shithub: opus

Download patch

ref: 8cbea177ba1a99c36d15adda77f615f8e80fe212
parent: 30e2493efa11aa09fd5faf4df6cd8711b32d187a
author: Jean-Marc Valin <[email protected]>
date: Thu Aug 5 11:22:57 EDT 2010

Removing pitch prediction code

--- a/libcelt/bands.c
+++ b/libcelt/bands.c
@@ -212,125 +212,6 @@
    }
 }
 
-int compute_pitch_gain(const CELTMode *m, const celt_sig *X, const celt_sig *P, int norm_rate, int *gain_id, int _C, celt_word16 *gain_prod, int M)
-{
-   int j, c;
-   celt_word16 g;
-   celt_word16 delta;
-   const int C = CHANNELS(_C);
-   celt_word32 Sxy=0, Sxx=0, Syy=0;
-   int len = M*m->pitchEnd;
-   int N = M*m->shortMdctSize;
-#ifdef FIXED_POINT
-   int shift = 0;
-   celt_word32 maxabs=0;
-
-   for (c=0;c<C;c++)
-   {
-      for (j=0;j<len;j++)
-      {
-         maxabs = MAX32(maxabs, ABS32(X[j+c*N]));
-         maxabs = MAX32(maxabs, ABS32(P[j+c*N]));
-      }
-   }
-   shift = celt_ilog2(maxabs)-12;
-   if (shift<0)
-      shift = 0;
-#endif
-   delta = PDIV32_16(Q15ONE, len);
-   for (c=0;c<C;c++)
-   {
-      celt_word16 gg = Q15ONE;
-      for (j=0;j<len;j++)
-      {
-         celt_word16 Xj, Pj;
-         Xj = EXTRACT16(SHR32(X[j+c*N], shift));
-         Pj = MULT16_16_P15(gg,EXTRACT16(SHR32(P[j+c*N], shift)));
-         Sxy = MAC16_16(Sxy, Xj, Pj);
-         Sxx = MAC16_16(Sxx, Pj, Pj);
-         Syy = MAC16_16(Syy, Xj, Xj);
-         gg = SUB16(gg, delta);
-      }
-   }
-#ifdef FIXED_POINT
-   {
-      celt_word32 num, den;
-      celt_word16 fact;
-      fact = MULT16_16(QCONST16(.04f, 14), norm_rate);
-      if (fact < QCONST16(1.f, 14))
-         fact = QCONST16(1.f, 14);
-      num = Sxy;
-      den = EPSILON+Sxx+MULT16_32_Q15(QCONST16(.03f,15),Syy);
-      shift = celt_zlog2(Sxy)-16;
-      if (shift < 0)
-         shift = 0;
-      if (Sxy < MULT16_32_Q15(fact, MULT16_16(celt_sqrt(EPSILON+Sxx),celt_sqrt(EPSILON+Syy))))
-         g = 0;
-      else
-         g = DIV32(SHL32(SHR32(num,shift),14),ADD32(EPSILON,SHR32(den,shift)));
-
-      /* This MUST round down so that we don't over-estimate the gain */
-      *gain_id = EXTRACT16(SHR32(MULT16_16(20,(g-QCONST16(.5f,14))),14));
-   }
-#else
-   {
-      float fact = .04f*norm_rate;
-      if (fact < 1)
-         fact = 1;
-      g = Sxy/(.1f+Sxx+.03f*Syy);
-      if (Sxy < .5f*fact*celt_sqrt(1+Sxx*Syy))
-         g = 0;
-      /* This MUST round down so that we don't over-estimate the gain */
-      *gain_id = floor(20*(g-.5f));
-   }
-#endif
-   /* This prevents the pitch gain from being above 1.0 for too long by bounding the 
-      maximum error amplification factor to 2.0 */
-   g = ADD16(QCONST16(.5f,14), MULT16_16_16(QCONST16(.05f,14),*gain_id));
-   *gain_prod = MAX16(QCONST32(1.f, 13), MULT16_16_Q14(*gain_prod,g));
-   if (*gain_prod>QCONST32(2.f, 13))
-   {
-      *gain_id=9;
-      *gain_prod = QCONST32(2.f, 13);
-   }
-
-   if (*gain_id < 0)
-   {
-      *gain_id = 0;
-      return 0;
-   } else {
-      if (*gain_id > 15)
-         *gain_id = 15;
-      return 1;
-   }
-}
-
-void apply_pitch(const CELTMode *m, celt_sig *X, const celt_sig *P, int gain_id, int pred, int _C, int M)
-{
-   int j, c, N;
-   celt_word16 gain;
-   celt_word16 delta;
-   const int C = CHANNELS(_C);
-   int len = M*m->pitchEnd;
-
-   N = M*m->shortMdctSize;
-   gain = ADD16(QCONST16(.5f,14), MULT16_16_16(QCONST16(.05f,14),gain_id));
-   delta = PDIV32_16(gain, len);
-   if (pred)
-      gain = -gain;
-   else
-      delta = -delta;
-   for (c=0;c<C;c++)
-   {
-      celt_word16 gg = gain;
-      for (j=0;j<len;j++)
-      {
-         X[j+c*N] += SHL32(MULT16_32_Q15(gg,P[j+c*N]),1);
-         gg = ADD16(gg, delta);
-      }
-   }
-}
-
 static void stereo_band_mix(const CELTMode *m, celt_norm *X, celt_norm *Y, const celt_ener *bank, int stereo_mode, int bandID, int dir, int N)
 {
    int i = bandID;
--- a/libcelt/bands.h
+++ b/libcelt/bands.h
@@ -66,17 +66,6 @@
  */
 void denormalise_bands(const CELTMode *m, const celt_norm * restrict X, celt_sig * restrict freq, const celt_ener *bands, int end, int _C, int M);
 
