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

   1 /*
   2  * QCELP decoder
   3  * Copyright (c) 2007 Reynaldo H. Verdejo Pinochet
   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 /**
  23  * @file qcelpdec.c
  24  * QCELP decoder
  25  * @author Reynaldo H. Verdejo Pinochet
  26  * @remark FFmpeg merging spearheaded by Kenan Gillet
  27  * @remark Development mentored by Benjamin Larson
  28  */
  29
  30 #include <stddef.h>
  31
  32 #include "avcodec.h"
  33 #include "bitstream.h"
  34
  35 #include "qcelpdata.h"
  36
  37 #include "celp_math.h"
  38 #include "celp_filters.h"
  39
  40 #undef NDEBUG
  41 #include <assert.h>
  42
  43 typedef enum
  44 {
  45     I_F_Q = -1,    /*!< insufficient frame quality */
  46     SILENCE,
  47     RATE_OCTAVE,
  48     RATE_QUARTER,
  49     RATE_HALF,
  50     RATE_FULL
  51 } qcelp_packet_rate;
  52
  53 typedef struct
  54 {
  55     GetBitContext     gb;
  56     qcelp_packet_rate bitrate;
  57     QCELPFrame        frame;    /*!< unpacked data frame */
  58
  59     uint8_t  erasure_count;
  60     uint8_t  octave_count;      /*!< count the consecutive RATE_OCTAVE frames */
  61     float    prev_lspf[10];
  62     float    predictor_lspf[10];/*!< LSP predictor for RATE_OCTAVE and I_F_Q */
  63     float    pitch_synthesis_filter_mem[303];
  64     float    pitch_pre_filter_mem[303];
  65     float    rnd_fir_filter_mem[180];
  66     float    formant_mem[170];
  67     float    last_codebook_gain;
  68     int      prev_g1[2];
  69     int      prev_bitrate;
  70     float    pitch_gain[4];
  71     uint8_t  pitch_lag[4];
  72     uint16_t first16bits;
  73 } QCELPContext;
  74
  75 /**
  76  * Reconstructs LPC coefficients from the line spectral pair frequencies.
  77  *
  78  * TIA/EIA/IS-733 2.4.3.3.5
  79  */
  80 void ff_qcelp_lspf2lpc(const float *lspf, float *lpc);
  81
  82 static void weighted_vector_sumf(float *out, const float *in_a,
  83                                  const float *in_b, float weight_coeff_a,
  84                                  float weight_coeff_b, int length)
  85 {
  86     int i;
  87
  88     for(i=0; i<length; i++)
  89         out[i] = weight_coeff_a * in_a[i]
  90                + weight_coeff_b * in_b[i];
  91 }
  92
  93 /**
  94  * Initialize the speech codec according to the specification.
  95  *
  96  * TIA/EIA/IS-733 2.4.9
  97  */
  98 static av_cold int qcelp_decode_init(AVCodecContext *avctx)
  99 {
 100     QCELPContext *q = avctx->priv_data;
 101     int i;
 102
 103     avctx->sample_fmt = SAMPLE_FMT_FLT;
 104
 105     for(i=0; i<10; i++)
 106         q->prev_lspf[i] = (i+1)/11.;
 107
 108     return 0;
 109 }
 110
 111 /**
 112  * Decodes the 10 quantized LSP frequencies from the LSPV/LSP
 113  * transmission codes of any bitrate and checks for badly received packets.
 114  *
 115  * @param q the context
 116  * @param lspf line spectral pair frequencies
 117  *
 118  * @return 0 on success, -1 if the packet is badly received
 119  *
 120  * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3
 121  */
 122 static int decode_lspf(QCELPContext *q, float *lspf)
 123 {
 124     int i;
 125     float tmp_lspf, smooth, erasure_coeff;
 126     const float *predictors;
 127
 128     if(q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q)
 129     {
 130         predictors = (q->prev_bitrate != RATE_OCTAVE &&
 131                        q->prev_bitrate != I_F_Q ?
