/src/opus/celt/quant_bands.c

Source (jump to first uncovered line)
/* Copyright (c) 2007-2008 CSIRO
   Copyright (c) 2007-2009 Xiph.Org Foundation
   Written by Jean-Marc Valin */
/*
   Redistribution and use in source and binary forms, with or without
   modification, are permitted provided that the following conditions
   are met:

   - Redistributions of source code must retain the above copyright
   notice, this list of conditions and the following disclaimer.

   - Redistributions in binary form must reproduce the above copyright
   notice, this list of conditions and the following disclaimer in the
   documentation and/or other materials provided with the distribution.

   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/

#ifdef HAVE_CONFIG_H
#include "config.h"
#endif

#include "quant_bands.h"
#include "laplace.h"
#include <math.h>
#include "os_support.h"
#include "arch.h"
#include "mathops.h"
#include "stack_alloc.h"
#include "rate.h"

#ifdef FIXED_POINT
/* Mean energy in each band quantized in Q4 */
const signed char eMeans[25] = {
      103,100, 92, 85, 81,
       77, 72, 70, 78, 75,
       73, 71, 78, 74, 69,
       72, 70, 74, 76, 71,
       60, 60, 60, 60, 60
};
#else
/* Mean energy in each band quantized in Q4 and converted back to float */
const opus_val16 eMeans[25] = {
      6.437500f, 6.250000f, 5.750000f, 5.312500f, 5.062500f,
      4.812500f, 4.500000f, 4.375000f, 4.875000f, 4.687500f,
      4.562500f, 4.437500f, 4.875000f, 4.625000f, 4.312500f,
      4.500000f, 4.375000f, 4.625000f, 4.750000f, 4.437500f,
      3.750000f, 3.750000f, 3.750000f, 3.750000f, 3.750000f
};
#endif
/* prediction coefficients: 0.9, 0.8, 0.65, 0.5 */
#ifdef FIXED_POINT
static const opus_val16 pred_coef[4] = {29440, 26112, 21248, 16384};
static const opus_val16 beta_coef[4] = {30147, 22282, 12124, 6554};
static const opus_val16 beta_intra = 4915;
#else
static const opus_val16 pred_coef[4] = {29440/32768., 26112/32768., 21248/32768., 16384/32768.};
static const opus_val16 beta_coef[4] = {30147/32768., 22282/32768., 12124/32768., 6554/32768.};
static const opus_val16 beta_intra = 4915/32768.;
#endif

/*Parameters of the Laplace-like probability models used for the coarse energy.
  There is one pair of parameters for each frame size, prediction type
   (inter/intra), and band number.
  The first number of each pair is the probability of 0, and the second is the
   decay rate, both in Q8 precision.*/
static const unsigned char e_prob_model[4][2][42] = {
   /*120 sample frames.*/
   {
      /*Inter*/
      {
          72, 127,  65, 129,  66, 128,  65, 128,  64, 128,  62, 128,  64, 128,
          64, 128,  92,  78,  92,  79,  92,  78,  90,  79, 116,  41, 115,  40,
         114,  40, 132,  26, 132,  26, 145,  17, 161,  12, 176,  10, 177,  11
      },
      /*Intra*/
      {
          24, 179,  48, 138,  54, 135,  54, 132,  53, 134,  56, 133,  55, 132,
          55, 132,  61, 114,  70,  96,  74,  88,  75,  88,  87,  74,  89,  66,
          91,  67, 100,  59, 108,  50, 120,  40, 122,  37,  97,  43,  78,  50
      }
   },
   /*240 sample frames.*/
   {
      /*Inter*/
      {
          83,  78,  84,  81,  88,  75,  86,  74,  87,  71,  90,  73,  93,  74,
          93,  74, 109,  40, 114,  36, 117,  34, 117,  34, 143,  17, 145,  18,
         146,  19, 162,  12, 165,  10, 178,   7, 189,   6, 190,   8, 177,   9
      },
      /*Intra*/
      {
          23, 178,  54, 115,  63, 102,  66,  98,  69,  99,  74,  89,  71,  91,
          73,  91,  78,  89,  86,  80,  92,  66,  93,  64, 102,  59, 103,  60,
         104,  60, 117,  52, 123,  44, 138,  35, 133,  31,  97,  38,  77,  45
      }
   },
   /*480 sample frames.*/
   {
      /*Inter*/
      {
          61,  90,  93,  60, 105,  42, 107,  41, 110,  45, 116,  38, 113,  38,
         112,  38, 124,  26, 132,  27, 136,  19, 140,  20, 155,  14, 159,  16,
         158,  18, 170,  13, 177,  10, 187,   8, 192,   6, 175,   9, 159,  10
      },
      /*Intra*/
      {
          21, 178,  59, 110,  71,  86,  75,  85,  84,  83,  91,  66,  88,  73,
          87,  72,  92,  75,  98,  72, 105,  58, 107,  54, 115,  52, 114,  55,
         112,  56, 129,  51, 132,  40, 150,  33, 140,  29,  98,  35,  77,  42
      }
   },
   /*960 sample frames.*/
   {
      /*Inter*/
      {
          42, 121,  96,  66, 108,  43, 111,  40, 117,  44, 123,  32, 120,  36,
         119,  33, 127,  33, 134,  34, 139,  21, 147,  23, 152,  20, 158,  25,
         154,  26, 166,  21, 173,  16, 184,  13, 184,  10, 150,  13, 139,  15
      },
      /*Intra*/
      {
          22, 178,  63, 114,  74,  82,  84,  83,  92,  82, 103,  62,  96,  72,
          96,  67, 101,  73, 107,  72, 113,  55, 118,  52, 125,  52, 118,  52,
         117,  55, 135,  49, 137,  39, 157,  32, 145,  29,  97,  33,  77,  40
      }
   }
};

static const unsigned char small_energy_icdf[3]={2,1,0};

