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 "internal.h"
  34 #include "bitstream.h"
  35
  36 #include "qcelpdata.h"
  37
  38 #include "celp_math.h"
  39 #include "celp_filters.h"
  40
  41 #undef NDEBUG
  42 #include <assert.h>
  43
  44 typedef enum
  45 {
  46     I_F_Q = -1,    /*!< insufficient frame quality */
  47     SILENCE,
  48     RATE_OCTAVE,
  49     RATE_QUARTER,
  50     RATE_HALF,
  51     RATE_FULL
  52 } qcelp_packet_rate;
  53
  54 typedef struct
  55 {
  56     GetBitContext     gb;
  57     qcelp_packet_rate bitrate;
  58     QCELPFrame        frame;    /*!< unpacked data frame */
  59
  60     uint8_t  erasure_count;
  61     uint8_t  octave_count;      /*!< count the consecutive RATE_OCTAVE frames */
  62     float    prev_lspf[10];
  63     float    predictor_lspf[10];/*!< LSP predictor for RATE_OCTAVE and I_F_Q */
  64     float    pitch_synthesis_filter_mem[303];
  65     float    pitch_pre_filter_mem[303];
  66     float    rnd_fir_filter_mem[180];
  67     float    formant_mem[170];
  68     float    last_codebook_gain;
  69     int      prev_g1[2];
  70     int      prev_bitrate;
  71     float    pitch_gain[4];
  72     uint8_t  pitch_lag[4];
  73     uint16_t first16bits;
  74     uint8_t  warned_buf_mismatch_bitrate;
  75 } QCELPContext;
  76
  77 /**
  78  * Reconstructs LPC coefficients from the line spectral pair frequencies.
  79  *
  80  * TIA/EIA/IS-733 2.4.3.3.5
  81  */
  82 void ff_qcelp_lspf2lpc(const float *lspf, float *lpc);
  83
  84 static void weighted_vector_sumf(float *out, const float *in_a,
  85                                  const float *in_b, float weight_coeff_a,
  86                                  float weight_coeff_b, int length)
  87 {
  88     int i;
  89
  90     for(i=0; i<length; i++)
  91         out[i] = weight_coeff_a * in_a[i]
  92                + weight_coeff_b * in_b[i];
  93 }
  94
  95 /**
  96  * Initialize the speech codec according to the specification.
  97  *
  98  * TIA/EIA/IS-733 2.4.9
  99  */
 100 static av_cold int qcelp_decode_init(AVCodecContext *avctx)
 101 {
 102     QCELPContext *q = avctx->priv_data;
 103     int i;
 104
 105     avctx->sample_fmt = SAMPLE_FMT_FLT;
 106
 107     for(i=0; i<10; i++)
 108         q->prev_lspf[i] = (i+1)/11.;
 109
 110     return 0;
 111 }
 112
 113 /**
 114  * Decodes the 10 quantized LSP frequencies from the LSPV/LSP
 115  * transmission codes of any bitrate and checks for badly received packets.
 116  *
 117  * @param q the context
 118  * @param lspf line spectral pair frequencies
 119  *
 120  * @return 0 on success, -1 if the packet is badly received
 121  *
 122  * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3
 123  */
 124 static int decode_lspf(QCELPContext *q, float *lspf)
 125 {
 126     int i;
 127     float tmp_lspf, smooth, erasure_coeff;
 128     const float *predictors;
 129
 130     if(q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q)
 131     {
 132         predictors = (q->prev_bitrate != RATE_OCTAVE &&
 133                        q->prev_bitrate != I_F_Q ?