 132                        q->prev_lspf : q->predictor_lspf);
 133
 134         if(q->bitrate == RATE_OCTAVE)
 135         {
 136             q->octave_count++;
 137
 138             for(i=0; i<10; i++)
 139             {
 140                 q->predictor_lspf[i] =
 141                              lspf[i] = (q->frame.lspv[i] ?  QCELP_LSP_SPREAD_FACTOR
 142                                                          : -QCELP_LSP_SPREAD_FACTOR)
 143                                      + predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR
 144                                      + (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR)/11);
 145             }
 146             smooth = (q->octave_count < 10 ? .875 : 0.1);
 147         }else
 148         {
 149             erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR;
 150
 151             assert(q->bitrate == I_F_Q);
 152
 153             if(q->erasure_count > 1)
 154                 erasure_coeff *= (q->erasure_count < 4 ? 0.9 : 0.7);
 155
 156             for(i=0; i<10; i++)
 157             {
 158                 q->predictor_lspf[i] =
 159                              lspf[i] = (i + 1) * ( 1 - erasure_coeff)/11
 160                                      + erasure_coeff * predictors[i];
 161             }
 162             smooth = 0.125;
 163         }
 164
 165         // Check the stability of the LSP frequencies.
 166         lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR);
 167         for(i=1; i<10; i++)
 168             lspf[i] = FFMAX(lspf[i], (lspf[i-1] + QCELP_LSP_SPREAD_FACTOR));
 169
 170         lspf[9] = FFMIN(lspf[9], (1.0 - QCELP_LSP_SPREAD_FACTOR));
 171         for(i=9; i>0; i--)
 172             lspf[i-1] = FFMIN(lspf[i-1], (lspf[i] - QCELP_LSP_SPREAD_FACTOR));
 173
 174         // Low-pass filter the LSP frequencies.
 175         weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0-smooth, 10);
 176     }else
 177     {
 178         q->octave_count = 0;
 179
 180         tmp_lspf = 0.;
 181         for(i=0; i<5 ; i++)
 182         {
 183             lspf[2*i+0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001;
 184             lspf[2*i+1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001;
 185         }
 186
 187         // Check for badly received packets.
 188         if(q->bitrate == RATE_QUARTER)
 189         {
 190             if(lspf[9] <= .70 || lspf[9] >=  .97)
 191                 return -1;
 192             for(i=3; i<10; i++)
 193                 if(fabs(lspf[i] - lspf[i-2]) < .08)
 194                     return -1;
 195         }else
 196         {
 197             if(lspf[9] <= .66 || lspf[9] >= .985)
 198                 return -1;
 199             for(i=4; i<10; i++)
 200                 if (fabs(lspf[i] - lspf[i-4]) < .0931)
 201                     return -1;
 202         }
 203     }
 204     return 0;
 205 }
 206
 207 /**
 208  * Converts codebook transmission codes to GAIN and INDEX.
 209  *
 210  * @param q the context
 211  * @param gain array holding the decoded gain
 212  *
 213  * TIA/EIA/IS-733 2.4.6.2
 214  */
 215 static void decode_gain_and_index(QCELPContext  *q,
 216                                   float *gain) {
 217     int   i, subframes_count, g1[16];
 218     float slope;
 219
 220     if(q->bitrate >= RATE_QUARTER)
 221     {
 222         switch(q->bitrate)
 223         {
 224             case RATE_FULL: subframes_count = 16; break;
 225             case RATE_HALF: subframes_count = 4;  break;
 226             default:        subframes_count = 5;
 227         }
 228         for(i=0; i<subframes_count; i++)
 229         {
 230             g1[i] = 4 * q->frame.cbgain[i];
 231             if(q->bitrate == RATE_FULL && !((i+1) & 3))
 232             {
 233                 g1[i] += av_clip((g1[i-1] + g1[i-2] + g1[i-3]) / 3 - 6, 0, 32);
 234             }
 235
 236             gain[i] = qcelp_g12ga[g1[i]];
 237
 238             if(q->frame.cbsign[i])
 239             {
 240                 gain[i] = -gain[i];
 241                 q->frame.cindex[i] = (q->frame.cindex[i]-89) & 127;
 242             }
 243         }
 244
 245         q->prev_g1[0] = g1[i-2];
 246         q->prev_g1[1] = g1[i-1];
 247         q->last_codebook_gain = qcelp_g12ga[g1[i-1]];
 248
 249         if(q->bitrate == RATE_QUARTER)
 250         {
 251             // Provide smoothing of the unvoiced excitation energy.