static opus_val32 loss_distortion(const celt_glog *eBands, celt_glog *oldEBands, int start, int end, int len, int C)
{
   int c, i;
   opus_val32 dist = 0;
   c=0; do {
      for (i=start;i<end;i++)
      {
         celt_glog d = PSHR32(SUB32(eBands[i+c*len], oldEBands[i+c*len]), DB_SHIFT-7);
         dist = MAC16_16(dist, d,d);
      }
   } while (++c<C);
   return MIN32(200,SHR32(dist,14));
}

static int quant_coarse_energy_impl(const CELTMode *m, int start, int end,
      const celt_glog *eBands, celt_glog *oldEBands,
      opus_int32 budget, opus_int32 tell,
      const unsigned char *prob_model, celt_glog *error, ec_enc *enc,
      int C, int LM, int intra, celt_glog max_decay, int lfe)
{
   int i, c;
   int badness = 0;
   opus_val32 prev[2] = {0,0};
   opus_val16 coef;
   opus_val16 beta;

   if (tell+3 <= budget)
      ec_enc_bit_logp(enc, intra, 3);
   if (intra)
   {
      coef = 0;
      beta = beta_intra;
   } else {
      beta = beta_coef[LM];
      coef = pred_coef[LM];
   }

   /* Encode at a fixed coarse resolution */
   for (i=start;i<end;i++)
   {
      c=0;
      do {
         int bits_left;
         int qi, qi0;
         opus_val32 q;
         celt_glog x;
         opus_val32 f, tmp;
         celt_glog oldE;
         celt_glog decay_bound;
         x = eBands[i+c*m->nbEBands];
         oldE = MAXG(-GCONST(9.f), oldEBands[i+c*m->nbEBands]);
#ifdef FIXED_POINT
         f = x - MULT16_32_Q15(coef,oldE) - prev[c];
         /* Rounding to nearest integer here is really important! */
         qi = (f+QCONST32(.5f,DB_SHIFT))>>DB_SHIFT;
         decay_bound = MAXG(-GCONST(28.f), SUB32((opus_val32)oldEBands[i+c*m->nbEBands],max_decay));
#else
         f = x-coef*oldE-prev[c];
         /* Rounding to nearest integer here is really important! */
         qi = (int)floor(.5f+f);
         decay_bound = MAXG(-GCONST(28.f), oldEBands[i+c*m->nbEBands]) - max_decay;
#endif
         /* Prevent the energy from going down too quickly (e.g. for bands
            that have just one bin) */
         if (qi < 0 && x < decay_bound)
         {
            qi += (int)SHR32(SUB32(decay_bound,x), DB_SHIFT);
            if (qi > 0)
               qi = 0;
         }
         qi0 = qi;
         /* If we don't have enough bits to encode all the energy, just assume
             something safe. */
         tell = ec_tell(enc);
         bits_left = budget-tell-3*C*(end-i);
         if (i!=start && bits_left < 30)
         {
            if (bits_left < 24)
               qi = IMIN(1, qi);
            if (bits_left < 16)
               qi = IMAX(-1, qi);
         }
         if (lfe && i>=2)
            qi = IMIN(qi, 0);
         if (budget-tell >= 15)
         {
            int pi;
            pi = 2*IMIN(i,20);
            ec_laplace_encode(enc, &qi,
                  prob_model[pi]<<7, prob_model[pi+1]<<6);
         }
         else if(budget-tell >= 2)
         {
            qi = IMAX(-1, IMIN(qi, 1));
            ec_enc_icdf(enc, 2*qi^-(qi<0), small_energy_icdf, 2);
         }
         else if(budget-tell >= 1)
         {
            qi = IMIN(0, qi);
            ec_enc_bit_logp(enc, -qi, 1);
         }
         else
            qi = -1;
         error[i+c*m->nbEBands] = f - SHL32(qi,DB_SHIFT);
         badness += abs(qi0-qi);
         q = (opus_val32)SHL32(EXTEND32(qi),DB_SHIFT);

         tmp = MULT16_32_Q15(coef,oldE) + prev[c] + q;
#ifdef FIXED_POINT
         tmp = MAX32(-GCONST(28.f), tmp);
#endif
         oldEBands[i+c*m->nbEBands] = tmp;
         prev[c] = prev[c] + q - MULT16_32_Q15(beta,q);
      } while (++c < C);
   }
   return lfe ? 0 : badness;
}

void quant_coarse_energy(const CELTMode *m, int start, int end, int effEnd,
      const celt_glog *eBands, celt_glog *oldEBands, opus_uint32 budget,
      celt_glog *error, ec_enc *enc, int C, int LM, int nbAvailableBytes,
      int force_intra, opus_val32 *delayedIntra, int two_pass, int loss_rate, int lfe)
{
   int intra;
   celt_glog max_decay;
   VARDECL(celt_glog, oldEBands_intra);
   VARDECL(celt_glog, error_intra);
   ec_enc enc_start_state;
   opus_uint32 tell;
   int badness1=0;
   opus_int32 intra_bias;
   opus_val32 new_distortion;
   SAVE_STACK;

   intra = force_intra || (!two_pass && *delayedIntra>2*C*(end-start) && nbAvailableBytes > (end-start)*C);
   intra_bias = (opus_int32)((budget**delayedIntra*loss_rate)/(C*512));
   new_distortion = loss_distortion(eBands, oldEBands, start, effEnd, m->nbEBands, C);

   tell = ec_tell(enc);
   if (tell+3 > budget)
      two_pass = intra = 0;

   max_decay = GCONST(16.f);
   if (end-start>10)
   {
#ifdef FIXED_POINT
      max_decay = SHL32(MIN32(SHR32(max_decay,DB_SHIFT-3), EXTEND32(nbAvailableBytes)),DB_SHIFT-3);
#else
      max_decay = MIN32(max_decay, .125f*nbAvailableBytes);
#endif
   }
   if (lfe)
      max_decay = GCONST(3.f);
   enc_start_state = *enc;

