opus_encoder.c

/* Copyright (c) 2010-2011 Xiph.Org Foundation, Skype Limited
   Written by Jean-Marc Valin and Koen Vos */
/*
   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 <stdarg.h>
#include "celt.h"
#include "entenc.h"
#include "modes.h"
#include "API.h"
#include "stack_alloc.h"
#include "float_cast.h"
#include "opus.h"
#include "arch.h"
#include "pitch.h"
#include "opus_private.h"
#include "os_support.h"
#include "cpu_support.h"
#include "analysis.h"
#include "mathops.h"
#include "tuning_parameters.h"
#ifdef FIXED_POINT
#include "fixed/structs_FIX.h"
#else
#include "float/structs_FLP.h"
#endif

#define MAX_ENCODER_BUFFER 480

typedef struct {
   opus_val32 XX, XY, YY;
   opus_val16 smoothed_width;
   opus_val16 max_follower;
} StereoWidthState;

struct OpusEncoder {
    int          celt_enc_offset;
    int          silk_enc_offset;
    silk_EncControlStruct silk_mode;
    int          application;
    int          channels;
    int          delay_compensation;
    int          force_channels;
    int          signal_type;
    int          user_bandwidth;
    int          max_bandwidth;
    int          user_forced_mode;
    int          voice_ratio;
    opus_int32   Fs;
    int          use_vbr;
    int          vbr_constraint;
    int          variable_duration;
    opus_int32   bitrate_bps;
    opus_int32   user_bitrate_bps;
    int          lsb_depth;
    int          encoder_buffer;
    int          lfe;

#define OPUS_ENCODER_RESET_START stream_channels
    int          stream_channels;
    opus_int16   hybrid_stereo_width_Q14;
    opus_int32   variable_HP_smth2_Q15;
    opus_val16   prev_HB_gain;
    opus_val32   hp_mem[4];
    int          mode;
    int          prev_mode;
    int          prev_channels;
    int          prev_framesize;
    int          bandwidth;
    int          silk_bw_switch;
    /* Sampling rate (at the API level) */
    int          first;
    opus_val16 * energy_masking;
    StereoWidthState width_mem;
    opus_val16   delay_buffer[MAX_ENCODER_BUFFER*2];
#ifndef DISABLE_FLOAT_API
    TonalityAnalysisState analysis;
    int          detected_bandwidth;
    int          analysis_offset;
#endif
    opus_uint32  rangeFinal;
    int          arch;
};

/* Transition tables for the voice and music. First column is the
   middle (memoriless) threshold. The second column is the hysteresis
   (difference with the middle) */
static const opus_int32 mono_voice_bandwidth_thresholds[8] = {
        11000, 1000, /* NB<->MB */
        14000, 1000, /* MB<->WB */
        17000, 1000, /* WB<->SWB */
        21000, 2000, /* SWB<->FB */
};
static const opus_int32 mono_music_bandwidth_thresholds[8] = {
        12000, 1000, /* NB<->MB */
        15000, 1000, /* MB<->WB */
        18000, 2000, /* WB<->SWB */
        22000, 2000, /* SWB<->FB */
};
static const opus_int32 stereo_voice_bandwidth_thresholds[8] = {
        11000, 1000, /* NB<->MB */
        14000, 1000, /* MB<->WB */
        21000, 2000, /* WB<->SWB */
        28000, 2000, /* SWB<->FB */
};
static const opus_int32 stereo_music_bandwidth_thresholds[8] = {
        12000, 1000, /* NB<->MB */
        18000, 2000, /* MB<->WB */
        21000, 2000, /* WB<->SWB */
        30000, 2000, /* SWB<->FB */
};
/* Threshold bit-rates for switching between mono and stereo */
static const opus_int32 stereo_voice_threshold = 30000;
static const opus_int32 stereo_music_threshold = 30000;

/* Threshold bit-rate for switching between SILK/hybrid and CELT-only */
static const opus_int32 mode_thresholds[2][2] = {
      /* voice */ /* music */
      {  64000,      16000}, /* mono */
      {  36000,      16000}, /* stereo */
};

int opus_encoder_get_size(int channels)
{
    int silkEncSizeBytes, celtEncSizeBytes;
    int ret;
    if (channels<1 || channels > 2)
        return 0;
    ret = silk_Get_Encoder_Size( &silkEncSizeBytes );
    if (ret)
        return 0;
    silkEncSizeBytes = align(silkEncSizeBytes);
    celtEncSizeBytes = celt_encoder_get_size(channels);
    return align(sizeof(OpusEncoder))+silkEncSizeBytes+celtEncSizeBytes;
}

int opus_encoder_init(OpusEncoder* st, opus_int32 Fs, int channels, int application)
{
    void *silk_enc;
    CELTEncoder *celt_enc;
    int err;
    int ret, silkEncSizeBytes;

   if((Fs!=48000&&Fs!=24000&&Fs!=16000&&Fs!=12000&&Fs!=8000)||(channels!=1&&channels!=2)||
        (application != OPUS_APPLICATION_VOIP && application != OPUS_APPLICATION_AUDIO
        && application != OPUS_APPLICATION_RESTRICTED_LOWDELAY))
        return OPUS_BAD_ARG;