 252             gain[7] =     gain[4];
 253             gain[6] = 0.4*gain[3] + 0.6*gain[4];
 254             gain[5] =     gain[3];
 255             gain[4] = 0.8*gain[2] + 0.2*gain[3];
 256             gain[3] = 0.2*gain[1] + 0.8*gain[2];
 257             gain[2] =     gain[1];
 258             gain[1] = 0.6*gain[0] + 0.4*gain[1];
 259         }
 260     }else
 261     {
 262         if(q->bitrate == RATE_OCTAVE)
 263         {
 264             g1[0] = 2 * q->frame.cbgain[0]
 265                   + av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54);
 266             subframes_count = 8;
 267         }else
 268         {
 269             assert(q->bitrate == I_F_Q);
 270
 271             g1[0] = q->prev_g1[1];
 272             switch(q->erasure_count)
 273             {
 274                 case 1 : break;
 275                 case 2 : g1[0] -= 1; break;
 276                 case 3 : g1[0] -= 2; break;
 277                 default: g1[0] -= 6;
 278             }
 279             if(g1[0] < 0)
 280                 g1[0] = 0;
 281             subframes_count = 4;
 282         }
 283         // This interpolation is done to produce smoother background noise.
 284         slope = 0.5*(qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count;
 285         for(i=1; i<=subframes_count; i++)
 286             gain[i-1] = q->last_codebook_gain + slope * i;
 287
 288         q->last_codebook_gain = gain[i-2];
 289         q->prev_g1[0] = q->prev_g1[1];
 290         q->prev_g1[1] = g1[0];
 291     }
 292 }
 293
 294 /**
 295  * If the received packet is Rate 1/4 a further sanity check is made of the
 296  * codebook gain.
 297  *
 298  * @param cbgain the unpacked cbgain array
 299  * @return -1 if the sanity check fails, 0 otherwise
 300  *
 301  * TIA/EIA/IS-733 2.4.8.7.3
 302  */
 303 static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)
 304 {
 305     int i, diff, prev_diff=0;
 306
 307     for(i=1; i<5; i++)
 308     {
 309         diff = cbgain[i] - cbgain[i-1];
 310         if(FFABS(diff) > 10)
 311             return -1;
 312         else if(FFABS(diff - prev_diff) > 12)
 313             return -1;
 314         prev_diff = diff;
 315     }
 316     return 0;
 317 }
 318
 319 /**
 320  * Computes the scaled codebook vector Cdn From INDEX and GAIN
 321  * for all rates.
 322  *
 323  * The specification lacks some information here.
 324  *
 325  * TIA/EIA/IS-733 has an omission on the codebook index determination
 326  * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says
 327  * you have to subtract the decoded index parameter from the given scaled
 328  * codebook vector index 'n' to get the desired circular codebook index, but
 329  * it does not mention that you have to clamp 'n' to [0-9] in order to get
 330  * RI-compliant results.
 331  *
 332  * The reason for this mistake seems to be the fact they forgot to mention you
 333  * have to do these calculations per codebook subframe and adjust given
 334  * equation values accordingly.