   ALLOC(oldEBands_intra, C*m->nbEBands, celt_glog);
   ALLOC(error_intra, C*m->nbEBands, celt_glog);
   OPUS_COPY(oldEBands_intra, oldEBands, C*m->nbEBands);

   if (two_pass || intra)
   {
      badness1 = quant_coarse_energy_impl(m, start, end, eBands, oldEBands_intra, budget,
            tell, e_prob_model[LM][1], error_intra, enc, C, LM, 1, max_decay, lfe);
   }

   if (!intra)
   {
      unsigned char *intra_buf;
      ec_enc enc_intra_state;
      opus_int32 tell_intra;
      opus_uint32 nstart_bytes;
      opus_uint32 nintra_bytes;
      opus_uint32 save_bytes;
      int badness2;
      VARDECL(unsigned char, intra_bits);

      tell_intra = ec_tell_frac(enc);

      enc_intra_state = *enc;

      nstart_bytes = ec_range_bytes(&enc_start_state);
      nintra_bytes = ec_range_bytes(&enc_intra_state);
      intra_buf = ec_get_buffer(&enc_intra_state) + nstart_bytes;
      save_bytes = nintra_bytes-nstart_bytes;
      if (save_bytes == 0)
         save_bytes = ALLOC_NONE;
      ALLOC(intra_bits, save_bytes, unsigned char);
      /* Copy bits from intra bit-stream */
      OPUS_COPY(intra_bits, intra_buf, nintra_bytes - nstart_bytes);

      *enc = enc_start_state;

      badness2 = quant_coarse_energy_impl(m, start, end, eBands, oldEBands, budget,
            tell, e_prob_model[LM][intra], error, enc, C, LM, 0, max_decay, lfe);

      if (two_pass && (badness1 < badness2 || (badness1 == badness2 && ((opus_int32)ec_tell_frac(enc))+intra_bias > tell_intra)))
      {
         *enc = enc_intra_state;
         /* Copy intra bits to bit-stream */
         OPUS_COPY(intra_buf, intra_bits, nintra_bytes - nstart_bytes);
         OPUS_COPY(oldEBands, oldEBands_intra, C*m->nbEBands);
         OPUS_COPY(error, error_intra, C*m->nbEBands);
         intra = 1;
      }
   } else {
      OPUS_COPY(oldEBands, oldEBands_intra, C*m->nbEBands);
      OPUS_COPY(error, error_intra, C*m->nbEBands);
   }

   if (intra)
      *delayedIntra = new_distortion;
   else
      *delayedIntra = ADD32(MULT16_32_Q15(MULT16_16_Q15(pred_coef[LM], pred_coef[LM]),*delayedIntra),
            new_distortion);

   RESTORE_STACK;
}

void quant_fine_energy(const CELTMode *m, int start, int end, celt_glog *oldEBands, celt_glog *error, int *fine_quant, ec_enc *enc, int C)
{
   int i, c;

   /* Encode finer resolution */
   for (i=start;i<end;i++)
   {
      opus_int16 frac = 1<<fine_quant[i];
      if (fine_quant[i] <= 0)
         continue;
      c=0;
      do {
         int q2;
         celt_glog offset;
#ifdef FIXED_POINT
         /* Has to be without rounding */
         q2 = (error[i+c*m->nbEBands]+GCONST(.5f))>>(DB_SHIFT-fine_quant[i]);
#else
         q2 = (int)floor((error[i+c*m->nbEBands]+.5f)*frac);
#endif
         if (q2 > frac-1)
            q2 = frac-1;
         if (q2<0)
            q2 = 0;
         ec_enc_bits(enc, q2, fine_quant[i]);
#ifdef FIXED_POINT
         offset = SUB32(VSHR32(2*q2+1, fine_quant[i]-DB_SHIFT+1), GCONST(.5f));
#else
         offset = (q2+.5f)*(1<<(14-fine_quant[i]))*(1.f/16384) - .5f;
#endif
         oldEBands[i+c*m->nbEBands] += offset;
         error[i+c*m->nbEBands] -= offset;
         /*printf ("%f ", error[i] - offset);*/
      } while (++c < C);
   }
}

void quant_energy_finalise(const CELTMode *m, int start, int end, celt_glog *oldEBands, celt_glog *error, int *fine_quant, int *fine_priority, int bits_left, ec_enc *enc, int C)
{
   int i, prio, c;

   /* Use up the remaining bits */
   for (prio=0;prio<2;prio++)
   {
      for (i=start;i<end && bits_left>=C ;i++)
      {
         if (fine_quant[i] >= MAX_FINE_BITS || fine_priority[i]!=prio)
            continue;
         c=0;
         do {
            int q2;
            celt_glog offset;
            q2 = error[i+c*m->nbEBands]<0 ? 0 : 1;
            ec_enc_bits(enc, q2, 1);
#ifdef FIXED_POINT
            offset = SHR32(SHL32(q2,DB_SHIFT)-GCONST(.5f),fine_quant[i]+1);
#else
            offset = (q2-.5f)*(1<<(14-fine_quant[i]-1))*(1.f/16384);
#endif
            oldEBands[i+c*m->nbEBands] += offset;
            error[i+c*m->nbEBands] -= offset;
            bits_left--;
         } while (++c < C);
      }
   }
}

void unquant_coarse_energy(const CELTMode *m, int start, int end, celt_glog *oldEBands, int intra, ec_dec *dec, int C, int LM)
{
   const unsigned char *prob_model = e_prob_model[LM][intra];
   int i, c;
   opus_val64 prev[2] = {0, 0};
   opus_val16 coef;
   opus_val16 beta;
   opus_int32 budget;
   opus_int32 tell;

   if (intra)
   {
      coef = 0;
      beta = beta_intra;
   } else {
      beta = beta_coef[LM];
      coef = pred_coef[LM];
   }

   budget = dec->storage*8;

