd_6k4[gc_2nd_index][0]; |
| 563 | gain_corr_factor = cb_gain_1st_6k4[gc_1st_index][1] + |
| 564 | cb_gain_2nd_6k4[gc_2nd_index][1]; |
| 565 | |
| 566 | /* Without check below overflow can occur in ff_acelp_update_past_gain. |
| 567 | It is not issue for G.729, because gain_corr_factor in it's case is always |
| 568 | greater than 1024, while in G.729D it can be even zero. */ |
| 569 | gain_corr_factor = FFMAX(gain_corr_factor, 1024); |
| 570 | #ifndef G729_BITEXACT |
| 571 | gain_corr_factor >>= 1; |
| 572 | #endif |
| 573 | } else { |
| 574 | ctx->past_gain_pitch[0] = cb_gain_1st_8k[gc_1st_index][0] + |
| 575 | cb_gain_2nd_8k[gc_2nd_index][0]; |
| 576 | gain_corr_factor = cb_gain_1st_8k[gc_1st_index][1] + |
| 577 | cb_gain_2nd_8k[gc_2nd_index][1]; |
| 578 | } |
| 579 | |
| 580 | /* Decode the fixed-codebook gain. */ |
| 581 | ctx->past_gain_code[0] = ff_acelp_decode_gain_code(&ctx->adsp, gain_corr_factor, |
| 582 | fc, MR_ENERGY, |
| 583 | ctx->quant_energy, |
| 584 | ma_prediction_coeff, |
| 585 | SUBFRAME_SIZE, 4); |
| 586 | #ifdef G729_BITEXACT |
| 587 | /* |
| 588 | This correction required to get bit-exact result with |
| 589 | reference code, because gain_corr_factor in G.729D is |
| 590 | two times larger than in original G.729. |
| 591 | |
| 592 | If bit-exact result is not issue then gain_corr_factor |
| 593 | can be simpler divided by 2 before call to g729_get_gain_code |
| 594 | instead of using correction below. |
| 595 | */ |
| 596 | if (packet_type == FORMAT_G729D_6K4) { |
| 597 | gain_corr_factor >>= 1; |
| 598 | ctx->past_gain_code[0] >>= 1; |
| 599 | } |
| 600 | #endif |
| 601 | } |
| 602 | ff_acelp_update_past_gain(ctx->quant_energy, gain_corr_factor, 2, frame_erasure); |
| 603 | |
| 604 | /* Routine requires rounding to lowest. */ |
| 605 | ff_acelp_interpolate(ctx->exc + i * SUBFRAME_SIZE, |
| 606 | ctx->exc + i * SUBFRAME_SIZE - pitch_delay_3x / 3, |
| 607 | ff_acelp_interp_filter, 6, |
| 608 | (pitch_delay_3x % 3) << 1, |
| 609 | 10, SUBFRAME_SIZE); |
| 610 | |
| 611 | ff_acelp_weighted_vector_sum(ctx->exc + i * SUBFRAME_SIZE, |
| 612 | ctx->exc + i * SUBFRAME_SIZE, fc, |
| 613 | (!ctx->was_periodic && frame_erasure) ? 0 : ctx->past_gain_pitch[0], |
| 614 | ( ctx->was_periodic && frame_erasure) ? 0 : ctx->past_gain_code[0], |
| 615 | 1 << 13, 14, SUBFRAME_SIZE); |
| 616 | |
| 617 | memcpy(synth, ctx->syn_filter_data, 10 * sizeof(int16_t)); |
| 618 | |
| 619 | if (ff_celp_lp_synthesis_filter( |
| 620 | synth+10, |
| 621 | &lp[i][1], |
| 622 | ctx->exc + i * SUBFRAME_SIZE, |
| 623 | SUBFRAME_SIZE, |
| 624 | 10, |
| 625 | 1, |
| 626 | 0, |
| 627 | 0x800)) |
| 628 | /* Overflow occurred, downscale excitation signal... */ |
| 629 | for (j = 0; j < 2 * SUBFRAME_SIZE + PITCH_DELAY_MAX + INTERPOL_LEN; j++) |
| 630 | ctx->exc_base[j] >>= 2; |
| 631 | |
| 632 | /* ... and make synthesis again. */ |
| 633 | if (packet_type == FORMAT_G729D_6K4) { |
| 634 | int16_t exc_new[SUBFRAME_SIZE]; |
| 635 | |
| 636 | ctx->onset = g729d_onset_decision(ctx->onset, ctx->past_gain_code); |