 335  *
 336  * @param q the context
 337  * @param gain array holding the 4 pitch subframe gain values
 338  * @param cdn_vector array for the generated scaled codebook vector
 339  */
 340 static void compute_svector(QCELPContext *q, const float *gain,
 341                             float *cdn_vector)
 342 {
 343     int      i, j, k;
 344     uint16_t cbseed, cindex;
 345     float    *rnd, tmp_gain, fir_filter_value;
 346
 347     switch(q->bitrate)
 348     {
 349         case RATE_FULL:
 350             for(i=0; i<16; i++)
 351             {
 352                 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
 353                 cindex = -q->frame.cindex[i];
 354                 for(j=0; j<10; j++)
 355                     *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cindex++ & 127];
 356             }
 357         break;
 358         case RATE_HALF:
 359             for(i=0; i<4; i++)
 360             {
 361                 tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO;
 362                 cindex = -q->frame.cindex[i];
 363                 for (j = 0; j < 40; j++)
 364                 *cdn_vector++ = tmp_gain * qcelp_rate_half_codebook[cindex++ & 127];
 365             }
 366         break;
 367         case RATE_QUARTER:
 368             cbseed = (0x0003 & q->frame.lspv[4])<<14 |
 369                      (0x003F & q->frame.lspv[3])<< 8 |
 370                      (0x0060 & q->frame.lspv[2])<< 1 |
 371                      (0x0007 & q->frame.lspv[1])<< 3 |
 372                      (0x0038 & q->frame.lspv[0])>> 3 ;
 373             rnd = q->rnd_fir_filter_mem + 20;
 374             for(i=0; i<8; i++)
 375             {
 376                 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
 377                 for(k=0; k<20; k++)
 378                 {
 379                     cbseed = 521 * cbseed + 259;
 380                     *rnd = (int16_t)cbseed;
 381
 382                     // FIR filter
 383                     fir_filter_value = 0.0;
 384                     for(j=0; j<10; j++)
 385                         fir_filter_value += qcelp_rnd_fir_coefs[j ]
 386                                           * (rnd[-j ] + rnd[-20+j]);
 387
 388                     fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10];
 389                     *cdn_vector++ = tmp_gain * fir_filter_value;
 390                     rnd++;
 391                 }
 392             }
 393             memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160, 20 * sizeof(float));
 394         break;
 395         case RATE_OCTAVE:
 396             cbseed = q->first16bits;
 397             for(i=0; i<8; i++)
 398             {
 399                 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
 400                 for(j=0; j<20; j++)
 401                 {
 402                     cbseed = 521 * cbseed + 259;
 403                     *cdn_vector++ = tmp_gain * (int16_t)cbseed;
 404                 }
 405             }
 406         break;
 407         case I_F_Q:
 408             cbseed = -44; // random codebook index
 409             for(i=0; i<4; i++)
 410             {
 411                 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
 412                 for(j=0; j<40; j++)
 413                     *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cbseed++ & 127];
 414             }
 415         break;
 416     }
 417 }
 418
 419 /**
 420  * Apply generic gain control.
 421  *
 422  * @param v_out output vector
 423  * @param v_in gain-controlled vector
 424  * @param v_ref vector to control gain of
 425  *
 426  * FIXME: If v_ref is a zero vector, it energy is zero
 427  *        and the behavior of the gain control is
 428  *        undefined in the specs.
 429  *
 430  * TIA/EIA/IS-733 2.4.8.3-2/3/4/5, 2.4.8.6
 431  */
 432 static void apply_gain_ctrl(float *v_out, const float *v_ref,
 433                             const float *v_in)
 434 {
 435     int   i, j, len;
 436     float scalefactor;
 437
 438     for(i=0, j=0; i<4; i++)
 439     {
 440         scalefactor = ff_dot_productf(v_in + j, v_in + j, 40);
 441         if(scalefactor)
 442             scalefactor = sqrt(ff_dot_productf(v_ref + j, v_ref + j, 40)
 443                         / scalefactor);
 444         else
 445             av_log_missing_feature(NULL, "Zero energy for gain control", 1);
 446         for(len=j+40; j<len; j++)
 447             v_out[j] = scalefactor * v_in[j];
 448     }
 449 }
 450
 451 /**
 452  * Apply filter in pitch-subframe steps.