   /* Decode at a fixed coarse resolution */
   for (i=start;i<end;i++)
   {
      c=0;
      do {
         int qi;
         opus_val32 q;
         opus_val32 tmp;
         /* It would be better to express this invariant as a
            test on C at function entry, but that isn't enough
            to make the static analyzer happy. */
         celt_sig_assert(c<2);
         tell = ec_tell(dec);
         if(budget-tell>=15)
         {
            int pi;
            pi = 2*IMIN(i,20);
            qi = ec_laplace_decode(dec,
                  prob_model[pi]<<7, prob_model[pi+1]<<6);
         }
         else if(budget-tell>=2)
         {
            qi = ec_dec_icdf(dec, small_energy_icdf, 2);
            qi = (qi>>1)^-(qi&1);
         }
         else if(budget-tell>=1)
         {
            qi = -ec_dec_bit_logp(dec, 1);
         }
         else
            qi = -1;
         q = (opus_val32)SHL32(EXTEND32(qi),DB_SHIFT);

         oldEBands[i+c*m->nbEBands] = MAXG(-GCONST(9.f), oldEBands[i+c*m->nbEBands]);
         tmp = MULT16_32_Q15(coef,oldEBands[i+c*m->nbEBands]) + prev[c] + q;
#ifdef FIXED_POINT
         tmp = MIN32(GCONST(28.f), MAX32(-GCONST(28.f), tmp));
#endif
         oldEBands[i+c*m->nbEBands] = tmp;
         prev[c] = prev[c] + q - MULT16_32_Q15(beta,q);
      } while (++c < C);
   }
}

void unquant_fine_energy(const CELTMode *m, int start, int end, celt_glog *oldEBands, int *fine_quant, ec_dec *dec, int C)
{
   int i, c;
   /* Decode finer resolution */
   for (i=start;i<end;i++)
   {
      if (fine_quant[i] <= 0)
         continue;
      c=0;
      do {
         int q2;
         celt_glog offset;
         q2 = ec_dec_bits(dec, fine_quant[i]);
#ifdef FIXED_POINT
         offset = SUB32(VSHR32(2*q2+1, fine_quant[i]-DB_SHIFT+1), GCONST(.5f));
#else
         offset = (q2+.5f)*(1<<(14-fine_quant[i]))*(1.f/16384) - .5f;
#endif
         oldEBands[i+c*m->nbEBands] += offset;
      } while (++c < C);
   }
}

void unquant_energy_finalise(const CELTMode *m, int start, int end, celt_glog *oldEBands, int *fine_quant,  int *fine_priority, int bits_left, ec_dec *dec, int C)
{
   int i, prio, c;

   /* Use up the remaining bits */
   for (prio=0;prio<2;prio++)
   {
      for (i=start;i<end && bits_left>=C ;i++)
      {
         if (fine_quant[i] >= MAX_FINE_BITS || fine_priority[i]!=prio)
            continue;
         c=0;
         do {
            int q2;
            celt_glog offset;
            q2 = ec_dec_bits(dec, 1);
#ifdef FIXED_POINT
            offset = SHR32(SHL32(q2,DB_SHIFT)-GCONST(.5f),fine_quant[i]+1);
#else
            offset = (q2-.5f)*(1<<(14-fine_quant[i]-1))*(1.f/16384);
#endif
            oldEBands[i+c*m->nbEBands] += offset;
            bits_left--;
         } while (++c < C);
      }
   }
}

void amp2Log2(const CELTMode *m, int effEnd, int end,
      celt_ener *bandE, celt_glog *bandLogE, int C)
{
   int c, i;
   c=0;
   do {
      for (i=0;i<effEnd;i++)
      {
         bandLogE[i+c*m->nbEBands] =
               celt_log2_db(bandE[i+c*m->nbEBands])
               - SHL32((celt_glog)eMeans[i],DB_SHIFT-4);
#ifdef FIXED_POINT
         /* Compensate for bandE[] being Q12 but celt_log2() taking a Q14 input. */
         bandLogE[i+c*m->nbEBands] += GCONST(2.f);
#endif
      }
      for (i=effEnd;i<end;i++)
         bandLogE[c*m->nbEBands+i] = -GCONST(14.f);
   } while (++c < C);
}