 134                        q->prev_lspf : q->predictor_lspf);
 135
 136         if(q->bitrate == RATE_OCTAVE)
 137         {
 138             q->octave_count++;
 139
 140             for(i=0; i<10; i++)
 141             {
 142                 q->predictor_lspf[i] =
 143                              lspf[i] = (q->frame.lspv[i] ?  QCELP_LSP_SPREAD_FACTOR
 144                                                          : -QCELP_LSP_SPREAD_FACTOR)
 145                                      + predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR
 146                                      + (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR)/11);
 147             }
 148             smooth = (q->octave_count < 10 ? .875 : 0.1);
 149         }else
 150         {
 151             erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR;
 152
 153             assert(q->bitrate == I_F_Q);
 154
 155             if(q->erasure_count > 1)
 156                 erasure_coeff *= (q->erasure_count < 4 ? 0.9 : 0.7);
 157
 158             for(i=0; i<10; i++)
 159             {
 160                 q->predictor_lspf[i] =
 161                              lspf[i] = (i + 1) * ( 1 - erasure_coeff)/11
 162                                      + erasure_coeff * predictors[i];
 163             }
 164             smooth = 0.125;
 165         }
 166
 167         // Check the stability of the LSP frequencies.
 168         lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR);
 169         for(i=1; i<10; i++)
 170             lspf[i] = FFMAX(lspf[i], (lspf[i-1] + QCELP_LSP_SPREAD_FACTOR));
 171
 172         lspf[9] = FFMIN(lspf[9], (1.0 - QCELP_LSP_SPREAD_FACTOR));
 173         for(i=9; i>0; i--)
 174             lspf[i-1] = FFMIN(lspf[i-1], (lspf[i] - QCELP_LSP_SPREAD_FACTOR));
 175
 176         // Low-pass filter the LSP frequencies.
 177         weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0-smooth, 10);
 178     }else
 179     {
 180         q->octave_count = 0;
 181
 182         tmp_lspf = 0.;
 183         for(i=0; i<5 ; i++)
 184         {
 185             lspf[2*i+0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001;
 186             lspf[2*i+1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001;
 187         }
 188
 189         // Check for badly received packets.
 190         if(q->bitrate == RATE_QUARTER)
 191         {
 192             if(lspf[9] <= .70 || lspf[9] >=  .97)
 193                 return -1;
 194             for(i=3; i<10; i++)
 195                 if(fabs(lspf[i] - lspf[i-2]) < .08)
 196                     return -1;
 197         }else
 198         {
 199             if(lspf[9] <= .66 || lspf[9] >= .985)
 200                 return -1;
 201             for(i=4; i<10; i++)
 202                 if (fabs(lspf[i] - lspf[i-4]) < .0931)
 203                     return -1;
 204         }
 205     }
 206     return 0;
 207 }
 208
 209 /**
 210  * Converts codebook transmission codes to GAIN and INDEX.
 211  *
 212  * @param q the context
 213  * @param gain array holding the decoded gain
 214  *
 215  * TIA/EIA/IS-733 2.4.6.2
 216  */
 217 static void decode_gain_and_index(QCELPContext  *q,
 218                                   float *gain) {
 219     int   i, subframes_count, g1[16];
 220     float slope;
 221
 222     if(q->bitrate >= RATE_QUARTER)
 223     {
 224         switch(q->bitrate)
 225         {
 226             case RATE_FULL: subframes_count = 16; break;
 227             case RATE_HALF: subframes_count = 4;  break;
 228             default:        subframes_count = 5;
 229         }
 230         for(i=0; i<subframes_count; i++)
 231         {
 232             g1[i] = 4 * q->frame.cbgain[i];
 233             if(q->bitrate == RATE_FULL && !((i+1) & 3))
 234             {
 235                 g1[i] += av_clip((g1[i-1] + g1[i-2] + g1[i-3]) / 3 - 6, 0, 32);
 236             }
 237
 238             gain[i] = qcelp_g12ga[g1[i]];
 239
 240             if(q->frame.cbsign[i])
 241             {
 242                 gain[i] = -gain[i];
 243                 q->frame.cindex[i] = (q->frame.cindex[i]-89) & 127;
 244             }
 245         }
 246
 247         q->prev_g1[0] = g1[i-2];
 248         q->prev_g1[1] = g1[i-1];
 249         q->last_codebook_gain = qcelp_g12ga[g1[i-1]];
 250
 251         if(q->bitrate == RATE_QUARTER)
 252         {
 253             // Provide smoothing of the unvoiced excitation energy.