    OPUS_CLEAR((char*)st, opus_encoder_get_size(channels));
    /* Create SILK encoder */
    ret = silk_Get_Encoder_Size( &silkEncSizeBytes );
    if (ret)
        return OPUS_BAD_ARG;
    silkEncSizeBytes = align(silkEncSizeBytes);
    st->silk_enc_offset = align(sizeof(OpusEncoder));
    st->celt_enc_offset = st->silk_enc_offset+silkEncSizeBytes;
    silk_enc = (char*)st+st->silk_enc_offset;
    celt_enc = (CELTEncoder*)((char*)st+st->celt_enc_offset);

    st->stream_channels = st->channels = channels;

    st->Fs = Fs;

    st->arch = opus_select_arch();

    ret = silk_InitEncoder( silk_enc, st->arch, &st->silk_mode );
    if(ret)return OPUS_INTERNAL_ERROR;

    /* default SILK parameters */
    st->silk_mode.nChannelsAPI              = channels;
    st->silk_mode.nChannelsInternal         = channels;
    st->silk_mode.API_sampleRate            = st->Fs;
    st->silk_mode.maxInternalSampleRate     = 16000;
    st->silk_mode.minInternalSampleRate     = 8000;
    st->silk_mode.desiredInternalSampleRate = 16000;
    st->silk_mode.payloadSize_ms            = 20;
    st->silk_mode.bitRate                   = 25000;
    st->silk_mode.packetLossPercentage      = 0;
    st->silk_mode.complexity                = 9;
    st->silk_mode.useInBandFEC              = 0;
    st->silk_mode.useDTX                    = 0;
    st->silk_mode.useCBR                    = 0;
    st->silk_mode.reducedDependency         = 0;

    /* Create CELT encoder */
    /* Initialize CELT encoder */
    err = celt_encoder_init(celt_enc, Fs, channels, st->arch);
    if(err!=OPUS_OK)return OPUS_INTERNAL_ERROR;

    celt_encoder_ctl(celt_enc, CELT_SET_SIGNALLING(0));
    celt_encoder_ctl(celt_enc, OPUS_SET_COMPLEXITY(st->silk_mode.complexity));

    st->use_vbr = 1;
    /* Makes constrained VBR the default (safer for real-time use) */
    st->vbr_constraint = 1;
    st->user_bitrate_bps = OPUS_AUTO;
    st->bitrate_bps = 3000+Fs*channels;
    st->application = application;
    st->signal_type = OPUS_AUTO;
    st->user_bandwidth = OPUS_AUTO;
    st->max_bandwidth = OPUS_BANDWIDTH_FULLBAND;
    st->force_channels = OPUS_AUTO;
    st->user_forced_mode = OPUS_AUTO;
    st->voice_ratio = -1;
    st->encoder_buffer = st->Fs/100;
    st->lsb_depth = 24;
    st->variable_duration = OPUS_FRAMESIZE_ARG;

    /* Delay compensation of 4 ms (2.5 ms for SILK's extra look-ahead
       + 1.5 ms for SILK resamplers and stereo prediction) */
    st->delay_compensation = st->Fs/250;

    st->hybrid_stereo_width_Q14 = 1 << 14;
    st->prev_HB_gain = Q15ONE;
    st->variable_HP_smth2_Q15 = silk_LSHIFT( silk_lin2log( VARIABLE_HP_MIN_CUTOFF_HZ ), 8 );
    st->first = 1;
    st->mode = MODE_HYBRID;
    st->bandwidth = OPUS_BANDWIDTH_FULLBAND;

    return OPUS_OK;
}

static unsigned char gen_toc(int mode, int framerate, int bandwidth, int channels)
{
   int period;
   unsigned char toc;
   period = 0;
   while (framerate < 400)
   {
       framerate <<= 1;
       period++;
   }
   if (mode == MODE_SILK_ONLY)
   {
       toc = (bandwidth-OPUS_BANDWIDTH_NARROWBAND)<<5;
       toc |= (period-2)<<3;
   } else if (mode == MODE_CELT_ONLY)
   {
       int tmp = bandwidth-OPUS_BANDWIDTH_MEDIUMBAND;
       if (tmp < 0)
           tmp = 0;
       toc = 0x80;
       toc |= tmp << 5;
       toc |= period<<3;
   } else /* Hybrid */
   {
       toc = 0x60;
       toc |= (bandwidth-OPUS_BANDWIDTH_SUPERWIDEBAND)<<4;
       toc |= (period-2)<<3;
   }
   toc |= (channels==2)<<2;
   return toc;
}