Coverage Report

Created: 2025-08-28 07:12

Line	Count	Source (jump to first uncovered line)
1		/* Copyright (c) 2007-2008 CSIRO
2		Copyright (c) 2007-2009 Xiph.Org Foundation
3		Written by Jean-Marc Valin */
4		/*
5		Redistribution and use in source and binary forms, with or without
6		modification, are permitted provided that the following conditions
7		are met:
8
9		- Redistributions of source code must retain the above copyright
10		notice, this list of conditions and the following disclaimer.
11
12		- Redistributions in binary form must reproduce the above copyright
13		notice, this list of conditions and the following disclaimer in the
14		documentation and/or other materials provided with the distribution.
15
16		THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
17		``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
18		LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
19		A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
20		OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
21		EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
22		PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
23		PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
24		LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
25		NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
26		SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27		*/
28
29		#ifdef HAVE_CONFIG_H
30		#include "config.h"
31		#endif
32
33		#include "quant_bands.h"
34		#include "laplace.h"
35		#include <math.h>
36		#include "os_support.h"
37		#include "arch.h"
38		#include "mathops.h"
39		#include "stack_alloc.h"
40		#include "rate.h"
41
42		#ifdef FIXED_POINT
43		/* Mean energy in each band quantized in Q4 */
44		const signed char eMeans[25] = {
45		103,100, 92, 85, 81,
46		77, 72, 70, 78, 75,
47		73, 71, 78, 74, 69,
48		72, 70, 74, 76, 71,
49		60, 60, 60, 60, 60
50		};
51		#else
52		/* Mean energy in each band quantized in Q4 and converted back to float */
53		const opus_val16 eMeans[25] = {
54		6.437500f, 6.250000f, 5.750000f, 5.312500f, 5.062500f,
55		4.812500f, 4.500000f, 4.375000f, 4.875000f, 4.687500f,
56		4.562500f, 4.437500f, 4.875000f, 4.625000f, 4.312500f,
57		4.500000f, 4.375000f, 4.625000f, 4.750000f, 4.437500f,
58		3.750000f, 3.750000f, 3.750000f, 3.750000f, 3.750000f
59		};
60		#endif
61		/* prediction coefficients: 0.9, 0.8, 0.65, 0.5 */
62		#ifdef FIXED_POINT
63		static const opus_val16 pred_coef[4] = {29440, 26112, 21248, 16384};
64		static const opus_val16 beta_coef[4] = {30147, 22282, 12124, 6554};
65		static const opus_val16 beta_intra = 4915;
66		#else
67		static const opus_val16 pred_coef[4] = {29440/32768., 26112/32768., 21248/32768., 16384/32768.};
68		static const opus_val16 beta_coef[4] = {30147/32768., 22282/32768., 12124/32768., 6554/32768.};
69		static const opus_val16 beta_intra = 4915/32768.;
70		#endif
71
72		/*Parameters of the Laplace-like probability models used for the coarse energy.
73		There is one pair of parameters for each frame size, prediction type
74		(inter/intra), and band number.
75		The first number of each pair is the probability of 0, and the second is the
76		decay rate, both in Q8 precision.*/
77		static const unsigned char e_prob_model[4][2][42] = {
78		/120 sample frames./
79		{
80		/Inter/
81		{
82		72, 127, 65, 129, 66, 128, 65, 128, 64, 128, 62, 128, 64, 128,
83		64, 128, 92, 78, 92, 79, 92, 78, 90, 79, 116, 41, 115, 40,
84		114, 40, 132, 26, 132, 26, 145, 17, 161, 12, 176, 10, 177, 11
85		},
86		/Intra/
87		{
88		24, 179, 48, 138, 54, 135, 54, 132, 53, 134, 56, 133, 55, 132,
89		55, 132, 61, 114, 70, 96, 74, 88, 75, 88, 87, 74, 89, 66,
90		91, 67, 100, 59, 108, 50, 120, 40, 122, 37, 97, 43, 78, 50
91		}
92		},
93		/240 sample frames./
94		{
95		/Inter/
96		{
97		83, 78, 84, 81, 88, 75, 86, 74, 87, 71, 90, 73, 93, 74,
98		93, 74, 109, 40, 114, 36, 117, 34, 117, 34, 143, 17, 145, 18,
99		146, 19, 162, 12, 165, 10, 178, 7, 189, 6, 190, 8, 177, 9
100		},
101		/Intra/
102		{
103		23, 178, 54, 115, 63, 102, 66, 98, 69, 99, 74, 89, 71, 91,
104		73, 91, 78, 89, 86, 80, 92, 66, 93, 64, 102, 59, 103, 60,
105		104, 60, 117, 52, 123, 44, 138, 35, 133, 31, 97, 38, 77, 45
106		}
107		},
108		/480 sample frames./
109		{
110		/Inter/
111		{
112		61, 90, 93, 60, 105, 42, 107, 41, 110, 45, 116, 38, 113, 38,
113		112, 38, 124, 26, 132, 27, 136, 19, 140, 20, 155, 14, 159, 16,
114		158, 18, 170, 13, 177, 10, 187, 8, 192, 6, 175, 9, 159, 10
115		},
116		/Intra/
117		{
118		21, 178, 59, 110, 71, 86, 75, 85, 84, 83, 91, 66, 88, 73,