-/** Compute the pitch predictor gain for each pitch band
- * @param m Mode data 
- * @param X Spectrum to predict
- * @param P Pitch vector (normalised)
- * @param gains Gain computed for each pitch band (returned)
- * @param bank Square root of the energy for each band
- */
-int compute_pitch_gain(const CELTMode *m, const celt_sig *X, const celt_sig *P, int norm_rate, int *gain_id, int _C, celt_word16 *gain_prod, int M);
-
-void apply_pitch(const CELTMode *m, celt_sig *X, const celt_sig *P, int gain_id, int pred, int _C, int M);
-
 int folding_decision(const CELTMode *m, celt_norm *X, celt_word16 *average, int *last_decision, int end, int _C, int M);
 
 #ifdef MEASURE_NORM_MSE
--- a/libcelt/celt.c
+++ b/libcelt/celt.c
@@ -78,9 +78,6 @@
    int overlap;
    int channels;
    
-   int pitch_enabled;       /* Complexity level is allowed to use pitch */
-   int pitch_permitted;     /*  Use of the LTP is permitted by the user */
-   int pitch_available;     /*  Amount of pitch buffer available */
    int force_intra;
    int delayedIntra;
    celt_word16 tonal_average;
@@ -101,8 +98,6 @@
 
    celt_sig *in_mem;
    celt_sig *out_mem;
-   celt_word16 *pitch_buf;
-   celt_sig xmem;
 
    celt_word16 *oldBandE;
 };
@@ -161,9 +156,6 @@
    st->end = st->mode->effEBands;
 
    st->vbr_rate_norm = 0;
-   st->pitch_enabled = 1;
-   st->pitch_permitted = 1;
-   st->pitch_available = 1;
    st->force_intra  = 0;
    st->delayedIntra = 1;
    st->tonal_average = QCONST16(1.f,8);
@@ -171,7 +163,6 @@
 
    st->in_mem = celt_alloc(st->overlap*C*sizeof(celt_sig));
    st->out_mem = celt_alloc((MAX_PERIOD+st->overlap)*C*sizeof(celt_sig));
-   st->pitch_buf = celt_alloc(((MAX_PERIOD>>1)+2)*sizeof(celt_word16));
 
    st->oldBandE = (celt_word16*)celt_alloc(C*mode->nbEBands*sizeof(celt_word16));
 
@@ -215,7 +206,6 @@
    
    celt_free(st->in_mem);
    celt_free(st->out_mem);
-   celt_free(st->pitch_buf);
    celt_free(st->oldBandE);
    
    st->marker = ENCODERFREED;
@@ -428,13 +418,12 @@
       7 /*111 */ | FLAG_INTRA|FLAG_SHORT|FLAG_FOLD
 };
 