#ifndef FIXED_POINT
static void silk_biquad_float(
    const opus_val16      *in,            /* I:    Input signal                   */
    const opus_int32      *B_Q28,         /* I:    MA coefficients [3]            */
    const opus_int32      *A_Q28,         /* I:    AR coefficients [2]            */
    opus_val32            *S,             /* I/O:  State vector [2]               */
    opus_val16            *out,           /* O:    Output signal                  */
    const opus_int32      len,            /* I:    Signal length (must be even)   */
    int stride
)
{
    /* DIRECT FORM II TRANSPOSED (uses 2 element state vector) */
    opus_int   k;
    opus_val32 vout;
    opus_val32 inval;
    opus_val32 A[2], B[3];

    A[0] = (opus_val32)(A_Q28[0] * (1.f/((opus_int32)1<<28)));
    A[1] = (opus_val32)(A_Q28[1] * (1.f/((opus_int32)1<<28)));
    B[0] = (opus_val32)(B_Q28[0] * (1.f/((opus_int32)1<<28)));
    B[1] = (opus_val32)(B_Q28[1] * (1.f/((opus_int32)1<<28)));
    B[2] = (opus_val32)(B_Q28[2] * (1.f/((opus_int32)1<<28)));

    /* Negate A_Q28 values and split in two parts */

    for( k = 0; k < len; k++ ) {
        /* S[ 0 ], S[ 1 ]: Q12 */
        inval = in[ k*stride ];
        vout = S[ 0 ] + B[0]*inval;

        S[ 0 ] = S[1] - vout*A[0] + B[1]*inval;

        S[ 1 ] = - vout*A[1] + B[2]*inval + VERY_SMALL;

        /* Scale back to Q0 and saturate */
        out[ k*stride ] = vout;
    }
}
#endif

static void hp_cutoff(const opus_val16 *in, opus_int32 cutoff_Hz, opus_val16 *out, opus_val32 *hp_mem, int len, int channels, opus_int32 Fs)
{
   opus_int32 B_Q28[ 3 ], A_Q28[ 2 ];
   opus_int32 Fc_Q19, r_Q28, r_Q22;

   silk_assert( cutoff_Hz <= silk_int32_MAX / SILK_FIX_CONST( 1.5 * 3.14159 / 1000, 19 ) );
   Fc_Q19 = silk_DIV32_16( silk_SMULBB( SILK_FIX_CONST( 1.5 * 3.14159 / 1000, 19 ), cutoff_Hz ), Fs/1000 );
   silk_assert( Fc_Q19 > 0 && Fc_Q19 < 32768 );

   r_Q28 = SILK_FIX_CONST( 1.0, 28 ) - silk_MUL( SILK_FIX_CONST( 0.92, 9 ), Fc_Q19 );

   /* b = r * [ 1; -2; 1 ]; */
   /* a = [ 1; -2 * r * ( 1 - 0.5 * Fc^2 ); r^2 ]; */
   B_Q28[ 0 ] = r_Q28;
   B_Q28[ 1 ] = silk_LSHIFT( -r_Q28, 1 );
   B_Q28[ 2 ] = r_Q28;

   /* -r * ( 2 - Fc * Fc ); */
   r_Q22  = silk_RSHIFT( r_Q28, 6 );
   A_Q28[ 0 ] = silk_SMULWW( r_Q22, silk_SMULWW( Fc_Q19, Fc_Q19 ) - SILK_FIX_CONST( 2.0,  22 ) );
   A_Q28[ 1 ] = silk_SMULWW( r_Q22, r_Q22 );

#ifdef FIXED_POINT
   silk_biquad_alt( in, B_Q28, A_Q28, hp_mem, out, len, channels );
   if( channels == 2 ) {
       silk_biquad_alt( in+1, B_Q28, A_Q28, hp_mem+2, out+1, len, channels );
   }
#else
   silk_biquad_float( in, B_Q28, A_Q28, hp_mem, out, len, channels );
   if( channels == 2 ) {
       silk_biquad_float( in+1, B_Q28, A_Q28, hp_mem+2, out+1, len, channels );
   }
#endif
}

#ifdef FIXED_POINT
static void dc_reject(const opus_val16 *in, opus_int32 cutoff_Hz, opus_val16 *out, opus_val32 *hp_mem, int len, int channels, opus_int32 Fs)
{
   int c, i;
   int shift;