119		87, 72, 92, 75, 98, 72, 105, 58, 107, 54, 115, 52, 114, 55,
120		112, 56, 129, 51, 132, 40, 150, 33, 140, 29, 98, 35, 77, 42
121		}
122		},
123		/960 sample frames./
124		{
125		/Inter/
126		{
127		42, 121, 96, 66, 108, 43, 111, 40, 117, 44, 123, 32, 120, 36,
128		119, 33, 127, 33, 134, 34, 139, 21, 147, 23, 152, 20, 158, 25,
129		154, 26, 166, 21, 173, 16, 184, 13, 184, 10, 150, 13, 139, 15
130		},
131		/Intra/
132		{
133		22, 178, 63, 114, 74, 82, 84, 83, 92, 82, 103, 62, 96, 72,
134		96, 67, 101, 73, 107, 72, 113, 55, 118, 52, 125, 52, 118, 52,
135		117, 55, 135, 49, 137, 39, 157, 32, 145, 29, 97, 33, 77, 40
136		}
137		}
138		};
139
140		static const unsigned char small_energy_icdf[3]={2,1,0};
141
142		static opus_val32 loss_distortion(const celt_glog eBands, celt_glog oldEBands, int start, int end, int len, int C)
143	0	{
144	0	int c, i;
145	0	opus_val32 dist = 0;
146	0	c=0; do {
147	0	for (i=start;i<end;i++)
148	0	{
149	0	celt_glog d = PSHR32(SUB32(eBands[i+clen], oldEBands[i+clen]), DB_SHIFT-7);
150	0	dist = MAC16_16(dist, d,d);
151	0	}
152	0	} while (++c<C);
153	0	return MIN32(200,SHR32(dist,14));
154	0	}
155
156		static int quant_coarse_energy_impl(const CELTMode *m, int start, int end,
157		const celt_glog eBands, celt_glog oldEBands,
158		opus_int32 budget, opus_int32 tell,
159		const unsigned char prob_model, celt_glog error, ec_enc *enc,
160		int C, int LM, int intra, celt_glog max_decay, int lfe)
161	0	{
162	0	int i, c;
163	0	int badness = 0;
164	0	opus_val32 prev[2] = {0,0};
165	0	opus_val16 coef;
166	0	opus_val16 beta;
167
168	0	if (tell+3 <= budget)
169	0	ec_enc_bit_logp(enc, intra, 3);
170	0	if (intra)
171	0	{
172	0	coef = 0;
173	0	beta = beta_intra;
174	0	} else {
175	0	beta = beta_coef[LM];
176	0	coef = pred_coef[LM];
177	0	}
178
179		/* Encode at a fixed coarse resolution */
180	0	for (i=start;i<end;i++)
181	0	{
182	0	c=0;
183	0	do {
184	0	int bits_left;
185	0	int qi, qi0;
186	0	opus_val32 q;
187	0	celt_glog x;
188	0	opus_val32 f, tmp;
189	0	celt_glog oldE;
190	0	celt_glog decay_bound;
191	0	x = eBands[i+c*m->nbEBands];
192	0	oldE = MAXG(-GCONST(9.f), oldEBands[i+c*m->nbEBands]);
193		#ifdef FIXED_POINT
194		f = x - MULT16_32_Q15(coef,oldE) - prev[c];
195		/* Rounding to nearest integer here is really important! */
196		qi = (f+QCONST32(.5f,DB_SHIFT))>>DB_SHIFT;
197		decay_bound = MAXG(-GCONST(28.f), SUB32((opus_val32)oldEBands[i+c*m->nbEBands],max_decay));
198		#else
199	0	f = x-coef*oldE-prev[c];
200		/* Rounding to nearest integer here is really important! */
201	0	qi = (int)floor(.5f+f);
202	0	decay_bound = MAXG(-GCONST(28.f), oldEBands[i+c*m->nbEBands]) - max_decay;
203	0	#endif
204		/* Prevent the energy from going down too quickly (e.g. for bands
205		that have just one bin) */
206	0	if (qi < 0 && x < decay_bound)
207	0	{
208	0	qi += (int)SHR32(SUB32(decay_bound,x), DB_SHIFT);
209	0	if (qi > 0)
210	0	qi = 0;
211	0	}
212	0	qi0 = qi;
213		/* If we don't have enough bits to encode all the energy, just assume
214		something safe. */
215	0	tell = ec_tell(enc);
216	0	bits_left = budget-tell-3C(end-i);
217	0	if (i!=start && bits_left < 30)
218	0	{
219	0	if (bits_left < 24)
220	0	qi = IMIN(1, qi);
221	0	if (bits_left < 16)
222	0	qi = IMAX(-1, qi);
223	0	}
224	0	if (lfe && i>=2)
225	0	qi = IMIN(qi, 0);
226	0	if (budget-tell >= 15)
227	0	{
228	0	int pi;
229	0	pi = 2*IMIN(i,20);
230	0	ec_laplace_encode(enc, &qi,
231	0	prob_model[pi]<<7, prob_model[pi+1]<<6);
232	0	}
233	0	else if(budget-tell >= 2)
234	0	{
235	0	qi = IMAX(-1, IMIN(qi, 1));
236	0	ec_enc_icdf(enc, 2*qi^-(qi<0), small_energy_icdf, 2);
237	0	}
238	0	else if(budget-tell >= 1)
239	0	{
240	0	qi = IMIN(0, qi);
241	0	ec_enc_bit_logp(enc, -qi, 1);
242	0	}
243	0	else
244	0	qi = -1;
245	0	error[i+c*m->nbEBands] = f - SHL32(qi,DB_SHIFT);
246	0	badness += abs(qi0-qi);
247	0	q = (opus_val32)SHL32(EXTEND32(qi),DB_SHIFT);
248
249	0	tmp = MULT16_32_Q15(coef,oldE) + prev[c] + q;
250		#ifdef FIXED_POINT
251		tmp = MAX32(-GCONST(28.f), tmp);
252		#endif
253	0	oldEBands[i+c*m->nbEBands] = tmp;
254	0	prev[c] = prev[c] + q - MULT16_32_Q15(beta,q);
255	0	} while (++c < C);
256	0	}
257	0	return lfe ? 0 : badness;
258	0	}
259
260		void quant_coarse_energy(const CELTMode *m, int start, int end, int effEnd,
261		const celt_glog eBands, celt_glog oldEBands, opus_uint32 budget,
262		celt_glog error, ec_enc enc, int C, int LM, int nbAvailableBytes,
263		int force_intra, opus_val32 *delayedIntra, int two_pass, int loss_rate, int lfe)
264	0	{
265	0	int intra;
266	0	celt_glog max_decay;
267	0	VARDECL(celt_glog, oldEBands_intra);
268	0	VARDECL(celt_glog, error_intra);
