git.sesse.net Git - ffmpeg/blob - libavcodec/g729dec.c

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