 453  *
 454  * @param memory buffer for the previous state of the filter
 455  *        - must be able to contain 303 elements
 456  *        - the 143 first elements are from the previous state
 457  *        - the next 160 are for output
 458  * @param v_in input filter vector
 459  * @param gain per-subframe gain array, each element is between 0.0 and 2.0
 460  * @param lag per-subframe lag array, each element is
 461  *        - between 16 and 143 if its corresponding pfrac is 0,
 462  *        - between 16 and 139 otherwise
 463  * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0
 464  *        otherwise
 465  *
 466  * @return filter output vector
 467  */
 468 static const float *do_pitchfilter(float memory[303], const float v_in[160],
 469                                    const float gain[4], const uint8_t *lag,
 470                                    const uint8_t pfrac[4])
 471 {
 472     int         i, j;
 473     float       *v_lag, *v_out;
 474     const float *v_len;
 475
 476     v_out = memory + 143; // Output vector starts at memory[143].
 477
 478     for(i=0; i<4; i++)
 479     {
 480         if(gain[i])
 481         {
 482             v_lag = memory + 143 + 40 * i - lag[i];
 483             for(v_len=v_in+40; v_in<v_len; v_in++)
 484             {
 485                 if(pfrac[i]) // If it is a fractional lag...
 486                 {
 487                     for(j=0, *v_out=0.; j<4; j++)
 488                         *v_out += qcelp_hammsinc_table[j] * (v_lag[j-4] + v_lag[3-j]);
 489                 }else
 490                     *v_out = *v_lag;
 491
 492                 *v_out = *v_in + gain[i] * *v_out;
 493
 494                 v_lag++;
 495                 v_out++;
 496             }
 497         }else
 498         {
 499             memcpy(v_out, v_in, 40 * sizeof(float));
 500             v_in  += 40;
 501             v_out += 40;
 502         }
 503     }
 504
 505     memmove(memory, memory + 160, 143 * sizeof(float));
 506     return memory + 143;
 507 }
 508
 509 /**
 510  * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector.
 511  * TIA/EIA/IS-733 2.4.5.2
 512  *
 513  * @param q the context
 514  * @param cdn_vector the scaled codebook vector
 515  */
 516 static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
 517 {
 518     int         i;
 519     const float *v_synthesis_filtered, *v_pre_filtered;
 520
 521     if(q->bitrate >= RATE_HALF ||
 522        (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF)))
 523     {
 524
 525         if(q->bitrate >= RATE_HALF)
 526         {
 527
 528             // Compute gain & lag for the whole frame.
 529             for(i=0; i<4; i++)
 530             {
 531                 q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0;
 532
 533                 q->pitch_lag[i] = q->frame.plag[i] + 16;
 534             }
 535         }else
 536         {
 537             float max_pitch_gain = q->erasure_count < 3 ? 0.9 - 0.3 * (q->erasure_count - 1) : 0.0;
 538             for(i=0; i<4; i++)
 539                 q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain);
 540
 541             memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac));
 542         }
 543
 544         // pitch synthesis filter
 545         v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem,
 546                                               cdn_vector, q->pitch_gain,
 547                                               q->pitch_lag, q->frame.pfrac);
 548
 549         // pitch prefilter update
 550         for(i=0; i<4; i++)
 551             q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0);
 552
 553         v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem,
 554                                         v_synthesis_filtered,
 555                                         q->pitch_gain, q->pitch_lag,
 556                                         q->frame.pfrac);
 557
 558         apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered);
 559     }else
 560     {
 561         memcpy(q->pitch_synthesis_filter_mem, cdn_vector + 17,
 562                143 * sizeof(float));
 563         memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float));
 564         memset(q->pitch_gain, 0, sizeof(q->pitch_gain));
 565         memset(q->pitch_lag,  0, sizeof(q->pitch_lag));
 566     }
 567 }
 568
 569 /**
 570  * Interpolates LSP frequencies and computes LPC coefficients
 571  * for a given bitrate & pitch subframe.