 254             gain[7] =     gain[4];
 255             gain[6] = 0.4*gain[3] + 0.6*gain[4];
 256             gain[5] =     gain[3];
 257             gain[4] = 0.8*gain[2] + 0.2*gain[3];
 258             gain[3] = 0.2*gain[1] + 0.8*gain[2];
 259             gain[2] =     gain[1];
 260             gain[1] = 0.6*gain[0] + 0.4*gain[1];
 261         }
 262     }else
 263     {
 264         if(q->bitrate == RATE_OCTAVE)
 265         {
 266             g1[0] = 2 * q->frame.cbgain[0]
 267                   + av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54);
 268             subframes_count = 8;
 269         }else
 270         {
 271             assert(q->bitrate == I_F_Q);
 272
 273             g1[0] = q->prev_g1[1];
 274             switch(q->erasure_count)
 275             {
 276                 case 1 : break;
 277                 case 2 : g1[0] -= 1; break;
 278                 case 3 : g1[0] -= 2; break;
 279                 default: g1[0] -= 6;
 280             }
 281             if(g1[0] < 0)
 282                 g1[0] = 0;
 283             subframes_count = 4;
 284         }
 285         // This interpolation is done to produce smoother background noise.
 286         slope = 0.5*(qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count;
 287         for(i=1; i<=subframes_count; i++)
 288             gain[i-1] = q->last_codebook_gain + slope * i;
 289
 290         q->last_codebook_gain = gain[i-2];
 291         q->prev_g1[0] = q->prev_g1[1];
 292         q->prev_g1[1] = g1[0];
 293     }
 294 }
 295
 296 /**
 297  * If the received packet is Rate 1/4 a further sanity check is made of the
 298  * codebook gain.
 299  *
 300  * @param cbgain the unpacked cbgain array
 301  * @return -1 if the sanity check fails, 0 otherwise
 302  *
 303  * TIA/EIA/IS-733 2.4.8.7.3
 304  */
 305 static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)
 306 {
 307     int i, diff, prev_diff=0;
 308
 309     for(i=1; i<5; i++)
 310     {
 311         diff = cbgain[i] - cbgain[i-1];
 312         if(FFABS(diff) > 10)
 313             return -1;
 314         else if(FFABS(diff - prev_diff) > 12)
 315             return -1;
 316         prev_diff = diff;
 317     }
 318     return 0;
 319 }
 320
 321 /**
 322  * Computes the scaled codebook vector Cdn From INDEX and GAIN
 323  * for all rates.
 324  *
 325  * The specification lacks some information here.
 326  *
 327  * TIA/EIA/IS-733 has an omission on the codebook index determination
 328  * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says
 329  * you have to subtract the decoded index parameter from the given scaled
 330  * codebook vector index 'n' to get the desired circular codebook index, but
 331  * it does not mention that you have to clamp 'n' to [0-9] in order to get
 332  * RI-compliant results.
 333  *
 334  * The reason for this mistake seems to be the fact they forgot to mention you
 335  * have to do these calculations per codebook subframe and adjust given
 336  * equation values accordingly.