269	0	ec_enc enc_start_state;
270	0	opus_uint32 tell;
271	0	int badness1=0;
272	0	opus_int32 intra_bias;
273	0	opus_val32 new_distortion;
274	0	SAVE_STACK;
275
276	0	intra = force_intra \|\| (!two_pass && delayedIntra>2C(end-start) && nbAvailableBytes > (end-start)C);
277	0	intra_bias = (opus_int32)((budget*delayedIntraloss_rate)/(C*512));
278	0	new_distortion = loss_distortion(eBands, oldEBands, start, effEnd, m->nbEBands, C);
279
280	0	tell = ec_tell(enc);
281	0	if (tell+3 > budget)
282	0	two_pass = intra = 0;
283
284	0	max_decay = GCONST(16.f);
285	0	if (end-start>10)
286	0	{
287		#ifdef FIXED_POINT
288		max_decay = SHL32(MIN32(SHR32(max_decay,DB_SHIFT-3), EXTEND32(nbAvailableBytes)),DB_SHIFT-3);
289		#else
290	0	max_decay = MIN32(max_decay, .125f*nbAvailableBytes);
291	0	#endif
292	0	}
293	0	if (lfe)
294	0	max_decay = GCONST(3.f);
295	0	enc_start_state = *enc;
296
297	0	ALLOC(oldEBands_intra, C*m->nbEBands, celt_glog);
298	0	ALLOC(error_intra, C*m->nbEBands, celt_glog);
299	0	OPUS_COPY(oldEBands_intra, oldEBands, C*m->nbEBands);
300
301	0	if (two_pass \|\| intra)
302	0	{
303	0	badness1 = quant_coarse_energy_impl(m, start, end, eBands, oldEBands_intra, budget,
304	0	tell, e_prob_model[LM][1], error_intra, enc, C, LM, 1, max_decay, lfe);
305	0	}
306
307	0	if (!intra)
308	0	{
309	0	unsigned char *intra_buf;
310	0	ec_enc enc_intra_state;
311	0	opus_int32 tell_intra;
312	0	opus_uint32 nstart_bytes;
313	0	opus_uint32 nintra_bytes;
314	0	opus_uint32 save_bytes;
315	0	int badness2;
316	0	VARDECL(unsigned char, intra_bits);
317
318	0	tell_intra = ec_tell_frac(enc);
319
320	0	enc_intra_state = *enc;
321
322	0	nstart_bytes = ec_range_bytes(&enc_start_state);
323	0	nintra_bytes = ec_range_bytes(&enc_intra_state);
324	0	intra_buf = ec_get_buffer(&enc_intra_state) + nstart_bytes;
325	0	save_bytes = nintra_bytes-nstart_bytes;
326	0	if (save_bytes == 0)
327	0	save_bytes = ALLOC_NONE;
328	0	ALLOC(intra_bits, save_bytes, unsigned char);
329		/* Copy bits from intra bit-stream */
330	0	OPUS_COPY(intra_bits, intra_buf, nintra_bytes - nstart_bytes);
331
332	0	*enc = enc_start_state;
333
334	0	badness2 = quant_coarse_energy_impl(m, start, end, eBands, oldEBands, budget,
335	0	tell, e_prob_model[LM][intra], error, enc, C, LM, 0, max_decay, lfe);
336
337	0	if (two_pass && (badness1 < badness2 \|\| (badness1 == badness2 && ((opus_int32)ec_tell_frac(enc))+intra_bias > tell_intra)))
338	0	{
339	0	*enc = enc_intra_state;
340		/* Copy intra bits to bit-stream */
341	0	OPUS_COPY(intra_buf, intra_bits, nintra_bytes - nstart_bytes);
342	0	OPUS_COPY(oldEBands, oldEBands_intra, C*m->nbEBands);
343	0	OPUS_COPY(error, error_intra, C*m->nbEBands);
344	0	intra = 1;
345	0	}
346	0	} else {
347	0	OPUS_COPY(oldEBands, oldEBands_intra, C*m->nbEBands);
348	0	OPUS_COPY(error, error_intra, C*m->nbEBands);
349	0	}
350
351	0	if (intra)
352	0	*delayedIntra = new_distortion;
353	0	else
354	0	delayedIntra = ADD32(MULT16_32_Q15(MULT16_16_Q15(pred_coef[LM], pred_coef[LM]),delayedIntra),
355	0	new_distortion);
356
357	0	RESTORE_STACK;
358	0	}
359
360		void quant_fine_energy(const CELTMode m, int start, int end, celt_glog oldEBands, celt_glog error, int fine_quant, ec_enc *enc, int C)
361	0	{
362	0	int i, c;
363
364		/* Encode finer resolution */
365	0	for (i=start;i<end;i++)
366	0	{
367	0	opus_int16 frac = 1<<fine_quant[i];
368	0	if (fine_quant[i] <= 0)
369	0	continue;
370	0	c=0;
371	0	do {
372	0	int q2;
373	0	celt_glog offset;
374		#ifdef FIXED_POINT
375		/* Has to be without rounding */
376		q2 = (error[i+c*m->nbEBands]+GCONST(.5f))>>(DB_SHIFT-fine_quant[i]);
377		#else
378	0	q2 = (int)floor((error[i+cm->nbEBands]+.5f)frac);
379	0	#endif
380	0	if (q2 > frac-1)
381	0	q2 = frac-1;
382	0	if (q2<0)
383	0	q2 = 0;
384	0	ec_enc_bits(enc, q2, fine_quant[i]);
385		#ifdef FIXED_POINT
386		offset = SUB32(VSHR32(2*q2+1, fine_quant[i]-DB_SHIFT+1), GCONST(.5f));
387		#else
388	0	offset = (q2+.5f)(1<<(14-fine_quant[i]))(1.f/16384) - .5f;
389	0	#endif
390	0	oldEBands[i+c*m->nbEBands] += offset;
391	0	error[i+c*m->nbEBands] -= offset;
392		/printf ("%f ", error[i] - offset);/
393	0	} while (++c < C);
394	0	}
395	0	}
396
397		void quant_energy_finalise(const CELTMode m, int start, int end, celt_glog oldEBands, celt_glog error, int fine_quant, int fine_priority, int bits_left, ec_enc enc, int C)
398	0	{
399	0	int i, prio, c;
400
401		/* Use up the remaining bits */
402	0	for (prio=0;prio<2;prio++)
403	0	{
404	0	for (i=start;i<end && bits_left>=C ;i++)
405	0	{
406	0	if (fine_quant[i] >= MAX_FINE_BITS \|\| fine_priority[i]!=prio)
407	0	continue;