 572  *
 573  * TIA/EIA/IS-733 2.4.3.3.4
 574  *
 575  * @param q the context
 576  * @param curr_lspf LSP frequencies vector of the current frame
 577  * @param lpc float vector for the resulting LPC
 578  * @param subframe_num frame number in decoded stream
 579  */
 580 void interpolate_lpc(QCELPContext *q, const float *curr_lspf, float *lpc,
 581                      const int subframe_num)
 582 {
 583     float interpolated_lspf[10];
 584     float weight;
 585
 586     if(q->bitrate >= RATE_QUARTER)
 587         weight = 0.25 * (subframe_num + 1);
 588     else if(q->bitrate == RATE_OCTAVE && !subframe_num)
 589         weight = 0.625;
 590     else
 591         weight = 1.0;
 592
 593     if(weight != 1.0)
 594     {
 595         weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
 596                              weight, 1.0 - weight, 10);
 597         ff_qcelp_lspf2lpc(interpolated_lspf, lpc);
 598     }else if(q->bitrate >= RATE_QUARTER ||
 599              (q->bitrate == I_F_Q && !subframe_num))
 600         ff_qcelp_lspf2lpc(curr_lspf, lpc);
 601 }
 602
 603 static qcelp_packet_rate buf_size2bitrate(const int buf_size)
 604 {
 605     switch(buf_size)
 606     {
 607         case 35: return RATE_FULL;
 608         case 17: return RATE_HALF;
 609         case  8: return RATE_QUARTER;
 610         case  4: return RATE_OCTAVE;
 611         case  1: return SILENCE;
 612     }
 613
 614     return I_F_Q;
 615 }
 616
 617 /**
 618  * Determine the bitrate from the frame size and/or the first byte of the frame.
 619  *
 620  * @param avctx the AV codec context
 621  * @param buf_size length of the buffer
 622  * @param buf the bufffer
 623  *
 624  * @return the bitrate on success,
 625  *         I_F_Q  if the bitrate cannot be satisfactorily determined
 626  *
 627  * TIA/EIA/IS-733 2.4.8.7.1
 628  */
 629 static int determine_bitrate(AVCodecContext *avctx, const int buf_size,
 630                              const uint8_t **buf)
 631 {
 632     qcelp_packet_rate bitrate;
 633
 634     if((bitrate = buf_size2bitrate(buf_size)) >= 0)
 635     {
 636         if(bitrate > **buf)
 637         {
 638             av_log(avctx, AV_LOG_WARNING,
 639                    "Claimed bitrate and buffer size mismatch.\n");
 640             bitrate = **buf;
 641         }else if(bitrate < **buf)
 642         {
 643             av_log(avctx, AV_LOG_ERROR,
 644                    "Buffer is too small for the claimed bitrate.\n");
 645             return I_F_Q;
 646         }
 647         (*buf)++;
 648     }else if((bitrate = buf_size2bitrate(buf_size + 1)) >= 0)
 649     {
 650         av_log(avctx, AV_LOG_WARNING,
 651                "Bitrate byte is missing, guessing the bitrate from packet size.\n");
 652     }else
 653         return I_F_Q;
 654
 655     if(bitrate == SILENCE)
 656     {
 657         // FIXME: the decoder should not handle SILENCE frames as I_F_Q frames
 658         av_log_missing_feature(avctx, "Blank frame", 1);
 659         bitrate = I_F_Q;
 660     }
 661     return bitrate;
 662 }
 663
 664 static void warn_insufficient_frame_quality(AVCodecContext *avctx,
 665                                             const char *message)
 666 {
 667     av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n", avctx->frame_number,
 668            message);
 669 }
 670
 671 static int qcelp_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
 672                               const uint8_t *buf, int buf_size)
 673 {
 674     QCELPContext *q = avctx->priv_data;
 675     float *outbuffer = data;
 676     int   i;
 677     float quantized_lspf[10], lpc[10];
 678     float gain[16];
 679     float *formant_mem;
 680
 681     if((q->bitrate = determine_bitrate(avctx, buf_size, &buf)) == I_F_Q)
 682     {
 683         warn_insufficient_frame_quality(avctx, "bitrate cannot be determined.");
 684         goto erasure;
 685     }
 686
 687     if(q->bitrate == RATE_OCTAVE &&
 688        (q->first16bits = AV_RB16(buf)) == 0xFFFF)
 689     {
 690         warn_insufficient_frame_quality(avctx, "Bitrate is 1/8 and first 16 bits are on.");
 691         goto erasure;
 692     }
 693
 694     if(q->bitrate > SILENCE)
 695     {
 696         const QCELPBitmap *bitmaps     = qcelp_unpacking_bitmaps_per_rate[q->bitrate];
 697         const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate]
 698                                        + qcelp_unpacking_bitmaps_lengths[q->bitrate];
 699         uint8_t           *unpacked_data = (uint8_t *)&q->frame;
 700
 701         init_get_bits(&q->gb, buf, 8*buf_size);
 702
 703         memset(&q->frame, 0, sizeof(QCELPFrame));
 704
 705         for(; bitmaps < bitmaps_end; bitmaps++)
 706             unpacked_data[bitmaps->index] |= get_bits(&q->gb, bitmaps->bitlen) << bitmaps->bitpos;
 707
 708         // Check for erasures/blanks on rates 1, 1/4 and 1/8.