| 637 | ctx->voice_decision = g729d_voice_decision(ctx->onset, ctx->voice_decision, ctx->past_gain_pitch); |
| 638 | |
| 639 | g729d_get_new_exc(exc_new, ctx->exc + i * SUBFRAME_SIZE, fc, ctx->voice_decision, ctx->past_gain_code[0], SUBFRAME_SIZE); |
| 640 | |
| 641 | ff_celp_lp_synthesis_filter( |
| 642 | synth+10, |
| 643 | &lp[i][1], |
| 644 | exc_new, |
| 645 | SUBFRAME_SIZE, |
| 646 | 10, |
| 647 | 0, |
| 648 | 0, |
| 649 | 0x800); |
| 650 | } else { |
| 651 | ff_celp_lp_synthesis_filter( |
| 652 | synth+10, |
| 653 | &lp[i][1], |
| 654 | ctx->exc + i * SUBFRAME_SIZE, |
| 655 | SUBFRAME_SIZE, |
| 656 | 10, |
| 657 | 0, |
| 658 | 0, |
| 659 | 0x800); |
| 660 | } |
| 661 | /* Save data (without postfilter) for use in next subframe. */ |
| 662 | memcpy(ctx->syn_filter_data, synth+SUBFRAME_SIZE, 10 * sizeof(int16_t)); |
| 663 | |
| 664 | /* Calculate gain of unfiltered signal for use in AGC. */ |
| 665 | gain_before = 0; |
| 666 | for (j = 0; j < SUBFRAME_SIZE; j++) |
| 667 | gain_before += FFABS(synth[j+10]); |
| 668 | |
| 669 | /* Call postfilter and also update voicing decision for use in next frame. */ |
| 670 | ff_g729_postfilter( |
| 671 | &ctx->adsp, |
| 672 | &ctx->ht_prev_data, |
| 673 | &is_periodic, |
| 674 | &lp[i][0], |
| 675 | pitch_delay_int[0], |
| 676 | ctx->residual, |
| 677 | ctx->res_filter_data, |
| 678 | ctx->pos_filter_data, |
| 679 | synth+10, |
| 680 | SUBFRAME_SIZE); |
| 681 | |
| 682 | /* Calculate gain of filtered signal for use in AGC. */ |
| 683 | gain_after = 0; |
| 684 | for(j=0; j<SUBFRAME_SIZE; j++) |
| 685 | gain_after += FFABS(synth[j+10]); |
| 686 | |
| 687 | ctx->gain_coeff = ff_g729_adaptive_gain_control( |
| 688 | gain_before, |
| 689 | gain_after, |
| 690 | synth+10, |
| 691 | SUBFRAME_SIZE, |
| 692 | ctx->gain_coeff); |
| 693 | |
| 694 | if (frame_erasure) |
| 695 | ctx->pitch_delay_int_prev = FFMIN(ctx->pitch_delay_int_prev + 1, PITCH_DELAY_MAX); |
| 696 | else |
| 697 | ctx->pitch_delay_int_prev = pitch_delay_int[i]; |
| 698 | |
| 699 | memcpy(synth+8, ctx->hpf_z, 2*sizeof(int16_t)); |
| 700 | ff_acelp_high_pass_filter( |
| 701 | out_frame + i*SUBFRAME_SIZE, |
| 702 | ctx->hpf_f, |
| 703 | synth+10, |
| 704 | SUBFRAME_SIZE); |
| 705 | memcpy(ctx->hpf_z, synth+8+SUBFRAME_SIZE, 2*sizeof(int16_t)); |
| 706 | } |
| 707 | |
| 708 | ctx->was_periodic = is_periodic; |
| 709 | |
| 710 | /* Save signal for use in next frame. */ |
| 711 | memmove(ctx->exc_base, ctx->exc_base + 2 * SUBFRAME_SIZE, (PITCH_DELAY_MAX+INTERPOL_LEN)*sizeof(int16_t)); |
| 712 | |
| 713 | *got_frame_ptr = 1; |
| 714 | return buf_size; |
| 715 | } |
| 716 | |
| 717 | AVCodec ff_g729_decoder = { |
| 718 | .name = "g729", |
| 719 | .long_name = NULL_IF_CONFIG_SMALL("G.729"), |
| 720 | .type = AVMEDIA_TYPE_AUDIO, |
| 721 | .id = AV_CODEC_ID_G729, |
| 722 | .priv_data_size = sizeof(G729Context), |
| 723 | .init = decoder_init, |
| 724 | .decode = decode_frame, |
| 725 | .capabilities = CODEC_CAP_DR1, |
| 726 | }; |