 337  *
 338  * @param q the context
 339  * @param gain array holding the 4 pitch subframe gain values
 340  * @param cdn_vector array for the generated scaled codebook vector
 341  */
 342 static void compute_svector(QCELPContext *q, const float *gain,
 343                             float *cdn_vector)
 344 {
 345     int      i, j, k;
 346     uint16_t cbseed, cindex;
 347     float    *rnd, tmp_gain, fir_filter_value;
 348
 349     switch(q->bitrate)
 350     {
 351         case RATE_FULL:
 352             for(i=0; i<16; i++)
 353             {
 354                 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
 355                 cindex = -q->frame.cindex[i];
 356                 for(j=0; j<10; j++)
 357                     *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cindex++ & 127];
 358             }
 359         break;
 360         case RATE_HALF:
 361             for(i=0; i<4; i++)
 362             {
 363                 tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO;
 364                 cindex = -q->frame.cindex[i];
 365                 for (j = 0; j < 40; j++)
 366                 *cdn_vector++ = tmp_gain * qcelp_rate_half_codebook[cindex++ & 127];
 367             }
 368         break;
 369         case RATE_QUARTER:
 370             cbseed = (0x0003 & q->frame.lspv[4])<<14 |
 371                      (0x003F & q->frame.lspv[3])<< 8 |
 372                      (0x0060 & q->frame.lspv[2])<< 1 |
 373                      (0x0007 & q->frame.lspv[1])<< 3 |
 374                      (0x0038 & q->frame.lspv[0])>> 3 ;
 375             rnd = q->rnd_fir_filter_mem + 20;
 376             for(i=0; i<8; i++)
 377             {
 378                 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
 379                 for(k=0; k<20; k++)
 380                 {
 381                     cbseed = 521 * cbseed + 259;
 382                     *rnd = (int16_t)cbseed;
 383
 384                     // FIR filter
 385                     fir_filter_value = 0.0;
 386                     for(j=0; j<10; j++)
 387                         fir_filter_value += qcelp_rnd_fir_coefs[j ]
 388                                           * (rnd[-j ] + rnd[-20+j]);
 389
 390                     fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10];
 391                     *cdn_vector++ = tmp_gain * fir_filter_value;
 392                     rnd++;
 393                 }
 394             }
 395             memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160, 20 * sizeof(float));
 396         break;
 397         case RATE_OCTAVE:
 398             cbseed = q->first16bits;
 399             for(i=0; i<8; i++)
 400             {
 401                 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
 402                 for(j=0; j<20; j++)
 403                 {
 404                     cbseed = 521 * cbseed + 259;
 405                     *cdn_vector++ = tmp_gain * (int16_t)cbseed;
 406                 }
 407             }
 408         break;
 409         case I_F_Q:
 410             cbseed = -44; // random codebook index
 411             for(i=0; i<4; i++)
 412             {
 413                 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
 414                 for(j=0; j<40; j++)
 415                     *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cbseed++ & 127];
 416             }
 417         break;
 418     }
 419 }
 420
 421 /**
 422  * Apply generic gain control.
 423  *
 424  * @param v_out output vector
 425  * @param v_in gain-controlled vector
 426  * @param v_ref vector to control gain of
 427  *
 428  * FIXME: If v_ref is a zero vector, it energy is zero
 429  *        and the behavior of the gain control is
 430  *        undefined in the specs.
 431  *
 432  * TIA/EIA/IS-733 2.4.8.3-2/3/4/5, 2.4.8.6
 433  */
 434 static void apply_gain_ctrl(float *v_out, const float *v_ref,
 435                             const float *v_in)
 436 {
 437     int   i, j, len;
 438     float scalefactor;
 439
 440     for(i=0, j=0; i<4; i++)
 441     {
 442         scalefactor = ff_dot_productf(v_in + j, v_in + j, 40);
 443         if(scalefactor)
 444             scalefactor = sqrt(ff_dot_productf(v_ref + j, v_ref + j, 40)
 445                         / scalefactor);
 446         else
 447             ff_log_missing_feature(NULL, "Zero energy for gain control", 1);
 448         for(len=j+40; j<len; j++)
 449             v_out[j] = scalefactor * v_in[j];
 450     }
 451 }
 452
 453 /**
 454  * Apply filter in pitch-subframe steps.
 455  *
 456  * @param memory buffer for the previous state of the filter
 457  *        - must be able to contain 303 elements
 458  *        - the 143 first elements are from the previous state
 459  *        - the next 160 are for output
 460  * @param v_in input filter vector
 461  * @param gain per-subframe gain array, each element is between 0.0 and 2.0
 462  * @param lag per-subframe lag array, each element is
 463  *        - between 16 and 143 if its corresponding pfrac is 0,
 464  *        - between 16 and 139 otherwise
 465  * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0
 466  *        otherwise
 467  *
 468  * @return filter output vector
 469  */
 470 static const float *do_pitchfilter(float memory[303], const float v_in[160],
 471                                    const float gain[4], const uint8_t *lag,
 472                                    const uint8_t pfrac[4])
 473 {
 474     int         i, j;
 475     float       *v_lag, *v_out;
 476     const float *v_len;
 477
 478     v_out = memory + 143; // Output vector starts at memory[143].