408	0	c=0;
409	0	do {
410	0	int q2;
411	0	celt_glog offset;
412	0	q2 = error[i+c*m->nbEBands]<0 ? 0 : 1;
413	0	ec_enc_bits(enc, q2, 1);
414		#ifdef FIXED_POINT
415		offset = SHR32(SHL32(q2,DB_SHIFT)-GCONST(.5f),fine_quant[i]+1);
416		#else
417	0	offset = (q2-.5f)(1<<(14-fine_quant[i]-1))(1.f/16384);
418	0	#endif
419	0	oldEBands[i+c*m->nbEBands] += offset;
420	0	error[i+c*m->nbEBands] -= offset;
421	0	bits_left--;
422	0	} while (++c < C);
423	0	}
424	0	}
425	0	}
426
427		void unquant_coarse_energy(const CELTMode m, int start, int end, celt_glog oldEBands, int intra, ec_dec *dec, int C, int LM)
428	116k	{
429	116k	const unsigned char *prob_model = e_prob_model[LM][intra];
430	116k	int i, c;
431	116k	opus_val64 prev[2] = {0, 0};
432	116k	opus_val16 coef;
433	116k	opus_val16 beta;
434	116k	opus_int32 budget;
435	116k	opus_int32 tell;
436
437	116k	if (intra)
438	6.48k	{
439	6.48k	coef = 0;
440	6.48k	beta = beta_intra;
441	110k	} else {
442	110k	beta = beta_coef[LM];
443	110k	coef = pred_coef[LM];
444	110k	}
445
446	116k	budget = dec->storage*8;
447
448		/* Decode at a fixed coarse resolution */
449	1.55M	for (i=start;i<end;i++)
450	1.44M	{
451	1.44M	c=0;
452	1.77M	do {
453	1.77M	int qi;
454	1.77M	opus_val32 q;
455	1.77M	opus_val32 tmp;
456		/* It would be better to express this invariant as a
457		test on C at function entry, but that isn't enough
458		to make the static analyzer happy. */
459	1.77M	celt_sig_assert(c<2);
460	1.77M	tell = ec_tell(dec);
461	1.77M	if(budget-tell>=15)
462	1.31M	{
463	1.31M	int pi;
464	1.31M	pi = 2*IMIN(i,20);
465	1.31M	qi = ec_laplace_decode(dec,
466	1.31M	prob_model[pi]<<7, prob_model[pi+1]<<6);
467	1.31M	}
468	467k	else if(budget-tell>=2)
469	194k	{
470	194k	qi = ec_dec_icdf(dec, small_energy_icdf, 2);
471	194k	qi = (qi>>1)^-(qi&1);
472	194k	}
473	272k	else if(budget-tell>=1)
474	8.88k	{
475	8.88k	qi = -ec_dec_bit_logp(dec, 1);
476	8.88k	}
477	263k	else
478	263k	qi = -1;
479	1.77M	q = (opus_val32)SHL32(EXTEND32(qi),DB_SHIFT);
480
481	1.77M	oldEBands[i+cm->nbEBands] = MAXG(-GCONST(9.f), oldEBands[i+cm->nbEBands]);
482	1.77M	tmp = MULT16_32_Q15(coef,oldEBands[i+c*m->nbEBands]) + prev[c] + q;
483		#ifdef FIXED_POINT
484		tmp = MIN32(GCONST(28.f), MAX32(-GCONST(28.f), tmp));
485		#endif
486	1.77M	oldEBands[i+c*m->nbEBands] = tmp;
487	1.77M	prev[c] = prev[c] + q - MULT16_32_Q15(beta,q);
488	1.77M	} while (++c < C);
489	1.44M	}
490	116k	}
491
492		void unquant_fine_energy(const CELTMode m, int start, int end, celt_glog oldEBands, int fine_quant, ec_dec dec, int C)
493	116k	{
494	116k	int i, c;
495		/* Decode finer resolution */
496	1.55M	for (i=start;i<end;i++)
497	1.44M	{
498	1.44M	if (fine_quant[i] <= 0)
499	732k	continue;
500	708k	c=0;
501	896k	do {
502	896k	int q2;
503	896k	celt_glog offset;
504	896k	q2 = ec_dec_bits(dec, fine_quant[i]);
505		#ifdef FIXED_POINT
506		offset = SUB32(VSHR32(2*q2+1, fine_quant[i]-DB_SHIFT+1), GCONST(.5f));
507		#else
508	896k	offset = (q2+.5f)(1<<(14-fine_quant[i]))(1.f/16384) - .5f;
509	896k	#endif
510	896k	oldEBands[i+c*m->nbEBands] += offset;
511	896k	} while (++c < C);
512	708k	}
513	116k	}
514
515		void unquant_energy_finalise(const CELTMode m, int start, int end, celt_glog oldEBands, int fine_quant, int fine_priority, int bits_left, ec_dec *dec, int C)
516	116k	{
517	116k	int i, prio, c;
518
519		/* Use up the remaining bits */
520	350k	for (prio=0;prio<2;prio++)
521	233k	{
522	1.08M	for (i=start;i<end && bits_left>=C ;i++)
523	846k	{
524	846k	if (fine_quant[i] >= MAX_FINE_BITS \|\| fine_priority[i]!=prio)
525	682k	continue;
526	164k	c=0;
527	218k	do {
528	218k	int q2;
529	218k	celt_glog offset;
530	218k	q2 = ec_dec_bits(dec, 1);
531		#ifdef FIXED_POINT
532		offset = SHR32(SHL32(q2,DB_SHIFT)-GCONST(.5f),fine_quant[i]+1);
533		#else
534	218k	offset = (q2-.5f)(1<<(14-fine_quant[i]-1))(1.f/16384);
535	218k	#endif
536	218k	oldEBands[i+c*m->nbEBands] += offset;
537	218k	bits_left--;
538	218k	} while (++c < C);
539	164k	}
540	233k	}
541	116k	}
542
543		void amp2Log2(const CELTMode *m, int effEnd, int end,
544		celt_ener bandE, celt_glog bandLogE, int C)
545	0	{
546	0	int c, i;
547	0	c=0;
548	0	do {
549	0	for (i=0;i<effEnd;i++)
550	0	{
551	0	bandLogE[i+c*m->nbEBands] =
552	0	celt_log2_db(bandE[i+c*m->nbEBands])
553	0	- SHL32((celt_glog)eMeans[i],DB_SHIFT-4);
554		#ifdef FIXED_POINT
555		/* Compensate for bandE[] being Q12 but celt_log2() taking a Q14 input. */
556		bandLogE[i+c*m->nbEBands] += GCONST(2.f);
557		#endif
558	0	}
559	0	for (i=effEnd;i<end;i++)
560	0	bandLogE[c*m->nbEBands+i] = -GCONST(14.f);
561	0	} while (++c < C);
562	0	}