-static void encode_flags(ec_enc *enc, int intra_ener, int has_pitch, int shortBlocks, int has_fold)
+static void encode_flags(ec_enc *enc, int intra_ener, int shortBlocks, int has_fold)
 {
    int i;
    int flags=FLAG_NONE;
    int flag_bits;
    flags |= intra_ener   ? FLAG_INTRA : 0;
-   flags |= has_pitch    ? FLAG_PITCH : 0;
    flags |= shortBlocks  ? FLAG_SHORT : 0;
    flags |= has_fold     ? FLAG_FOLD  : 0;
    for (i=0;i<8;i++)
@@ -459,7 +448,7 @@
    }
 }
 
-static void decode_flags(ec_dec *dec, int *intra_ener, int *has_pitch, int *shortBlocks, int *has_fold)
+static void decode_flags(ec_dec *dec, int *intra_ener, int *shortBlocks, int *has_fold)
 {
    int i;
    int flag_bits;
@@ -474,7 +463,6 @@
          break;
    celt_assert(i<8);
    *intra_ener  = (flaglist[i]&FLAG_INTRA) != 0;
-   *has_pitch   = (flaglist[i]&FLAG_PITCH) != 0;
    *shortBlocks = (flaglist[i]&FLAG_SHORT) != 0;
    *has_fold    = (flaglist[i]&FLAG_FOLD ) != 0;
    /*printf ("dec %d: %d %d %d %d\n", flag_bits, *intra_ener, *has_pitch, *shortBlocks, *has_fold);*/
@@ -684,8 +672,6 @@
 {
 #endif
    int i, c, N, NN, N4;
-   int has_pitch;
-   int pitch_index;
    int bits;
    int has_fold=1;
    ec_byte_buffer buf;
@@ -692,7 +678,6 @@
    ec_enc         _enc;
    VARDECL(celt_sig, in);
    VARDECL(celt_sig, freq);
-   VARDECL(celt_sig, pitch_freq);
    VARDECL(celt_norm, X);
    VARDECL(celt_ener, bandE);
    VARDECL(celt_word16, bandLogE);
@@ -711,8 +696,6 @@
    const int C = CHANNELS(st->channels);
    int mdct_weight_shift = 0;
    int mdct_weight_pos=0;
-   int gain_id=0;
-   int norm_rate;
    int LM, M;
    int tf_select;
    celt_int32 vbr_rate=0;
@@ -780,7 +763,7 @@
    transient_shift = 0;
    isTransient = 0;
 
-   resynth = st->pitch_available>0 || optional_resynthesis!=NULL;
+   resynth = optional_resynthesis!=NULL;
 
    if (M > 1 && transient_analysis(in, N+st->overlap, C, &transient_time, &transient_shift, &st->frame_max, st->overlap))
    {
@@ -824,38 +807,8 @@
    /* Compute MDCTs */
    compute_mdcts(st->mode, shortBlocks, in, freq, C, LM);
 
-
-   norm_rate = (nbAvailableBytes-5)*8*(celt_uint32)st->mode->Fs/(C*N)>>10;
-   /* Pitch analysis: we do it early to save on the peak stack space */
-   /* Don't use pitch if there isn't enough data available yet, 
-      or if we're using shortBlocks */
-   has_pitch = st->pitch_enabled && st->pitch_permitted && (N <= 512) 
-            && (st->pitch_available >= MAX_PERIOD) && (!shortBlocks)
-            && norm_rate < 50;
-   if (has_pitch)
-   {
-      VARDECL(celt_word16, x_lp);
-      SAVE_STACK;
-      ALLOC(x_lp, (2*N-2*N4)>>1, celt_word16);
-      pitch_downsample(in, x_lp, 2*N-2*N4, N, C, &st->xmem, &st->pitch_buf[MAX_PERIOD>>1]);
-      pitch_search(st->mode, x_lp, st->pitch_buf, 2*N-2*N4, MAX_PERIOD-(2*N-2*N4), &pitch_index, &st->xmem, M);
-      RESTORE_STACK;
-   }
-
-   /* Deferred allocation after find_spectral_pitch() to reduce 
-      the peak memory usage */
    ALLOC(X, C*N, celt_norm);         /**< Interleaved normalised MDCTs */
 
-   ALLOC(pitch_freq, C*N, celt_sig); /**< Interleaved signal MDCTs */
-   if (has_pitch)
-   {
-      compute_mdcts(st->mode, 0, st->out_mem+pitch_index*C, pitch_freq, C, LM);
-      has_pitch = compute_pitch_gain(st->mode, freq, pitch_freq, norm_rate, &gain_id, C, &st->gain_prod, M);
-   }
-   
-   if (has_pitch)
-      apply_pitch(st->mode, freq, pitch_freq, gain_id, 1, C, M);
-
    compute_band_energies(st->mode, freq, bandE, effEnd, C, M);
    for (c=0;c<C;c++)
    {
@@ -871,7 +824,7 @@
       has_fold = 0;
 
