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