   /* Approximates -round(log2(4.*cutoff_Hz/Fs)) */
   shift=celt_ilog2(Fs/(cutoff_Hz*3));
   for (c=0;c<channels;c++)
   {
      for (i=0;i<len;i++)
      {
         opus_val32 x, tmp, y;
         x = SHL32(EXTEND32(in[channels*i+c]), 15);
         /* First stage */
         tmp = x-hp_mem[2*c];
         hp_mem[2*c] = hp_mem[2*c] + PSHR32(x - hp_mem[2*c], shift);
         /* Second stage */
         y = tmp - hp_mem[2*c+1];
         hp_mem[2*c+1] = hp_mem[2*c+1] + PSHR32(tmp - hp_mem[2*c+1], shift);
         out[channels*i+c] = EXTRACT16(SATURATE(PSHR32(y, 15), 32767));
      }
   }
}

#else
static void dc_reject(const opus_val16 *in, opus_int32 cutoff_Hz, opus_val16 *out, opus_val32 *hp_mem, int len, int channels, opus_int32 Fs)
{
   int c, i;
   float coef;

   coef = 4.0f*cutoff_Hz/Fs;
   for (c=0;c<channels;c++)
   {
      for (i=0;i<len;i++)
      {
         opus_val32 x, tmp, y;
         x = in[channels*i+c];
         /* First stage */
         tmp = x-hp_mem[2*c];
         hp_mem[2*c] = hp_mem[2*c] + coef*(x - hp_mem[2*c]) + VERY_SMALL;
         /* Second stage */
         y = tmp - hp_mem[2*c+1];
         hp_mem[2*c+1] = hp_mem[2*c+1] + coef*(tmp - hp_mem[2*c+1]) + VERY_SMALL;
         out[channels*i+c] = y;
      }
   }
}
#endif

static void stereo_fade(const opus_val16 *in, opus_val16 *out, opus_val16 g1, opus_val16 g2,
        int overlap48, int frame_size, int channels, const opus_val16 *window, opus_int32 Fs)
{
    int i;
    int overlap;
    int inc;
    inc = 48000/Fs;
    overlap=overlap48/inc;
    g1 = Q15ONE-g1;
    g2 = Q15ONE-g2;
    for (i=0;i<overlap;i++)
    {
       opus_val32 diff;
       opus_val16 g, w;
       w = MULT16_16_Q15(window[i*inc], window[i*inc]);
       g = SHR32(MAC16_16(MULT16_16(w,g2),
             Q15ONE-w, g1), 15);
       diff = EXTRACT16(HALF32((opus_val32)in[i*channels] - (opus_val32)in[i*channels+1]));
       diff = MULT16_16_Q15(g, diff);
       out[i*channels] = out[i*channels] - diff;
       out[i*channels+1] = out[i*channels+1] + diff;
    }
    for (;i<frame_size;i++)
    {
       opus_val32 diff;
       diff = EXTRACT16(HALF32((opus_val32)in[i*channels] - (opus_val32)in[i*channels+1]));
       diff = MULT16_16_Q15(g2, diff);
       out[i*channels] = out[i*channels] - diff;
       out[i*channels+1] = out[i*channels+1] + diff;
    }
}

static void gain_fade(const opus_val16 *in, opus_val16 *out, opus_val16 g1, opus_val16 g2,
        int overlap48, int frame_size, int channels, const opus_val16 *window, opus_int32 Fs)
{
    int i;
    int inc;
    int overlap;
    int c;
    inc = 48000/Fs;
    overlap=overlap48/inc;
    if (channels==1)
    {
       for (i=0;i<overlap;i++)
       {
          opus_val16 g, w;
          w = MULT16_16_Q15(window[i*inc], window[i*inc]);
          g = SHR32(MAC16_16(MULT16_16(w,g2),
                Q15ONE-w, g1), 15);
          out[i] = MULT16_16_Q15(g, in[i]);
       }
    } else {
       for (i=0;i<overlap;i++)
       {
          opus_val16 g, w;
          w = MULT16_16_Q15(window[i*inc], window[i*inc]);
          g = SHR32(MAC16_16(MULT16_16(w,g2),
                Q15ONE-w, g1), 15);
          out[i*2] = MULT16_16_Q15(g, in[i*2]);
          out[i*2+1] = MULT16_16_Q15(g, in[i*2+1]);
       }
    }
    c=0;do {
       for (i=overlap;i<frame_size;i++)
       {
          out[i*channels+c] = MULT16_16_Q15(g2, in[i*channels+c]);
       }
    }
    while (++c<channels);
}