    /* Don't use intra energy when we're operating at low bit-rate */
-   intra_ener = st->force_intra || (!has_pitch && st->delayedIntra && nbAvailableBytes > st->end);
+   intra_ener = st->force_intra || (st->delayedIntra && nbAvailableBytes > st->end);
    if (shortBlocks || intra_decision(bandLogE, st->oldBandE, st->start, effEnd, st->mode->nbEBands, C))
       st->delayedIntra = 1;
    else
@@ -925,12 +878,7 @@
    }
 
 
-   encode_flags(enc, intra_ener, has_pitch, shortBlocks, has_fold);
-   if (has_pitch)
-   {
-      ec_enc_uint(enc, pitch_index, MAX_PERIOD-(2*N-2*N4));
-      ec_enc_uint(enc, gain_id, 16);
-   }
+   encode_flags(enc, intra_ener, shortBlocks, has_fold);
    if (shortBlocks)
    {
       if (transient_shift)
@@ -1070,9 +1018,6 @@
       measure_norm_mse(st->mode, X, X0, bandE, bandE0, M, N, C);
 #endif
 
-      if (st->pitch_available>0 && st->pitch_available<MAX_PERIOD)
-        st->pitch_available+=N;
-
       if (mdct_weight_shift)
       {
          mdct_shape(st->mode, X, 0, mdct_weight_pos+1, N, mdct_weight_shift, effEnd, C, 1, M);
@@ -1083,9 +1028,6 @@
 
       CELT_MOVE(st->out_mem, st->out_mem+C*N, C*(MAX_PERIOD+st->overlap-N));
 
-      if (has_pitch)
-         apply_pitch(st->mode, freq, pitch_freq, gain_id, 0, C, M);
-      
       for (c=0;c<C;c++)
          for (i=0;i<M*st->mode->eBands[st->start];i++)
             freq[c*N+i] = 0;
@@ -1251,14 +1193,6 @@
          int value = va_arg(ap, celt_int32);
          if (value<0 || value>10)
             goto bad_arg;
-         if (value<=2) {
-            st->pitch_enabled = 0; 
-            st->pitch_available = 0;
-         } else {
-              st->pitch_enabled = 1;
-              if (st->pitch_available<1)
-                st->pitch_available = 1;
-         }   
       }
       break;
       case CELT_SET_START_BAND_REQUEST:
@@ -1285,13 +1219,10 @@
          if (value==0)
          {
             st->force_intra   = 1;
-            st->pitch_permitted = 0;
          } else if (value==1) {
             st->force_intra   = 0;
-            st->pitch_permitted = 0;
          } else {
             st->force_intra   = 0;
-            st->pitch_permitted = 1;
          }   
       }
       break;
@@ -1313,8 +1244,6 @@
          const CELTMode *mode = st->mode;
          int C = st->channels;
 
-         if (st->pitch_available > 0) st->pitch_available = 1;
-
          CELT_MEMSET(st->in_mem, 0, st->overlap*C);
          CELT_MEMSET(st->out_mem, 0, (MAX_PERIOD+st->overlap)*C);
 
@@ -1331,9 +1260,7 @@
          st->vbr_drift = 0;
          st->vbr_offset = 0;
          st->vbr_count = 0;
-         st->xmem = 0;
          st->frame_max = 0;
-         CELT_MEMSET(st->pitch_buf, 0, (MAX_PERIOD>>1)+2);
       }
       break;
       default:
@@ -1668,13 +1595,11 @@
 {
 #endif
    int c, i, N, N4;
-   int has_pitch, has_fold;
-   int pitch_index;
+   int has_fold;
    int bits;
    ec_dec _dec;
    ec_byte_buffer buf;
    VARDECL(celt_sig, freq);
-   VARDECL(celt_sig, pitch_freq);
    VARDECL(celt_norm, X);
    VARDECL(celt_ener, bandE);
    VARDECL(int, fine_quant);
@@ -1691,7 +1616,6 @@
    int mdct_weight_shift=0;
    const int C = CHANNELS(st->channels);
    int mdct_weight_pos=0;
-   int gain_id=0;
    int LM, M;
    int nbFilledBytes, nbAvailableBytes;
    int effEnd;
@@ -1752,7 +1676,7 @@
    }
    nbAvailableBytes = len-nbFilledBytes;
 