 479
 480     for(i=0; i<4; i++)
 481     {
 482         if(gain[i])
 483         {
 484             v_lag = memory + 143 + 40 * i - lag[i];
 485             for(v_len=v_in+40; v_in<v_len; v_in++)
 486             {
 487                 if(pfrac[i]) // If it is a fractional lag...
 488                 {
 489                     for(j=0, *v_out=0.; j<4; j++)
 490                         *v_out += qcelp_hammsinc_table[j] * (v_lag[j-4] + v_lag[3-j]);
 491                 }else
 492                     *v_out = *v_lag;
 493
 494                 *v_out = *v_in + gain[i] * *v_out;
 495
 496                 v_lag++;
 497                 v_out++;
 498             }
 499         }else
 500         {
 501             memcpy(v_out, v_in, 40 * sizeof(float));
 502             v_in  += 40;
 503             v_out += 40;
 504         }
 505     }
 506
 507     memmove(memory, memory + 160, 143 * sizeof(float));
 508     return memory + 143;
 509 }
 510
 511 /**
 512  * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector.
 513  * TIA/EIA/IS-733 2.4.5.2
 514  *
 515  * @param q the context
 516  * @param cdn_vector the scaled codebook vector
 517  */
 518 static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
 519 {
 520     int         i;
 521     const float *v_synthesis_filtered, *v_pre_filtered;
 522
 523     if(q->bitrate >= RATE_HALF ||
 524        (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF)))
 525     {
 526
 527         if(q->bitrate >= RATE_HALF)
 528         {
 529
 530             // Compute gain & lag for the whole frame.
 531             for(i=0; i<4; i++)
 532             {
 533                 q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0;
 534
 535                 q->pitch_lag[i] = q->frame.plag[i] + 16;
 536             }
 537         }else
 538         {
 539             float max_pitch_gain = q->erasure_count < 3 ? 0.9 - 0.3 * (q->erasure_count - 1) : 0.0;
 540             for(i=0; i<4; i++)
 541                 q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain);
 542
 543             memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac));
 544         }
 545
 546         // pitch synthesis filter
 547         v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem,
 548                                               cdn_vector, q->pitch_gain,
 549                                               q->pitch_lag, q->frame.pfrac);
 550
 551         // pitch prefilter update
 552         for(i=0; i<4; i++)
 553             q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0);
 554
 555         v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem,
 556                                         v_synthesis_filtered,
 557                                         q->pitch_gain, q->pitch_lag,
 558                                         q->frame.pfrac);
 559
 560         apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered);
 561     }else
 562     {
 563         memcpy(q->pitch_synthesis_filter_mem, cdn_vector + 17,
 564                143 * sizeof(float));
 565         memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float));
 566         memset(q->pitch_gain, 0, sizeof(q->pitch_gain));
 567         memset(q->pitch_lag,  0, sizeof(q->pitch_lag));
 568     }
 569 }
 570
 571 /**
 572  * Interpolates LSP frequencies and computes LPC coefficients
 573  * for a given bitrate & pitch subframe.