OpusEncoder *opus_encoder_create(opus_int32 Fs, int channels, int application, int *error)
{
   int ret;
   OpusEncoder *st;
   if((Fs!=48000&&Fs!=24000&&Fs!=16000&&Fs!=12000&&Fs!=8000)||(channels!=1&&channels!=2)||
       (application != OPUS_APPLICATION_VOIP && application != OPUS_APPLICATION_AUDIO
       && application != OPUS_APPLICATION_RESTRICTED_LOWDELAY))
   {
      if (error)
         *error = OPUS_BAD_ARG;
      return NULL;
   }
   st = (OpusEncoder *)opus_alloc(opus_encoder_get_size(channels));
   if (st == NULL)
   {
      if (error)
         *error = OPUS_ALLOC_FAIL;
      return NULL;
   }
   ret = opus_encoder_init(st, Fs, channels, application);
   if (error)
      *error = ret;
   if (ret != OPUS_OK)
   {
      opus_free(st);
      st = NULL;
   }
   return st;
}

static opus_int32 user_bitrate_to_bitrate(OpusEncoder *st, int frame_size, int max_data_bytes)
{
  if(!frame_size)frame_size=st->Fs/400;
  if (st->user_bitrate_bps==OPUS_AUTO)
    return 60*st->Fs/frame_size + st->Fs*st->channels;
  else if (st->user_bitrate_bps==OPUS_BITRATE_MAX)
    return max_data_bytes*8*st->Fs/frame_size;
  else
    return st->user_bitrate_bps;
}

#ifndef DISABLE_FLOAT_API
/* Don't use more than 60 ms for the frame size analysis */
#define MAX_DYNAMIC_FRAMESIZE 24
/* Estimates how much the bitrate will be boosted based on the sub-frame energy */
static float transient_boost(const float *E, const float *E_1, int LM, int maxM)
{
   int i;
   int M;
   float sumE=0, sumE_1=0;
   float metric;

   M = IMIN(maxM, (1<<LM)+1);
   for (i=0;i<M;i++)
   {
      sumE += E[i];
      sumE_1 += E_1[i];
   }
   metric = sumE*sumE_1/(M*M);
   /*if (LM==3)
      printf("%f\n", metric);*/
   /*return metric>10 ? 1 : 0;*/
   /*return MAX16(0,1-exp(-.25*(metric-2.)));*/
   return MIN16(1,(float)sqrt(MAX16(0,.05f*(metric-2))));
}

/* Viterbi decoding trying to find the best frame size combination using look-ahead

   State numbering:
    0: unused
    1:  2.5 ms
    2:  5 ms (#1)
    3:  5 ms (#2)
    4: 10 ms (#1)
    5: 10 ms (#2)
    6: 10 ms (#3)
    7: 10 ms (#4)
    8: 20 ms (#1)
    9: 20 ms (#2)
   10: 20 ms (#3)
   11: 20 ms (#4)
   12: 20 ms (#5)
   13: 20 ms (#6)
   14: 20 ms (#7)
   15: 20 ms (#8)
*/
static int transient_viterbi(const float *E, const float *E_1, int N, int frame_cost, int rate)
{
   int i;
   float cost[MAX_DYNAMIC_FRAMESIZE][16];
   int states[MAX_DYNAMIC_FRAMESIZE][16];
   float best_cost;
   int best_state;
   float factor;
   /* Take into account that we damp VBR in the 32 kb/s to 64 kb/s range. */
   if (rate<80)
      factor=0;
   else if (rate>160)
      factor=1;
   else
      factor = (rate-80.f)/80.f;
   /* Makes variable framesize less aggressive at lower bitrates, but I can't
      find any valid theoretical justification for this (other than it seems
      to help) */
   for (i=0;i<16;i++)
   {
      /* Impossible state */
      states[0][i] = -1;
      cost[0][i] = 1e10;
   }
   for (i=0;i<4;i++)
   {
      cost[0][1<<i] = (frame_cost + rate*(1<<i))*(1+factor*transient_boost(E, E_1, i, N+1));
      states[0][1<<i] = i;
   }
   for (i=1;i<N;i++)
   {
      int j;

      /* Follow continuations */
      for (j=2;j<16;j++)
      {
         cost[i][j] = cost[i-1][j-1];
         states[i][j] = j-1;
      }

      /* New frames */
      for(j=0;j<4;j++)
      {
         int k;
         float min_cost;
         float curr_cost;
         states[i][1<<j] = 1;
         min_cost = cost[i-1][1];
         for(k=1;k<4;k++)
         {
            float tmp = cost[i-1][(1<<(k+1))-1];
            if (tmp < min_cost)
            {
               states[i][1<<j] = (1<<(k+1))-1;
               min_cost = tmp;
            }
         }
         curr_cost = (frame_cost + rate*(1<<j))*(1+factor*transient_boost(E+i, E_1+i, j, N-i+1));
         cost[i][1<<j] = min_cost;
         /* If part of the frame is outside the analysis window, only count part of the cost */
         if (N-i < (1<<j))
            cost[i][1<<j] += curr_cost*(float)(N-i)/(1<<j);
         else
            cost[i][1<<j] += curr_cost;
      }
   }

   best_state=1;
   best_cost = cost[N-1][1];
   /* Find best end state (doesn't force a frame to end at N-1) */
   for (i=2;i<16;i++)
   {
      if (cost[N-1][i]<best_cost)
      {
         best_cost = cost[N-1][i];
         best_state = i;
      }
   }