-   decode_flags(dec, &intra_ener, &has_pitch, &isTransient, &has_fold);
+   decode_flags(dec, &intra_ener, &isTransient, &has_fold);
    if (isTransient)
       shortBlocks = M;
    else
@@ -1779,24 +1703,6 @@
       transient_shift = 0;
    }
    
-   if (has_pitch)
-   {
-      int maxpitch = MAX_PERIOD-(2*N-2*N4);
-      if (maxpitch<0)
-      {
-         celt_notify("detected pitch when not allowed, bit corruption suspected");
-         dec->error |= 1;
-         pitch_index = 0;
-         has_pitch = 0;
-      } else {
-         pitch_index = ec_dec_uint(dec, maxpitch);
-         gain_id = ec_dec_uint(dec, 16);
-      }
-   } else {
-      pitch_index = 0;
-   }
-
-
    ALLOC(fine_quant, st->mode->nbEBands, int);
    /* Get band energies */
    unquant_coarse_energy(st->mode, st->start, st->end, bandE, st->oldBandE, intra_ener, st->mode->prob, dec, C, LM);
@@ -1818,14 +1724,7 @@
    
    unquant_fine_energy(st->mode, st->start, st->end, bandE, st->oldBandE, fine_quant, dec, C);
 
-   ALLOC(pitch_freq, C*N, celt_sig); /**< Interleaved signal MDCTs */
-   if (has_pitch) 
-   {
-      /* Pitch MDCT */
-      compute_mdcts(st->mode, 0, st->out_mem+pitch_index*C, pitch_freq, C, LM);
-   }
-
-   /* Decode fixed codebook and merge with pitch */
+   /* Decode fixed codebook */
    quant_all_bands(0, st->mode, st->start, st->end, X, C==2 ? X+N : NULL, NULL, pulses, shortBlocks, has_fold, tf_res, 1, len*8, dec, LM);
 
    unquant_energy_finalise(st->mode, st->start, st->end, bandE, st->oldBandE, fine_quant, fine_priority, len*8-ec_dec_tell(dec, 0), dec, C);
@@ -1842,9 +1741,6 @@
 
 
    CELT_MOVE(st->decode_mem, st->decode_mem+C*N, C*(DECODE_BUFFER_SIZE+st->overlap-N));
-
-   if (has_pitch)
-      apply_pitch(st->mode, freq, pitch_freq, gain_id, 0, C, M);
 
    for (c=0;c<C;c++)
       for (i=0;i<M*st->mode->eBands[st->start];i++)
--- a/libcelt/dump_modes.c
+++ b/libcelt/dump_modes.c
@@ -151,7 +151,6 @@
       fprintf(file, "%d,\t/* overlap */\n", mode->overlap);
       fprintf(file, "%d,\t/* nbEBands */\n", mode->nbEBands);
       fprintf(file, "%d,\t/* effEBands */\n", mode->effEBands);
-      fprintf(file, "%d,\t/* pitchEnd */\n", mode->pitchEnd);
       fprintf(file, "{");
       for (i=0;i<4;i++)
          fprintf(file, WORD16 ", ", mode->preemph[i]);
--- a/libcelt/modes.c
+++ b/libcelt/modes.c
@@ -388,7 +388,6 @@
    mode->effEBands = mode->nbEBands;
    while (mode->eBands[mode->effEBands] > mode->shortMdctSize)
       mode->effEBands--;
-   mode->pitchEnd = 4000*(celt_int32)mode->shortMdctSize/Fs;
    
    /* Overlap must be divisible by 4 */
    if (mode->nbShortMdcts > 1)
--- a/libcelt/modes.h
+++ b/libcelt/modes.h
@@ -38,7 +38,6 @@
 #include "celt.h"
 #include "arch.h"
 #include "mdct.h"
-#include "pitch.h"
 #include "entenc.h"
 #include "entdec.h"
 
@@ -78,7 +77,6 @@
 
    int          nbEBands;
    int          effEBands;
-   int          pitchEnd;
    celt_word16    preemph[4];
    const celt_int16   *eBands;   /**< Definition for each "pseudo-critical band" */