 574  *
 575  * TIA/EIA/IS-733 2.4.3.3.4
 576  *
 577  * @param q the context
 578  * @param curr_lspf LSP frequencies vector of the current frame
 579  * @param lpc float vector for the resulting LPC
 580  * @param subframe_num frame number in decoded stream
 581  */
 582 void interpolate_lpc(QCELPContext *q, const float *curr_lspf, float *lpc,
 583                      const int subframe_num)
 584 {
 585     float interpolated_lspf[10];
 586     float weight;
 587
 588     if(q->bitrate >= RATE_QUARTER)
 589         weight = 0.25 * (subframe_num + 1);
 590     else if(q->bitrate == RATE_OCTAVE && !subframe_num)
 591         weight = 0.625;
 592     else
 593         weight = 1.0;
 594
 595     if(weight != 1.0)
 596     {
 597         weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
 598                              weight, 1.0 - weight, 10);
 599         ff_qcelp_lspf2lpc(interpolated_lspf, lpc);
 600     }else if(q->bitrate >= RATE_QUARTER ||
 601              (q->bitrate == I_F_Q && !subframe_num))
 602         ff_qcelp_lspf2lpc(curr_lspf, lpc);
 603 }
 604
 605 static qcelp_packet_rate buf_size2bitrate(const int buf_size)
 606 {
 607     switch(buf_size)
 608     {
 609         case 35: return RATE_FULL;
 610         case 17: return RATE_HALF;
 611         case  8: return RATE_QUARTER;
 612         case  4: return RATE_OCTAVE;
 613         case  1: return SILENCE;
 614     }
 615
 616     return I_F_Q;
 617 }
 618
 619 /**
 620  * Determine the bitrate from the frame size and/or the first byte of the frame.
 621  *
 622  * @param avctx the AV codec context
 623  * @param buf_size length of the buffer
 624  * @param buf the bufffer
 625  *
 626  * @return the bitrate on success,
 627  *         I_F_Q  if the bitrate cannot be satisfactorily determined
 628  *
 629  * TIA/EIA/IS-733 2.4.8.7.1
 630  */
 631 static int determine_bitrate(AVCodecContext *avctx, const int buf_size,
 632                              const uint8_t **buf)
 633 {
 634     qcelp_packet_rate bitrate;
 635
 636     if((bitrate = buf_size2bitrate(buf_size)) >= 0)
 637     {
 638         if(bitrate > **buf)
 639         {
 640             QCELPContext *q = avctx->priv_data;
 641             if (!q->warned_buf_mismatch_bitrate)
 642             {
 643             av_log(avctx, AV_LOG_WARNING,
 644                    "Claimed bitrate and buffer size mismatch.\n");
 645                 q->warned_buf_mismatch_bitrate = 1;
 646             }
 647             bitrate = **buf;
 648         }else if(bitrate < **buf)
 649         {
 650             av_log(avctx, AV_LOG_ERROR,
 651                    "Buffer is too small for the claimed bitrate.\n");
 652             return I_F_Q;
 653         }
 654         (*buf)++;
 655     }else if((bitrate = buf_size2bitrate(buf_size + 1)) >= 0)
 656     {
 657         av_log(avctx, AV_LOG_WARNING,
 658                "Bitrate byte is missing, guessing the bitrate from packet size.\n");
 659     }else
 660         return I_F_Q;
 661
 662     if(bitrate == SILENCE)
 663     {
 664         // FIXME: the decoder should not handle SILENCE frames as I_F_Q frames
 665         ff_log_missing_feature(avctx, "Blank frame", 1);
 666         bitrate = I_F_Q;
 667     }
 668     return bitrate;
 669 }
 670
 671 static void warn_insufficient_frame_quality(AVCodecContext *avctx,
 672                                             const char *message)
 673 {
 674     av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n", avctx->frame_number,
 675            message);
 676 }
 677
 678 static int qcelp_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
 679                               const uint8_t *buf, int buf_size)
 680 {
 681     QCELPContext *q = avctx->priv_data;
 682     float *outbuffer = data;
 683     int   i;
 684     float quantized_lspf[10], lpc[10];
 685     float gain[16];
 686     float *formant_mem;
 687
 688     if((q->bitrate = determine_bitrate(avctx, buf_size, &buf)) == I_F_Q)
 689     {
 690         warn_insufficient_frame_quality(avctx, "bitrate cannot be determined.");
 691         goto erasure;
 692     }
 693
 694     if(q->bitrate == RATE_OCTAVE &&
 695        (q->first16bits = AV_RB16(buf)) == 0xFFFF)
 696     {
 697         warn_insufficient_frame_quality(avctx, "Bitrate is 1/8 and first 16 bits are on.");
 698         goto erasure;
 699     }
 700
 701     if(q->bitrate > SILENCE)
 702     {
 703         const QCELPBitmap *bitmaps     = qcelp_unpacking_bitmaps_per_rate[q->bitrate];
 704         const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate]
 705                                        + qcelp_unpacking_bitmaps_lengths[q->bitrate];
 706         uint8_t           *unpacked_data = (uint8_t *)&q->frame;
 707
 708         init_get_bits(&q->gb, buf, 8*buf_size);
 709
 710         memset(&q->frame, 0, sizeof(QCELPFrame));
 711
 712         for(; bitmaps < bitmaps_end; bitmaps++)
 713             unpacked_data[bitmaps->index] |= get_bits(&q->gb, bitmaps->bitlen) << bitmaps->bitpos;
 714
 715         // Check for erasures/blanks on rates 1, 1/4 and 1/8.