   /* Follow transitions back */
   for (i=N-1;i>=0;i--)
   {
      /*printf("%d ", best_state);*/
      best_state = states[i][best_state];
   }
   /*printf("%d\n", best_state);*/
   return best_state;
}

static int optimize_framesize(const void *x, int len, int C, opus_int32 Fs,
                int bitrate, opus_val16 tonality, float *mem, int buffering,
                downmix_func downmix)
{
   int N;
   int i;
   float e[MAX_DYNAMIC_FRAMESIZE+4];
   float e_1[MAX_DYNAMIC_FRAMESIZE+3];
   opus_val32 memx;
   int bestLM=0;
   int subframe;
   int pos;
   int offset;
   VARDECL(opus_val32, sub);

   subframe = Fs/400;
   ALLOC(sub, subframe, opus_val32);
   e[0]=mem[0];
   e_1[0]=1.f/(EPSILON+mem[0]);
   if (buffering)
   {
      /* Consider the CELT delay when not in restricted-lowdelay */
      /* We assume the buffering is between 2.5 and 5 ms */
      offset = 2*subframe - buffering;
      celt_assert(offset>=0 && offset <= subframe);
      len -= offset;
      e[1]=mem[1];
      e_1[1]=1.f/(EPSILON+mem[1]);
      e[2]=mem[2];
      e_1[2]=1.f/(EPSILON+mem[2]);
      pos = 3;
   } else {
      pos=1;
      offset=0;
   }
   N=IMIN(len/subframe, MAX_DYNAMIC_FRAMESIZE);
   /* Just silencing a warning, it's really initialized later */
   memx = 0;
   for (i=0;i<N;i++)
   {
      float tmp;
      opus_val32 tmpx;
      int j;
      tmp=EPSILON;

      downmix(x, sub, subframe, i*subframe+offset, 0, -2, C);
      if (i==0)
         memx = sub[0];
      for (j=0;j<subframe;j++)
      {
         tmpx = sub[j];
         tmp += (tmpx-memx)*(float)(tmpx-memx);
         memx = tmpx;
      }
      e[i+pos] = tmp;
      e_1[i+pos] = 1.f/tmp;
   }
   /* Hack to get 20 ms working with APPLICATION_AUDIO
      The real problem is that the corresponding memory needs to use 1.5 ms
      from this frame and 1 ms from the next frame */
   e[i+pos] = e[i+pos-1];
   if (buffering)
      N=IMIN(MAX_DYNAMIC_FRAMESIZE, N+2);
   bestLM = transient_viterbi(e, e_1, N, (int)((1.f+.5f*tonality)*(60*C+40)), bitrate/400);
   mem[0] = e[1<<bestLM];
   if (buffering)
   {
      mem[1] = e[(1<<bestLM)+1];
      mem[2] = e[(1<<bestLM)+2];
   }
   return bestLM;
}

#endif

#ifndef DISABLE_FLOAT_API
#ifdef FIXED_POINT
#define PCM2VAL(x) FLOAT2INT16(x)
#else
#define PCM2VAL(x) SCALEIN(x)
#endif
void downmix_float(const void *_x, opus_val32 *sub, int subframe, int offset, int c1, int c2, int C)
{
   const float *x;
   opus_val32 scale;
   int j;
   x = (const float *)_x;
   for (j=0;j<subframe;j++)
      sub[j] = PCM2VAL(x[(j+offset)*C+c1]);
   if (c2>-1)
   {
      for (j=0;j<subframe;j++)
         sub[j] += PCM2VAL(x[(j+offset)*C+c2]);
   } else if (c2==-2)
   {
      int c;
      for (c=1;c<C;c++)
      {
         for (j=0;j<subframe;j++)
            sub[j] += PCM2VAL(x[(j+offset)*C+c]);
      }
   }
#ifdef FIXED_POINT
   scale = (1<<SIG_SHIFT);
#else
   scale = 1.f;
#endif
   if (C==-2)
      scale /= C;
   else
      scale /= 2;
   for (j=0;j<subframe;j++)
      sub[j] *= scale;
}
#endif