 709         if(q->frame.reserved)
 710         {
 711             warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area.");
 712             goto erasure;
 713         }
 714         if(q->bitrate == RATE_QUARTER &&
 715            codebook_sanity_check_for_rate_quarter(q->frame.cbgain))
 716         {
 717             warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed.");
 718             goto erasure;
 719         }
 720
 721         if(q->bitrate >= RATE_HALF)
 722         {
 723             for(i=0; i<4; i++)
 724             {
 725                 if(q->frame.pfrac[i] && q->frame.plag[i] >= 124)
 726                 {
 727                     warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter.");
 728                     goto erasure;
 729                 }
 730             }
 731         }
 732     }
 733
 734     decode_gain_and_index(q, gain);
 735     compute_svector(q, gain, outbuffer);
 736
 737     if(decode_lspf(q, quantized_lspf) < 0)
 738     {
 739         warn_insufficient_frame_quality(avctx, "Badly received packets in frame.");
 740         goto erasure;
 741     }
 742
 743
 744     apply_pitch_filters(q, outbuffer);
 745
 746     if(q->bitrate == I_F_Q)
 747     {
 748 erasure:
 749         q->bitrate = I_F_Q;
 750         q->erasure_count++;
 751         decode_gain_and_index(q, gain);
 752         compute_svector(q, gain, outbuffer);
 753         decode_lspf(q, quantized_lspf);
 754         apply_pitch_filters(q, outbuffer);
 755     }else
 756         q->erasure_count = 0;
 757
 758     formant_mem = q->formant_mem + 10;
 759     for(i=0; i<4; i++)
 760     {
 761         interpolate_lpc(q, quantized_lspf, lpc, i);
 762         ff_celp_lp_synthesis_filterf(formant_mem, lpc, outbuffer + i * 40, 40,
 763                                      10);
 764         formant_mem += 40;
 765     }
 766     memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float));
 767
 768     // FIXME: postfilter and final gain control should be here.
 769     // TIA/EIA/IS-733 2.4.8.6
 770
 771     formant_mem = q->formant_mem + 10;
 772     for(i=0; i<160; i++)
 773         *outbuffer++ = av_clipf(*formant_mem++, QCELP_CLIP_LOWER_BOUND,
 774                                 QCELP_CLIP_UPPER_BOUND);
 775
 776     memcpy(q->prev_lspf, quantized_lspf, sizeof(q->prev_lspf));
 777     q->prev_bitrate = q->bitrate;
 778
 779     *data_size = 160 * sizeof(*outbuffer);
 780
 781     return *data_size;
 782 }
 783
 784 AVCodec qcelp_decoder =
 785 {
 786     .name   = "qcelp",
 787     .type   = CODEC_TYPE_AUDIO,
 788     .id     = CODEC_ID_QCELP,
 789     .init   = qcelp_decode_init,
 790     .decode = qcelp_decode_frame,
 791     .priv_data_size = sizeof(QCELPContext),
 792     .long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"),
 793 };