 716         if(q->frame.reserved)
 717         {
 718             warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area.");
 719             goto erasure;
 720         }
 721         if(q->bitrate == RATE_QUARTER &&
 722            codebook_sanity_check_for_rate_quarter(q->frame.cbgain))
 723         {
 724             warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed.");
 725             goto erasure;
 726         }
 727
 728         if(q->bitrate >= RATE_HALF)
 729         {
 730             for(i=0; i<4; i++)
 731             {
 732                 if(q->frame.pfrac[i] && q->frame.plag[i] >= 124)
 733                 {
 734                     warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter.");
 735                     goto erasure;
 736                 }
 737             }
 738         }
 739     }
 740
 741     decode_gain_and_index(q, gain);
 742     compute_svector(q, gain, outbuffer);
 743
 744     if(decode_lspf(q, quantized_lspf) < 0)
 745     {
 746         warn_insufficient_frame_quality(avctx, "Badly received packets in frame.");
 747         goto erasure;
 748     }
 749
 750
 751     apply_pitch_filters(q, outbuffer);
 752
 753     if(q->bitrate == I_F_Q)
 754     {
 755 erasure:
 756         q->bitrate = I_F_Q;
 757         q->erasure_count++;
 758         decode_gain_and_index(q, gain);
 759         compute_svector(q, gain, outbuffer);
 760         decode_lspf(q, quantized_lspf);
 761         apply_pitch_filters(q, outbuffer);
 762     }else
 763         q->erasure_count = 0;
 764
 765     formant_mem = q->formant_mem + 10;
 766     for(i=0; i<4; i++)
 767     {
 768         interpolate_lpc(q, quantized_lspf, lpc, i);
 769         ff_celp_lp_synthesis_filterf(formant_mem, lpc, outbuffer + i * 40, 40,
 770                                      10);
 771         formant_mem += 40;
 772     }
 773     memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float));
 774
 775     // FIXME: postfilter and final gain control should be here.
 776     // TIA/EIA/IS-733 2.4.8.6
 777
 778     formant_mem = q->formant_mem + 10;
 779     for(i=0; i<160; i++)
 780         *outbuffer++ = av_clipf(*formant_mem++, QCELP_CLIP_LOWER_BOUND,
 781                                 QCELP_CLIP_UPPER_BOUND);
 782
 783     memcpy(q->prev_lspf, quantized_lspf, sizeof(q->prev_lspf));
 784     q->prev_bitrate = q->bitrate;
 785
 786     *data_size = 160 * sizeof(*outbuffer);
 787
 788     return *data_size;
 789 }
 790
 791 AVCodec qcelp_decoder =
 792 {
 793     .name   = "qcelp",
 794     .type   = CODEC_TYPE_AUDIO,
 795     .id     = CODEC_ID_QCELP,
 796     .init   = qcelp_decode_init,
 797     .decode = qcelp_decode_frame,
 798     .priv_data_size = sizeof(QCELPContext),
 799     .long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"),
 800 };