void downmix_int(const void *_x, opus_val32 *sub, int subframe, int offset, int c1, int c2, int C)
{
   const opus_int16 *x;
   opus_val32 scale;
   int j;
   x = (const opus_int16 *)_x;
   for (j=0;j<subframe;j++)
      sub[j] = x[(j+offset)*C+c1];
   if (c2>-1)
   {
      for (j=0;j<subframe;j++)
         sub[j] += x[(j+offset)*C+c2];
   } else if (c2==-2)
   {
      int c;
      for (c=1;c<C;c++)
      {
         for (j=0;j<subframe;j++)
            sub[j] += x[(j+offset)*C+c];
      }
   }
#ifdef FIXED_POINT
   scale = (1<<SIG_SHIFT);
#else
   scale = 1.f/32768;
#endif
   if (C==-2)
      scale /= C;
   else
      scale /= 2;
   for (j=0;j<subframe;j++)
      sub[j] *= scale;
}

opus_int32 frame_size_select(opus_int32 frame_size, int variable_duration, opus_int32 Fs)
{
   int new_size;
   if (frame_size<Fs/400)
      return -1;
   if (variable_duration == OPUS_FRAMESIZE_ARG)
      new_size = frame_size;
   else if (variable_duration == OPUS_FRAMESIZE_VARIABLE)
      new_size = Fs/50;
   else if (variable_duration >= OPUS_FRAMESIZE_2_5_MS && variable_duration <= OPUS_FRAMESIZE_60_MS)
      new_size = IMIN(3*Fs/50, (Fs/400)<<(variable_duration-OPUS_FRAMESIZE_2_5_MS));
   else
      return -1;
   if (new_size>frame_size)
      return -1;
   if (400*new_size!=Fs && 200*new_size!=Fs && 100*new_size!=Fs &&
            50*new_size!=Fs && 25*new_size!=Fs && 50*new_size!=3*Fs)
      return -1;
   return new_size;
}

opus_int32 compute_frame_size(const void *analysis_pcm, int frame_size,
      int variable_duration, int C, opus_int32 Fs, int bitrate_bps,
      int delay_compensation, downmix_func downmix
#ifndef DISABLE_FLOAT_API
      , float *subframe_mem
#endif
      )
{
#ifndef DISABLE_FLOAT_API
   if (variable_duration == OPUS_FRAMESIZE_VARIABLE && frame_size >= Fs/200)
   {
      int LM = 3;
      LM = optimize_framesize(analysis_pcm, frame_size, C, Fs, bitrate_bps,
            0, subframe_mem, delay_compensation, downmix);
      while ((Fs/400<<LM)>frame_size)
         LM--;
      frame_size = (Fs/400<<LM);
   } else
#else
   (void)analysis_pcm;
   (void)C;
   (void)bitrate_bps;
   (void)delay_compensation;
   (void)downmix;
#endif
   {
      frame_size = frame_size_select(frame_size, variable_duration, Fs);
   }
   if (frame_size<0)
      return -1;
   return frame_size;
}

opus_val16 compute_stereo_width(const opus_val16 *pcm, int frame_size, opus_int32 Fs, StereoWidthState *mem)
{
   opus_val16 corr;
   opus_val16 ldiff;
   opus_val16 width;
   opus_val32 xx, xy, yy;
   opus_val16 sqrt_xx, sqrt_yy;
   opus_val16 qrrt_xx, qrrt_yy;
   int frame_rate;
   int i;
   opus_val16 short_alpha;

   frame_rate = Fs/frame_size;
   short_alpha = Q15ONE - 25*Q15ONE/IMAX(50,frame_rate);
   xx=xy=yy=0;
   for (i=0;i<frame_size;i+=4)
   {
      opus_val32 pxx=0;
      opus_val32 pxy=0;
      opus_val32 pyy=0;
      opus_val16 x, y;
      x = pcm[2*i];
      y = pcm[2*i+1];
      pxx = SHR32(MULT16_16(x,x),2);
      pxy = SHR32(MULT16_16(x,y),2);
      pyy = SHR32(MULT16_16(y,y),2);
      x = pcm[2*i+2];
      y = pcm[2*i+3];
      pxx += SHR32(MULT16_16(x,x),2);
      pxy += SHR32(MULT16_16(x,y),2);
      pyy += SHR32(MULT16_16(y,y),2);
      x = pcm[2*i+4];
      y = pcm[2*i+5];
      pxx += SHR32(MULT16_16(x,x),2);
      pxy += SHR32(MULT16_16(x,y),2);
      pyy += SHR32(MULT16_16(y,y),2);
      x = pcm[2*i+6];
      y = pcm[2*i+7];
      pxx += SHR32(MULT16_16(x,x),2);
      pxy += SHR32(MULT16_16(x,y),2);
      pyy += SHR32(MULT16_16(y,y),2);

      xx += SHR32(pxx, 10);
      xy += SHR32(pxy, 10);
      yy += SHR32(pyy, 10);
   }
   mem->XX += MULT16_32_Q15(short_alpha, xx-mem->XX);
   mem->XY += MULT16_32_Q15(short_alpha, xy-mem->XY);
   mem->YY += MULT16_32_Q15(short_alpha, yy-mem->YY);
   mem->XX = MAX32(0, mem->XX);
   mem->XY = MAX32(0, mem->XY);
   mem->YY = MAX32(0, mem->YY);
   if (MAX32(mem->XX, mem->YY)>QCONST16(8e-4f, 18))
   {
      sqrt_xx = celt_sqrt(mem->XX);
      sqrt_yy = celt_sqrt(mem->YY);
      qrrt_xx = celt_sqrt(sqrt_xx);
      qrrt_yy = celt_sqrt(sqrt_yy);
      /* Inter-channel correlation */
      mem->XY = MIN32(mem->XY, sqrt_xx*sqrt_yy);
      corr = SHR32(frac_div32(mem->XY,EPSILON+MULT16_16(sqrt_xx,sqrt_yy)),16);
      /* Approximate loudness difference */
      ldiff = Q15ONE*ABS16(qrrt_xx-qrrt_yy)/(EPSILON+qrrt_xx+qrrt_yy);
      width = MULT16_16_Q15(celt_sqrt(QCONST32(1.f,30)-MULT16_16(corr,corr)), ldiff);
      /* Smoothing over one second */
      mem->smoothed_width += (width-mem->smoothed_width)/frame_rate;
      /* Peak follower */
      mem->max_follower = MAX16(mem->max_follower-QCONST16(.02f,15)/frame_rate, mem->smoothed_width);
   } else {
      width = 0;
      corr=Q15ONE;
      ldiff=0;
   }
   /*printf("%f %f %f %f %f ", corr/(float)Q15ONE, ldiff/(float)Q15ONE, width/(float)Q15ONE, mem->smoothed_width/(float)Q15ONE, mem->max_follower/(float)Q15ONE);*/
   return EXTRACT16(MIN32(Q15ONE,20*mem->max_follower));
}

opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_size,
                unsigned char *data, opus_int32 out_data_bytes, int lsb_depth,
                const void *analysis_pcm, opus_int32 analysis_size, int c1, int c2,
                int analysis_channels, downmix_func downmix, int float_api)
{
    void *silk_enc;
    CELTEncoder *celt_enc;
    int i;
    int ret=0;
    opus_int32 nBytes;
    ec_enc enc;
    int bytes_target;
    int prefill=0;
    int start_band = 0;
    int redundancy = 0;
    int redundancy_bytes = 0; /* Number of bytes to use for redundancy frame */
    int celt_to_silk = 0;
    VARDECL(opus_val16, pcm_buf);
    int nb_compr_bytes;
    int to_celt = 0;
    opus_uint32 redundant_rng = 0;
    int cutoff_Hz, hp_freq_smth1;
    int voice_est; /* Probability of voice in Q7 */
    opus_int32 equiv_rate;
    int delay_compensation;
    int frame_rate;
    opus_int32 max_rate; /* Max bitrate we're allowed to use */
    int curr_bandwidth;
    opus_val16 HB_gain;
    opus_int32 max_data_bytes; /* Max number of bytes we're allowed to use */
    int total_buffer;
    opus_val16 stereo_width;
    const CELTMode *celt_mode;
#ifndef DISABLE_FLOAT_API
    AnalysisInfo analysis_info;
    int analysis_read_pos_bak=-1;
    int analysis_read_subframe_bak=-1;
#endif
    VARDECL(opus_val16, tmp_prefill);

    ALLOC_STACK;

    max_data_bytes = IMIN(1276, out_data_bytes);

    st->rangeFinal = 0;
    if ((!st->variable_duration && 400*frame_size != st->Fs && 200*frame_size != st->Fs && 100*frame_size != st->Fs &&
         50*frame_size != st->Fs &&  25*frame_size != st->Fs &&  50*frame_size != 3*st->Fs)
         || (400*frame_size < st->Fs)
         || max_data_bytes<=0
         )
    {
       RESTORE_STACK;
       return OPUS_BAD_ARG;
    }
    silk_enc = (char*)st+st->silk_enc_offset;
    celt_enc = (CELTEncoder*)((char*)st+st->celt_enc_offset);
    if (st->application == OPUS_APPLICATION_RESTRICTED_LOWDELAY)
       delay_compensation = 0;
    else
       delay_compensation = st->delay_compensation;

    lsb_depth = IMIN(lsb_depth, st->lsb_depth);

    celt_encoder_ctl(celt_enc, CELT_GET_MODE(&celt_mode));
#ifndef DISABLE_FLOAT_API
    analysis_info.valid = 0;
#ifdef FIXED_POINT
    if (st->silk_mode.complexity >= 10 && st->Fs==48000)